Catharsis as a Service™

After a traumatic event, do not force early disclosure; seek structured help if symptoms persist

2026-09-14T00:00:00+00:00

Why early forced disclosure is not recommended

Psychological debriefing, a single session soon after a traumatic event in which people describe the event and express their reactions, was once widely used because it seemed consistent with the idea that early emotional release prevents later disorder. Rose, Bisson, Churchill and Wessely (2002) systematically reviewed randomised controlled trials of single-session individual debriefing. They found no evidence that it prevented post-traumatic stress disorder or reduced general distress, and in trials with longer follow-up there was some evidence of an increased risk of post-traumatic stress disorder among people who had been debriefed. The authors concluded that compulsory debriefing should cease.

What effective treatment involves

Emotional processing theory (Foa & Kozak, 1986) explains why structure matters. Change requires activating the traumatic memory together with information that corrects its threatening meaning, repeated over time within a safe, guided process. Treatments derived from this principle, such as prolonged exposure, differ fundamentally from one-off emotional release: they are planned, repeated and focused on new learning.

Other structured approaches also have evidence. For example, mindfulness-based therapies moderately reduce symptoms of anxiety and depression, with larger effects in people with diagnosed anxiety and mood disorders (Hofmann et al., 2010). The right approach depends on clinical assessment.

What further research adds

A later systematic review and meta-analysis of 61 randomised controlled trials examined multiple-session psychological interventions aimed at preventing or treating traumatic stress symptoms beginning within three months of a traumatic event (Roberts et al., 2019). For people exposed to trauma who had not been screened for traumatic stress symptoms, no intervention showed a clinically important difference from usual care. For people who reported traumatic stress symptoms, trauma-focused cognitive-behavioural therapy, cognitive therapy without exposure and eye movement desensitisation and reprocessing (EMDR) showed clinically important benefits, which were greatest for people diagnosed with acute stress disorder or PTSD. The evidence was strongest for trauma-focused cognitive-behavioural therapy.

For chronic PTSD, a Cochrane review of 70 studies with 4,761 participants found that individual trauma-focused cognitive-behavioural therapy and EMDR reduced clinician-assessed symptoms more than a waiting list or usual care (Bisson et al., 2013). The same review rated the evidence for every comparison as very low quality, found more drop-out in active treatment groups and noted that limited follow-up data compromise conclusions about long-term effects.

A meta-analysis of 75 studies found large associations between posttraumatic stress symptoms and both rumination (r = .52) and experiential avoidance (r = .48), and a medium association with thought suppression (r = .29) (Miethe et al., 2023). These are correlations, and the few longitudinal studies suggested somewhat smaller but still substantial associations. They do not show that any of these patterns causes symptoms.

In short

After trauma, release is neither required nor sufficient. Time, safety, support and, when symptoms persist, structured professional treatment are what the evidence supports.

After relief, ask which condition actually changed

2026-09-14T00:00:00+00:00

Why it matters

Any behaviour that reliably removes an unpleasant internal state is negatively reinforced and becomes more likely. This is adaptive when the behaviour also deals with the source of the state and self-perpetuating when it does not. Hayes and colleagues (1996) described experiential avoidance, the attempt to escape or change unwanted internal experiences even when this causes harm over time, as a process common to many psychological difficulties.

Why memory makes it harder to see

Retrospective evaluations of experiences are dominated by their most intense moment and by how they end, while their duration is largely neglected (Fredrickson & Kahneman, 1993). In one demonstration, participants chose to repeat a longer unpleasant experience that ended slightly better over a shorter one (Kahneman et al., 1993). A strong peak and a good ending make an evening remembered as more restorative than its lasting effect justifies.

What the review does

The four questions in this entry separate the effect of an experience on feeling from its effect on understanding, decisions and actions. They do not judge the experience. They make visible whether relief is standing in for resolution, which is the pattern the artifact describes as problem_solved: false.

What further research adds

Two lines of research sharpen the review. Kashdan, Barrios, Forsyth and Steger (2006) followed participants over a 21-day monitoring period. Experiential avoidance was associated with fewer positive daily experiences and positive events, more frequent negative events and greater negative affective experiences, and cognitive reappraisal was much less predictive of the quality of everyday experiences than experiential avoidance was. The design shows associations, not that avoidance caused the difference.

The peak-end pattern is also not universal. In three studies of meals, two with actual meals and one with an imagined meal, remembered liking showed consistent duration neglect, but none of the studies found peak, primacy or recency effects (Rode, Rozin & Durlach, 2007). How much a remembered peak or ending shapes the evaluation of an experience may therefore depend on the kind of experience, which is one more reason to write down the next day what actually changed rather than rely on memory.

Invest in support that is available when stress arrives

2026-09-14T00:00:00+00:00

Why it matters

Stress responses protect the body in the short term and contribute to wear when they are activated too often or do not switch off (McEwen, 1998). Anything that reduces how strongly or how long stress responses are engaged has long-term value.

What the evidence shows

Cohen and Wills (1985) reviewed studies of social support and well-being. Social integration had a general positive association with health, and perceived availability of support protected people specifically against the effects of stressful events. The buffering effect was strongest when the kind of support available matched the demands of the stressor.

Rimé (2009) reviewed research showing that people share most of their emotional experiences with others. Sharing reliably strengthens social bonds and brings attention and empathy, even though, on its own, it does not reduce the emotional impact of a memory. Sharing therefore contributes to support indirectly, by maintaining the relationships that make support available.

What further research adds

The long-term stakes are large. A meta-analysis of 148 studies with 308,849 participants found that people with stronger social relationships had a 50% increased likelihood of survival (odds ratio 1.50), a result that was consistent across age, sex, initial health status, cause of death and follow-up period (Holt-Lunstad, Smith & Layton, 2010). The association was strongest for complex measures of social integration and weakest for whether a person lived alone or with others. The authors judged the influence comparable with well-established risk factors for mortality. The estimate describes an association between relationships and survival, not a controlled test of building relationships.

How support is given also matters. Bolger, Zuckerman and Kessler (2000) note that although perceived availability of support buffers the effects of stressors, research had failed to show an association between actually receiving support and adjustment. In a daily diary study of couples, many support transactions reported by the supporter were not reported by the recipient, and these invisible transactions promoted adjustment to a major stressor. A review of 100 studies of social support interventions found some support for their overall usefulness, but not enough evidence to conclude which interventions work best for which problems (Hogan, Linden & Najarian, 2002).

Relation to collective experiences

A crowd can create a strong feeling of belonging for one night. Buffering depends on relationships that persist into the following week, when a specific stressor arrives. Both are valuable. Only the second is available on demand.

Turn abstract brooding into concrete next steps

2026-09-14T00:00:00+00:00

Why it matters

Rumination is a repetitive, passive focus on distress, its causes and its consequences. It feels like problem solving because it is about the problem, but it rarely produces decisions. Nolen-Hoeksema, Wisco and Lyubomirsky (2008) reviewed evidence that rumination prolongs negative mood, impairs problem solving and predicts the onset of depressive episodes.

What the evidence shows

Aldao, Nolen-Hoeksema and Schweizer (2010) meta-analysed studies relating emotion regulation strategies to symptoms across anxiety, depression, eating and substance-related disorders. Rumination had the largest association with psychopathology of all strategies examined. Problem solving and reappraisal were associated with fewer symptoms.

Watkins (2008) showed that the harm lies in a particular style of repetitive thought. Abstract, evaluative thinking (“why am I like this?”) tends to be unconstructive. Concrete, process-focused thinking (“how exactly did this unfold, and what could I do differently?”) can support problem solving and recovery. The same topic, thought about in a different mode, has different effects.

What further research adds

The shift from abstract to concrete thinking has been tested directly. Watkins, Baeyens and Read (2009) randomly allocated 60 people with dysphoria to concreteness training, to a bogus version matched for rationale, experimenter contact and duration but without active practice in concrete thinking, or to a waiting list. Concreteness training reduced depressive symptoms and increased concrete thinking more than both control conditions, and it reduced rumination more than the waiting list.

A later Phase II randomised controlled trial in primary care allocated 121 people with current major depression to treatment as usual, to treatment as usual plus concreteness training as guided self-help, or to treatment as usual plus relaxation training (Watkins et al., 2012). Adding concreteness training improved depressive symptoms after treatment and at 3- and 6-month follow-ups. It did not reduce symptoms more than relaxation training, although it reduced rumination and overgeneralisation more after treatment. The steps in this entry follow the same principle, but they are not the tested training programme.

Relation to catharsis

An intense cathartic experience can interrupt rumination for hours. The loop tends to return when the mode of thinking has not changed. Moving from abstract brooding to concrete steps is what turns interruption into change.

Interrupt rumination by changing how you think, not by thinking harder

2026-09-14T00:00:00+00:00

Why thinking harder does not help

Rumination is about the problem, so it feels like work on the problem. In three experiments, dysphoric students induced to ruminate on their feelings made more negative interpretations, were more pessimistic and produced less effective solutions to interpersonal problems than those induced to distract, who performed as well as nondysphoric participants (Lyubomirsky & Nolen-Hoeksema, 1995). In daily life, rumination and negative affect predicted each other from one moment to the next, a loop that can sustain itself (Moberly & Watkins, 2008).

What changes the loop

The mode of thought matters more than its topic. Watkins (2008) reviewed evidence that abstract, evaluative repetitive thought tends to be unconstructive, while concrete, process-focused thought can be constructive. In depressed patients, concrete rather than abstract self-focus improved social problem solving (Watkins & Moulds, 2005). Reflecting on an upsetting experience from a self-distanced perspective reduced emotional reactivity and rumination compared with immersed recall (Kross, Ayduk & Mischel, 2005).

Containing repetitive thought in time also has preliminary support: two weeks of a daily, time- and place-restricted 30-minute worry period reduced worry, anxiety, negative affect and insomnia more than a control condition in people with high trait worry (McGowan & Behar, 2013). In clinical settings, adding rumination-focused cognitive-behavioural therapy to usual care improved residual depression in a small trial (Watkins et al., 2011).

How this differs from brooding to next steps

The advice on turning brooding into next steps focuses on problem solving. This entry adds the recognition step, the observer perspective and the explanation of why rumination is so easily mistaken for processing. The evidence and its limits are in Research note 15.

Let anger cool instead of venting it

2026-09-14T00:00:00+00:00

Why venting backfires

The hydraulic model predicts that expressing anger discharges it. The experiments that tested this prediction found the opposite. Venting while thinking about a provocation rehearses the anger: it keeps anger-related thoughts active and makes aggressive responses more accessible. Physical exertion raises arousal, and when attention remains on the grievance, that arousal is experienced as more anger.

What the evidence shows

Bushman, Baumeister and Stack (1999) found that people who were told that hitting a punching bag relieves anger were more inclined to choose that activity, and that venting was followed by more aggression toward the person who had provoked them. Bushman (2002) compared hitting a punching bag while thinking about the provocateur, hitting it while thinking about fitness, and sitting quietly. The rumination condition produced the most anger and aggression, and sitting quietly produced the least.

Kross, Ayduk and Mischel (2005) showed that reflecting on an anger-provoking experience from a self-distanced perspective reduced emotional reactivity compared with reliving it from an immersed perspective. Distance allows the meaning of the event to be examined without re-igniting the emotion.

What further research adds

Rumination makes anger worse even without physical venting. Ray, Wilhelm and Gross (2008) asked participants in two studies to recall a recent anger-provoking event and then either ruminate on it or reappraise it. Compared with reappraisal, rumination led to greater anger experience, more cognitive perseveration and greater sympathetic nervous system activation. Denson’s (2013) review describes angry rumination as a risk factor for aggression that temporarily reduces self-control, which can increase aggression.

For people whose anger is a recurring problem, a meta-analysis of 23 controlled treatment studies in non-institutionalised adults with anger established by standardised measures found medium to large effects across therapies (Del Vecchio & O’Leary, 2004). Analyses by type of anger supported cognitive therapies for driving anger, anger suppression and trait anger, and relaxation for state anger. These are structured treatments delivered in studies, not self-help, which is why frequent or harmful anger remains a reason to speak with a professional.

In short

The sense of release after venting is real, but it is not evidence that anger has been reduced. Time, lower arousal and a distanced view reduce anger more reliably than discharge.

Mark a loss with a ritual that gives structure, not pressure to express

2026-09-14T00:00:00+00:00

Why it can help

Norton and Gino (2014) found in three experiments that people who reflected on rituals they had performed after past losses, or who completed a novel ritual after a loss, reported lower grief. The link was mediated by increased feelings of control, and it held for people who did not believe in rituals as well as for those who did. A loss removes control over something that mattered; a ritual, with its fixed sequence, gives back a small, deliberate act.

Hobson and colleagues (2018) describe rituals as regulating emotions, goal states and social connection, partly through their physical structure and partly through their meaning. A mourning ritual can do several things at once: mark that something has ended, gather people, and give a form to a relationship that continues in memory.

Why it need not be cathartic

The supported mechanism is control and structure, not discharge. The dual process model of bereavement (Stroebe & Schut, 1999) describes adaptive coping as an oscillation between confronting the loss and attending to ordinary life, including respite from grieving. That is why the steps above include stepping away afterwards. It is also why there is no reason to judge a ritual by how much emotion it produces.

What the evidence does not show

A Dutch questionnaire study found that bereaved people valued funerals and rituals, yet their evaluations and use of rituals were not significantly associated with grief reactions six months and three years later (Mitima-Verloop et al., 2021). Rituals can be experienced as helpful without measurably changing the course of grief. Most people adjust to bereavement without professional intervention (Stroebe, Schut & Stroebe, 2007), and psychological interventions show larger effects in trials that enrolled people with high grief symptoms and in trials of people at least six months after the loss (Johannsen et al., 2019). The evidence and the clinical boundary are set out in Research note 11.

Move to lift your mood, not to burn off anger

2026-09-14T00:00:00+00:00

Why it works for mood

In studies of everyday life, physical activity consistently predicted higher positive affect over the following few hours (Liao, Shonkoff & Dunton, 2015). Over weeks, randomised trials pooled in meta-analyses associate physical activity with lower depressive and anxiety symptoms: a medium effect on depression and a small effect on anxiety in non-clinical adults (Rebar et al., 2015), moderate effects for walking or jogging, yoga and strength training in major depression (Noetel et al., 2024), and medium effects across populations in an umbrella review (Singh et al., 2023).

Intensity matters for how activity feels. Pleasure tends to decline above the ventilatory or lactate threshold, while most people experience pleasant changes below it, and self-selected intensity appears to foster tolerance of harder effort (Ekkekakis, Parfitt & Petruzzello, 2011). Activity that feels tolerable is more likely to be repeated, and repetition is where the symptom benefits in trials came from.

Why it does not work for anger

Across 154 studies, arousal-decreasing activities reduced anger and aggression (g = -0.63), while arousal-increasing activities such as hitting a bag, jogging or cycling were ineffective overall (g = -0.02) (Kjærvik & Bushman, 2024). In an earlier experiment, hitting a punching bag while thinking about the provocateur produced the most anger and aggression (Bushman, 2002). Exertion raises arousal; if attention stays on the grievance, that arousal does not drain away.

The two findings are compatible because they measure different outcomes. Activity can lift energetic, positive feeling and still leave anger where it was. The reasoning is set out in Research note 12 and Research note 02.

Use shared rhythm for connection, and be clear about what it gives you

2026-09-14T00:00:00+00:00

Why it works

Moving in time with others coordinates attention and action. It produces a sense of being part of a group and, with sustained effort, physiological changes associated with bonding and analgesia. Music and exertion add reward and an acute lift in mood.

What the evidence shows

Synchronous tapping increased liking for a partner (Hove & Risen, 2009), and groups that walked, sang or moved in synchrony cooperated more afterwards than groups that did not (Wiltermuth & Heath, 2009). In group dance, synchrony and exertion each independently raised pain thresholds, an indirect marker of endorphin activity, and synchrony increased reported closeness to the group (Tarr et al., 2015). A meta-analysis found that single sessions of aerobic exercise reliably increase positive activated affect (Reed & Ones, 2006).

Outside the laboratory, a meta-analytic review of 50 studies found that reports of collective effervescence, intense shared emotion and a sense of unison during gatherings such as rituals, celebrations and demonstrations, were associated with group identity, collective positive emotions, well-being and collective efficacy (Pizarro et al., 2022). These are correlations based largely on self-report from people who chose to attend, so they describe what tends to go together rather than what attending causes (Research note 03).

What further research adds

A meta-analysis of 60 published and unpublished experiments that compared an interpersonal synchrony condition with at least one control condition found a medium effect of synchrony on prosociality, in both attitudes and behaviour (Rennung & Göritz, 2016). Experimenter effects and intentionality moderated these effects, so the size of the benefit depends on how synchrony is produced and studied.

Structured dance programmes have also been synthesised. A meta-analysis of 41 controlled intervention studies with 2,374 participants found a medium overall effect of dance movement therapy and dance interventions on health-related psychological outcomes (d = 0.60), with high heterogeneity (Koch et al., 2019). Larger effects came from observational measures, possibly indicating bias, and methodological shortcomings of many included studies limit the results; where follow-up data existed, most effects remained stable or slightly increased 22 weeks after the intervention. These are organised programmes, many in healthcare or preventive settings, not a night out dancing.

What it does not do

None of these studies measured whether personal problems were resolved. That is not a weakness of shared rhythm. It is a description of what it is for. Collective movement builds connection and restores energy, which are genuine resources for dealing with difficulties. The dealing itself usually happens through other means.

Put the feeling into precise words

2026-09-14T00:00:00+00:00

Why it works

Naming an emotion shifts processing from an immediate, bodily response toward a verbal, conceptual one. In the process model of emotion regulation, this changes how the situation and the response are represented, which can alter the emotion’s intensity and course (Gross, 1998).

What the evidence shows

Lieberman and colleagues (2007) asked participants to label the emotion shown in photographs of faces, to label another attribute such as gender, or simply to look. Labelling the emotion was associated with lower amygdala activity and higher activity in the right ventrolateral prefrontal cortex, and the two were inversely related. The authors describe affect labelling as a form of implicit emotion regulation: it helps even though people do not intend to regulate.

Naming also makes distance possible. Kross and Ayduk (2011) reviewed evidence that reflecting on negative experiences from a self-distanced perspective helps people make meaning of them without being overwhelmed. A precise label is often the first step in taking that perspective.

What further research adds

The effect of labelling is not limited to brain imaging. In four studies, self-reported distress was lower while participants labelled the emotion in negative pictures than while they passively watched them, and the reduction was similar to that produced by reappraisal and by distraction (Lieberman et al., 2011). Participants nonetheless predicted that labelling would increase their distress, even after experiencing the opposite. With positive pictures, labelling was associated with less self-reported pleasure, which suggests that it dampens affective responses in general rather than only relieving negative feelings. Torre and Lieberman (2018) review this work as evidence that affect labelling is a form of implicit emotion regulation that may not feel like regulation while it happens.

In a clinical application, spider-fearful people underwent a brief course of exposure to a live spider (Kircanski, Lieberman & Craske, 2012). One week later, with a different spider in another context, the group that had labelled their feelings during exposure showed a smaller skin conductance response than the groups that had reappraised, distracted themselves or received exposure alone, but it did not differ from them in self-reported fear. Greater use of anxiety and fear words during exposure was associated with greater reductions in fear responding.

In short

Before trying to release a feeling, give it a name. It costs a few seconds, requires no discharge, and changes what happens next.

Protect your sleep after an emotional strain instead of processing it all night

2026-09-14T00:00:00+00:00

Why it matters

Across 154 experiments with 5,717 healthy participants, all forms of sleep loss reduced positive affect and increased anxiety symptoms (Palmer et al., 2024). A second set of meta-analyses found that sleep loss increased negative mood and markedly decreased positive mood, and that sleep restriction modestly reduced adaptive emotion regulation in young people (Tomaso, Johnson & Nelson, 2021). After one night without sleep, healthy adults reported more stress, anxiety and anger than rested controls in response to a mild stressor, consistent with a lowered threshold for perceiving stress (Minkel et al., 2012).

This means that staying up late to resolve an emotional strain can work against the aim. The next day begins with fewer positive resources and a shorter fuse, and the strain is likely to feel larger.

What better sleep is associated with

Across 65 randomised trials, interventions that improved sleep also improved composite mental health, depression, anxiety and rumination, with greater sleep improvement linked to greater benefit (Scott et al., 2021). Over longer periods, insomnia was associated with about twice the odds of later depression (Baglioni et al., 2011), though that association cannot show direction.

Why the steps are modest

Individual sleep hygiene recommendations are associated with better sleep, but their direct effects in the general population remain largely untested (Irish et al., 2015). The steps above are therefore low-risk suggestions consistent with the evidence, not a validated protocol. Parking a concern with a concrete next step borrows from research on repetitive thought (Research note 15). The full evidence is in Research note 13.

Change the interpretation before trying to release the emotion

2026-09-14T00:00:00+00:00

Why it works

Gross’s process model of emotion regulation distinguishes strategies by the point at which they act on an emotion as it unfolds (Gross, 1998). Reappraisal acts early: it changes the meaning of the situation, and with it the emotional response the situation produces. Suppression acts late: it inhibits the outward expression of an emotion that has already been generated.

The difference has measurable consequences. Neuroimaging studies show that reappraisal recruits prefrontal and cingulate control systems that modulate activity in emotion-generating regions such as the amygdala (Ochsner & Gross, 2005). In everyday life, people who habitually reappraise report experiencing more positive and less negative emotion and have better interpersonal functioning and well-being than people who habitually suppress (Gross & John, 2003).

What the evidence shows

Webb, Miles and Sheeran (2012) meta-analysed experiments in which participants were instructed to use a specific regulation strategy. Strategies based on cognitive change, and reappraisal in particular, had a small-to-medium effect on emotional outcomes. Suppressing the expression of an emotion reduced its outward expression, but instructions to suppress the experience itself were not effective. Concentrating on one’s feelings did not improve emotional outcomes.

For catharsis, the implication is direct. The feeling that an emotion must be released assumes that its intensity is fixed. Reappraisal shows that intensity depends partly on interpretation, and that changing the interpretation can reduce the need for release altogether.

Reappraisal is a family of tactics

“Think about it differently” can sound like a single move. Uusberg, Ford, Uusberg and Gross (2023) propose that it is better understood as a family of tactics. In their expanded reAppraisal framework, reappraisal consists of shifts in appraisal that arise either from changing the mental construal of a situation (reconstrual) or from changing the goals used to evaluate it (repurposing). Both can target the situation itself (object level) or one’s own mental states, such as the emotion or thoughts about it (meta level). And both can work by loosening commitment to the dominant construal or goal (decommitment) or by building commitment to an alternative (commitment). Crossing these three distinctions gives a map of eight tactics, which the authors illustrate with examples from basic and clinical research.

The framework is a theoretical proposal, not a test of which tactic works best. It is practically useful because it widens the options. The steps above mostly use reconstrual of the situation: finding other interpretations of the same facts. When no alternative interpretation is credible, repurposing, asking which goal matters most now, or reappraising the emotion itself (“this anger is information, not an instruction”) are other routes within the same family.

When reappraisal is not the right tool

Reappraisal does not help everywhere. Troy, Shallcross and Mauss (2013) measured cognitive-reappraisal ability, the severity of recent life stressors, how controllable those stressors were and the level of depression in 170 participants. For participants with uncontrollable stress, higher reappraisal ability was associated with lower depression; for participants with controllable stress, it was associated with greater depression. The authors conclude that the adaptiveness of a strategy depends on the context. The data are correlational, but they are a reason not to use reappraisal in place of action when a situation can be changed.

Ford and Troy (2019) highlight two further drawbacks: people are often unable to use reappraisal successfully, and even when it succeeds, using reappraisal to feel better is not always functional. Whether it helps may depend on individual factors such as skill and on situational factors such as the intensity of the stressor.

What imaging studies add, and what they do not

Monachesi, Grecucci, Ahmadi Ghomroudi and Messina (2023) used activation likelihood estimation to meta-analyse fMRI studies of task-related activity during reappraisal and during acceptance. Both strategies were associated with increased activity in the left inferior frontal gyrus and the insula. Reappraisal was also associated with decreased activity in the basal ganglia, and acceptance with decreased activity in limbic regions. The authors suggest that emotion regulation may rely on a shared inhibitory circuit together with strategy-specific top-down and bottom-up processes, which would extend earlier dual-route models.

These are associations between tasks and patterns of measured activity across imaging studies. They are consistent with reappraisal and acceptance working partly through different processes, but they do not establish the mechanism, and they do not show that one strategy is better for a given person or problem. The practical reading is modest: acceptance is a distinct option rather than a failed attempt at reappraisal, which matters when a situation cannot be credibly reinterpreted.

Reflect from a distance instead of reliving

2026-09-14T00:00:00+00:00

Why it works

There are two ways of thinking about a painful memory. From an immersed perspective, a person re-enters the scene and re-experiences it. From a distanced perspective, the person views the scene from outside. Both involve thinking about the same event, but they have different consequences.

What the evidence shows

Kross, Ayduk and Mischel (2005) asked participants to recall an anger-provoking experience and to analyse their feelings from either an immersed or a distanced perspective. Distanced analysis reduced emotional reactivity and the tendency to recount the event in a rehearsing way, while still allowing participants to understand why they had felt as they did. Kross and Ayduk (2011) reviewed later studies that replicated and extended these effects to other negative emotions and outcomes.

The finding fits the distinction drawn by Watkins (2008) between unconstructive and constructive repetitive thought. Abstract, immersed brooding tends to maintain distress, while reflective, concrete processing can support recovery.

What further research adds

Two meta-analyses tested manipulations of psychological distance (k = 230) and of level of abstraction (k = 98) on emotional experience (Moran & Eyal, 2022). A distant perspective attenuated emotional experience (g = 0.52), with weaker effects for high-level emotions such as self-conscious emotions (g = 0.29) than for low-level emotions such as basic emotions (g = 0.64). Abstraction attenuated only low-level emotions (g = 0.2) and showed a reverse, non-significant effect for high-level emotions. Distance and abstraction are therefore not interchangeable.

Language offers a simple route to distance. Across seven studies with 585 participants, using non-first-person pronouns and one’s own name during introspection enhanced self-distancing (Kross et al., 2014). Under social stress, people who used this kind of self-talk performed better according to objective raters, displayed less distress and engaged in less maladaptive post-event processing, and none of these findings was moderated by trait social anxiety. A review of distancing research notes that its neurocognitive mechanisms remain unclear and that inconsistent terminology and methods make the literature difficult to synthesise (Powers & LaBar, 2019).

In short

Returning to a painful experience is not the problem. Re-living it is. Viewing it from a distance keeps the benefit of reflection without the cost of re-experiencing.

Write about difficult experiences to make sense of them, not to empty them

2026-09-14T00:00:00+00:00

Why it can help

Expressive writing gives an experience a structure. Putting events into sequence, naming causes and noticing what they mean turns a set of intrusive fragments into an account that can be thought about. This is different from catharsis in the hydraulic sense. The aim is not to discharge as much emotion as possible but to change how the experience is represented.

What the evidence shows

Pennebaker and Beall (1986) asked students to write for several consecutive days about a traumatic event or about superficial topics. Participants who wrote about both the facts of the event and their emotions about it made fewer visits to the health centre in the following months than control participants, although they reported more distress immediately after writing. Writing only about emotions, or only about facts, did not produce the same pattern.

Pennebaker (1997) reviewed the growing literature and highlighted evidence that benefits were associated with changes in language across sessions, particularly increasing use of words related to causation and insight. This suggests that cognitive processing, rather than emotional expression alone, carries the effect.

Research on talking about emotional experiences points in the same direction: talking is perceived as helpful, but the emotional impact of a memory is reduced when the conversation supports cognitive work on the experience rather than only re-expression (Zech & Rimé, 2005).

In short

Writing helps when it builds understanding. Its value is not in how much emotion leaves the page but in how the story of the experience changes.

Catharsis as a Service™

2026-09-14T00:00:00+00:00

A visual study that presents collective euphoria at first glance and a diagnosis of it at second.

Photography of an industrial rave, brutalist typography and a Prismatic telemetry overlay share one surface. The crowd gets its release, the system logs a successful response, and the underlying condition stays exactly where it was.

First glance, second glance

Seen from across a street, it is a campaign: a red-lit stage, a sea of raised hands, a headline big enough to promise something. That reading is intended and it is not a trap. The night it advertises is real, and so is the relief.

Seen up close, the poster starts filing a report. The overlay names an input, a process and an output. The status line reads as a success: 200 OK. Underneath it, one field refuses to cooperate: problem_solved: false.

Why an API

Service language is how we describe things meant to be consumed on demand, repeatably, with a predictable response. Putting catharsis behind an endpoint makes one uncomfortable property visible: the transaction can succeed completely without the request ever being resolved. The endpoint does exactly what it promises. The promise was just smaller than the need.

Relief is not resolution

The artifact does not argue that dancing is useless or that collective release is fake. Repetition, synchronization and discharge are old and effective human technology. The claim is narrower: a temporary state transition and a removed root cause are different outcomes, and it is worth being able to tell which one you received.

Why the telemetry is conceptual

Every metric on this page is an artistic value. Nothing is measured, collected or inferred about anyone who visits, and there is no analytics script to contradict that. A piece about instrumenting people should not quietly instrument its own audience.

How to use it

As a mirror: after the next night that felt like it fixed something, ask which field it actually changed.
As a conversation starter: the raw response is short enough to quote and specific enough to disagree with.
As a design reference: euphoria and diagnosis on one surface, with neither cancelling the other.

Closure as a Service™

2026-09-14T00:00:00+00:00

A visual study of closure sold on demand: one session, one letter, one ending. The poster makes the offer in full; the page records what the request actually returns.

Accepted, not completed

The endpoint answers 202 Accepted. In service language that means the request was received and will be processed, not that anything has finished. On the poster the ring stops just short of closing, the timer keeps counting, and one field stays open: grief_resolved: pending.

Why a ritual still matters

The work does not mock the letter that gets burned or the last conversation. A ritual gives a loss a form, a date and witnesses, and the sense of order it brings can be real and welcome. Research on rituals and on continuing bonds, discussed in the library, is consistent with a quieter reading: many people keep a relationship with the person or the life they lost, and that is not a failure to finish. A ritual can mark a threshold without being a finish line.

Why the telemetry is conceptual

Every metric on this page is an artistic value. Nothing measures, collects or infers anything about a visitor or their grief.

No deadline

Grief does not run on a service level agreement, and there is no correct date by which it should be over. If a loss keeps making everyday life unmanageable, or it stays this heavy for a long time, talking with people you trust, a doctor, a grief counsellor or a therapist can help. If you feel unsafe, contact local emergency services.

Case study: AI-assisted delivery under a supervisory control layer

2026-09-14T00:00:00+00:00

What this case study is (and is not)

This page documents one delivery: the site you are reading, built on 14 September 2026 by an AI coding agent (Claude Code) that one person directed, with Majordomus 0.6.0 as the supervisory control layer between the agent and the repository. It is written as a tech demo. Every number on it is computed from records the delivery left behind: the Majordomus ledger, the decision log and handovers of the working checkout, the Git history, and GitHub’s records of pipeline runs and releases. The timeline renders those records one by one, and each event links to the commit, run, release or file it came from.

What the page is not matters as much:

Not a controlled comparison. There was one session, one person and one agent, and no comparable session without supervision. No speed-up factor follows from these records, and none is claimed here. Measured durations are reported; estimates of time saved, with their assumptions labelled, are left to the time accounting article.
Not a general claim about Majordomus. It shows how one repository configured it, which commands ran, what they recorded and what they refused. Another configuration behaves differently.
Not a record of everything. The ledger stores accepted operations. A command that Majordomus refused leaves no entry, so friction is visible here only where a handover, a checkpoint or a commit mentions it.
Not a transcript. Prompts, checkpoint bodies and local paths stay in the checkout. The published snapshot is curated, and docs/CASE-STUDY.md describes what it contains and why.

The method note covers the first release in detail, including the kinds of instructions that drove the work. This page covers the supervision layer itself: how it was set up, how each part was used and where it stopped helping.

The setup

The agent and the person. The agent read the repository, ran commands, edited files, committed and pushed. In the first phases the person sent about fifteen messages: requests, art direction, long structured briefs and short directives such as “deploy”, “release automatically” and “use Majordomus and enforce it”. Implementation choices were the agent’s. Later in the day, further agents and a second session worked on the same checkout in parallel; the section on continuity describes how.

The repository. A Zola 0.23.6 site in English and Czech, styled with Tailwind CSS and enhanced with Alpine.js and Flowbite; Python validators, a Playwright suite and pytest; a small Rust workspace that reads the site’s JSON API; and GitHub Actions workflows that deploy to GitHub Pages and cut releases. The about page describes the site, and the status page shows the build that is running.

The Majordomus layer. Majordomus keeps its state in a directory named .ai/, which has two halves:

.ai/repo/ is tracked in Git and shared by every checkout. Its manifest registers the sections. The policy sets context budgets, the checkpoint interval of the default profile, the requirements of the finish contract (scope respected, verification ran, state updated, no open blockers, note present), the sections a handover must have (Objective, Current State, Next Action) and two enforcement entries that bind commands to Git hooks. Four execution profiles exist (implementation, debugging, deep work, routine), and every task here ran under implementation. Rules live under .ai/repo/rules/: a pinned, read-only vendored baseline and six project rules. Architecture decisions live under .ai/repo/adrs/.
.ai/local/ is ignored by Git. It holds the active task, the append-only ledger, checkpoints, handovers, the decision log and open questions. Nothing in it may be published by a generator, which is why this page ships a curated export instead of the directory.

From the policy, Majordomus generated three instruction files, AGENTS.md, CLAUDE.md and GEMINI.md. Each one points an agent at .ai/README.md and its discovery protocol. The first ledger event, at 10:26:23 UTC, is that projection.

Hooks and CI. Running npm ci points Git at the repository’s .githooks/ directory. The CI workflow repeats the Majordomus health check in a job of its own, so a clone without hooks is still checked. The wiring table lists every enforcement point.

The task lifecycle in practice

Majordomus allows one active task per checkout. The lifecycle diagram above walks through each step with its registered command and a real recorded example; this section tells the same story in order.

Start with a scope. majordomus start opens a task with a title, a profile and the paths it may touch. The first task started at 10:26:23 UTC and claimed 20 paths. Six tasks were started in the recorded period, with scopes from one path to 31.

Work, then check. majordomus check is read-only: it reports whether the active task is consistent with the policy, its scope and the recorded state. majordomus watch reports drift between the task, its scope, the policy and the generated instruction files. Between them they caught two scope problems before the first push (see what supervision caught).

Checkpoint. majordomus checkpoint records a short progress note, capped at 40 lines by the policy. Seven checkpoints were recorded. The agent wrote four of them; the checkpoint at 10:43:27, for example, was written right after the first push, while the pipeline was running. Two were derived from Git without any authored text and state only how many files had changed since the task started: 65 at 12:16:46 and 144 at 12:52:41. The last one belongs to the task that was still active at the end of the snapshot.

Record decisions. majordomus decision appends a dated entry with the rationale, the rejected alternatives and the evidence. Four decisions were recorded; they are listed under decisions and ADRs below.

Hand over. majordomus handover writes a continuation record with the three sections the policy requires. Four handovers were written. Three of them came one or two seconds before a task was finished as partial, which is the pattern the finish contract asks for: an outcome other than completed needs a note that says what comes next.

Finish with a verification command. majordomus finish evaluates the finish contract. An outcome of completed requires a verification command that ran and exited zero, and the ledger stores the command, its exit code and its duration. For this repository the agent chose the production smoke test, the script behind npm run smoke, run against the live URL. The first completed task closed at 10:47:25 UTC, less than two minutes after release v0.1.0 was published, and its verification command exited 0 after 7 seconds. Its contract recorded seven requirements as pass, four as skipped and none as failed; requirements that did not apply to the outcome or the configuration are recorded as skipped. The second completed task closed the same way at 13:08:34, after release v0.5.0.

Read it back. majordomus history prints the ledger. The snapshot behind this page was exported from its JSON output.

Enforcement wiring

A rule that nothing runs is a suggestion. Majordomus reconciles the enforcement entries of its policy: majordomus doctor fails when a declared command is missing, is not executable, is not invoked by the hook the policy names, or is invoked in a way that swallows its exit code. The vendored rule behind this is enforcement-wiring, and the doctor enforces it.

Pre-commit. .githooks/pre-commit refuses to run when Majordomus is not installed and then runs majordomus doctor. A commit therefore requires a healthy, wired layer.
Pre-push. .githooks/pre-push runs majordomus finish --check, which evaluates the active task’s finish contract without closing the task. A push that carries work outside the claimed scope is refused before it leaves the machine.
Commit messages. .githooks/commit-msg runs scripts/lint-commits.sh. Conventional Commits are enforced because release versions are computed from them.
The CI supervision job. The Majordomus supervision job in ci.yml installs Majordomus at the pinned version 0.6.0 (the installer verifies the SHA-256 of the release), points Git at the hooks and runs majordomus doctor. It passed in every Pages run in the snapshot.
Deployment gates. pages.yml runs CI, then Deploy, then Verify production, then Release. Deploy asserts that the Pages URL equals the base URL in the Zola configuration. Verify production waits until the live site serves the pushed commit, then runs the smoke test and the browser suite against it with npm run test:production. Release runs only after Verify production has passed.
Repository rulesets. Three rulesets protect the repository on GitHub: main requires CI, the history of main cannot be rewritten or deleted, and v* tags cannot be moved or deleted. The CI requirement allows an admin bypass, and the first push used it.

Only the two hooks are Majordomus enforcement in the strict sense. The rest are conventional safeguards of the same repository, and this account does not separate their effects from those of Majordomus.

Rules next to the code

Rules are Markdown documents with front matter: an id, a version, a class (blocking or advisory) and optional dependencies written as exact id@version references. The effective set is additive, the vendored Majordomus baseline plus the project’s own rules, and there is no override mechanism. majordomus rules list prints the set with each rule’s class and the commands that enforce it.

The six project rules, each linked to its file:

verified-deployment: nothing is reported as deployed until the verify job of the Pages workflow has passed against the live URL that GitHub returns.
zola-source-of-truth: Zola generates all public HTML; content lives in Markdown, interface strings in translation dictionaries, and templates are shared by both languages.
release-on-verified-main: Conventional Commits are mandatory, and every verified push to main produces a release whose version is derived from the commits.
evidence-ledger: every public claim is recorded with its sources, evidence level, confidence and a review date that has not passed.
no-neuro-overreach: no sentence may claim that a signal, a model or a brain region reveals what a person feels.
medical-claims: findings are stated at group level with their limits, nobody is assessed individually, and every research note and advice entry carries the clinical disclaimer.

All six are blocking, and their dependencies are explicit: evidence-ledger depends on verified-deployment, no-neuro-overreach on evidence-ledger, and medical-claims on both. None of them has an x-majordomus block, and the rule listing reports them as having no validator. That is deliberate and written in each file: Majordomus resolves them, lists them and loads them into an agent’s context, while the project’s validators, CI jobs and review enforce them. The distinction matters when reading this case study. When a prohibited phrase or an expired claim fails the build, the project’s content validator or ledger validator caught it, not Majordomus.

Majordomus enforcement lives in the vendored baseline. Four of its rules did visible work in this delivery. Scope-integrity treats work outside the claimed paths as not belonging to the task; check, finish and watch enforce it. Verification-integrity requires a completed outcome to have a verification command that exited zero, with its exit code and duration recorded; finish enforces it. Note-integrity requires every outcome to carry a note with the sections that outcome needs; finish enforces it too. Enforcement-wiring, described above, is enforced by the doctor.

Decisions and ADRs

Two mechanisms separate a working decision from a durable one.

The decision log is local, append-only and cheap to write. majordomus decision records what was decided, why, what was rejected and where the evidence is. Four entries were recorded:

Structured data is built as Tera data and serialised with json_encode. Hand-written JSON-LD inside templates is HTML-escaped and breaks on quotes. Rejected: hand-written JSON-LD strings, and a script that injects them in the browser.
Link previews are localized per language, and the sitemap lists every language edition. Rejected: one English preview card for every page, and language alternates only in the HTML.
Python validators are tested with pytest in a project virtual environment. Rejected: the standard library’s unittest, and a global pytest installation.
Evidence is tracked as a dated ledger of appraised sources and graded claims, with validators that fail the build on stale or inconsistent entries. Rejected: free-text evidence notes on each page, and a single file that mixes bibliographic facts with editorial judgement.

The first three were recorded within one second of each other (11:08:30 to 11:08:31 UTC), shortly before a handover. Decisions were written down when a task was about to close, not continuously while it ran.

The architecture decision record is tracked and durable. The fourth decision was written up as ADR-0001 with its context, the decision, the rejected alternatives and the consequences, and the ledger records it as proposed at 12:55:12. The agent did not accept it, because acceptance is a person’s act. The last handover of that task names accepting or amending ADR-0001 as the next action, and majordomus adr list still shows it as proposed. The decision itself is visible on this site: the evidence page and the methods page render grades from the ledger it describes.

What supervision caught

The findings on this page are generated from the records, not written by hand. The export script declares what to look for and publishes a finding only when the ledger, a handover, a checkpoint or the pipeline data confirms it. Each card links to its evidence in the timeline. In order:

Scope drift after a file move. The first task claimed 20 paths but not the root files that were moved into the artwork and static directories. The handover records that majordomus watch reported the drift. The task was finished as partial 2 minutes 53 seconds after it started and restarted with a scope that named those files.
A scope of “.” before the first push. The restarted task declared a single dot as its scope. Majordomus does not read that as the whole repository, so majordomus check reported every changed file as out of scope. The task was restarted at 10:39:39 with 26 explicit paths. The first push came about three minutes later, and its pipeline passed all seven jobs.
A brief that outgrew its scope. When the work expanded to a Rust workspace, data files and Python tooling, the handover recorded that the new top-level paths fell outside the claimed scope. A new task claimed 31 paths.
A ruleset bypass. The checkpoint written after the first push records that the push went through the admin bypass of the ruleset that requires CI on main. Deployment itself stayed gated by the job dependencies of the workflow. This was recorded, not prevented.
Completion only with a live check. Both tasks closed as completed ran the production smoke test as their verification command, and both exited 0. The three partial tasks claimed no completion.
A failed reference check stopped a deployment. CI caught this one, not Majordomus. Run 34842607207 failed in the References (Crossref) job, and Deploy, Verify production and Release were skipped. A fix to title matching (commit 94d734c) followed, and the next run published v0.3.0 at 12:24:28, seven minutes after the failing run had started.
A failed production verification. Also CI. Run 34860959250 deployed and then failed in Verify production, when one of 84 browser tests timed out against the live site; Release was skipped. Because live verification needs a deployment, the unverified build stayed online until the next run. A follow-up commit made that test deterministic, and the next run, for a later commit that included it, passed and published v0.6.0.
An ADR left to a person. Described under decisions and ADRs above.

Three events mentioned in other accounts of this delivery are not among the findings, because the records do not confirm them: a pre-commit health check that flagged the README for not linking the agent bootstrap, a finish command refused until a handover existed, and a rule dependency the resolver rejected until it was written as an id@version reference. Refusals are not written to the ledger, and none of the three appears in a handover, a checkpoint or a commit. They may have happened; this page does not count them.

Continuity across restarts and parallel workers

Restarting a task was cheap in recorded time. Between a handover, the partial finish and the next start, the ledger never shows more than two seconds. Six tasks were started and three of them were restarts. What made restarts cheap is that the state lived in files rather than in the agent’s conversation. A handover carries the objective, the current state and the next action, together with fields computed from Git (branch, commit, clean or dirty tree, changed files) that an author may not write. The policy also has the session hooks write a briefing when an episode starts, a derived checkpoint when the conversation is compacted, and a continuation record when an episode ends while a task is still active.

Later in the day the delivery tested a harder case: several workers on one checkout at the same time. At the end of the snapshot one task is active. It started at 14:54:07 UTC and claims 31 paths, which is the whole working tree. Content agents, a second Claude session and the main session worked inside it at once, each told which files it owned. This page was written by one of those workers, which was told not to start, checkpoint, hand over or finish tasks and not to edit files that other workers held.

That arrangement exposes a limit. Majordomus scopes work per task and per checkout. Inside one checkout, parallel workers share the active task, and Majordomus cannot tell which worker changed which file. Coordination came from explicit path ownership in each worker’s instructions, and from Git. The vendored baseline states the Majordomus answer as principles: one worker, one scope, with a scope claimed by another worktree treated as a boundary, and parallel work isolated from each other. This delivery did not apply them inside one checkout.

Evidence of delivery

Everything in this section is linked, and the timeline can be filtered to show only commits, pipeline runs or releases.

Commits. Every commit subject is a Conventional Commit. The first, the poster, is dated 09:54:05 UTC, one second before the GitHub repository was created; the site followed at 10:42:45.
Pipeline runs. The first Pages run started at 10:43:02 and finished at 10:45:45, with seven jobs that all passed. Later runs grew to ten jobs as Python tests, a Rust job and the Crossref check joined CI. Two of the first seven runs failed, and both failures are shown with what followed them.
Releases. Release v0.1.0 was published at 10:45:42, after live verification: 51 minutes 37 seconds after the first commit and 19 minutes 19 seconds after Majordomus was initialised. Releases v0.2.0 to v0.5.0 followed by 13:07:10, each cut by the Release job of a run whose production verification had passed.
Completed tasks. Both completed tasks store their verification command, exit code 0 and duration in the ledger.

Two loops can be measured directly. From the site commit to the published v0.1.0 took 2 minutes 57 seconds. From the start of the run that failed on references to the published v0.3.0 took 7 minutes 9 seconds. The releases page and the runs of the Pages workflow are the primary records, and gh run list reads the same data.

What it cost and what it did not solve

Friction. Three of six tasks ended as partial and were restarted because of how their scope was declared. Two of those restarts happened within the first fourteen minutes of supervision. A scope has to be written as explicit top-level paths, and in this delivery a grown brief meant closing the task and opening a new one, which matches the baseline principle of claiming paths when a task starts instead of widening the scope silently. The records themselves cost attention: seven checkpoints, four handovers and four decisions between 10:26 and 13:08 UTC, plus the task that followed. And because the finish contract refuses a completed outcome without a passing verification command, a task could not be closed as completed before the pipeline had deployed and the live site answered.

What Majordomus did not check.

Whether content is true, well sourced or safely worded. The project’s validators and a reviewer own that.
Whether the verification command is adequate. The contract checks that the command ran and exited zero; choosing a live smoke test was the agent’s decision.
Whether tests are meaningful, or whether a failing test is flaky or real.
Which worker did what inside one checkout shared by parallel workers.
The six project rules, which it lists but does not enforce.

What the records cannot show. Refused commands are not in the ledger. The time a person spent reading, deciding and writing briefs is recorded nowhere. Durations depend on hardware, the network and the CI queue. And because conventional safeguards (local validation, browser tests, CI gates and rulesets) ran alongside Majordomus, their effects cannot be separated here.

What it did change, as far as the records show. Every claim of completion in this project is tied to a recorded command and its exit code. Scope mistakes surfaced locally, before any push, instead of as failed pipeline runs. And the state needed to resume work sat in files that a new worker could read, which is also what made this case study possible.

How to reproduce

The records are in the public repository, except .ai/local/, which never leaves a checkout. To inspect them in your own clone:

Install Majordomus from majordomus.dev and run npm ci so that the hooks are active.
Check the layer with majordomus doctor, list the effective rules with majordomus rules list and the architecture decisions with majordomus adr list.
Read the pipeline history with gh run list and the task lifecycle in .ai/repo/workflows/task-lifecycle.md.
In a checkout that has its own ledger, majordomus history prints it, and python3 scripts/export-case-study.py regenerates the snapshot behind this page. With the check option, the same script validates the committed snapshot without local state or network access, which is how CI treats it.

The snapshot format, what it includes and what it leaves out are documented in docs/CASE-STUDY.md. For how the rest of the site was engineered, read the engineering articles.

Detection sandbox: what signals do not settle

2026-09-14T00:00:00+00:00

Why a sandbox rather than a demonstration of accuracy

A system that reports an emotional state has done two separable things: it recorded something, and it concluded something. The first is engineering and the second is inference, and the gap between them is where almost every overclaim about emotion recognition lives.

This page takes that gap apart. The signals are yours to set, so nothing about the reader is involved. The features are computed by rules printed beside their outputs, so nothing is hidden in a model. And the readings are shown all at once, ranked but never reduced to one, because the honest output of this kind of inference is a spread rather than a verdict.

The demonstration

Set the signals to something a concert would plausibly produce: a high heart rate, high movement, high alignment with the people nearby. Read the ranked list. Then change only the setting to an emergency. The signals have not moved by a single point, and the leading reading changes.

Nothing about the body changed. What changed is the prior that comes from knowing where the body is, and that prior is not available to any sensor in the list. A system deployed in a workplace, a school or a stadium inherits the same problem and usually hides it, because a single label reads as an answer while a ranked spread reads as an admission.

What the research says about the real version of this

The sandbox is a caricature, and the underlying findings are not kinder to detection than the caricature is. A large review concluded that facial movements are not reliable, specific indicators of emotional states across people and contexts (Barrett et al., 2019). A meta-analysis of 202 studies found that autonomic responses did not clearly distinguish one emotion category from another (Siegel et al., 2018). There is no gold-standard measure to fall back on: experiential, physiological and behavioural measures converge only modestly and cannot be treated as interchangeable (Mauss & Robinson, 2009).

Systems built on those signals therefore rest on particular conceptual models of emotion and on proxy data, and those choices shape what the systems imply about people (Stark & Hoey, 2021). When repeated physiological sampling is done carefully in daily life, the patterns that recur within one person vary in number and in the labels attached to them (Hoemann et al., 2020).

What to take from it

An inference can be honest about three things at once: what it saw, what it computed, and what it still does not know. The sandbox prints all three, and the third list is the longest. That is not a failure of the demonstration. It is the result.

The evidence ledger as code: claims, sources and review dates as validated data

2026-09-14T00:00:00+00:00

The problem a bibliography does not solve

The first versions of the library cited sources in the usual way: a list of references per page, each with a DOI, and a validator that compared every DOI, title and year with Crossref. That catches invented or mistyped references. It does not answer the questions a careful reader of a psychology site actually has:

Does the cited study support this sentence, or a nearby one?
How strong is the support compared with the claim on the next page?
Who was studied, and does the sentence stay within that population?
When did anyone last check that the sentence still reflects the literature?

These are editorial judgements, and they age. The decision record that introduced the ledger gives a concrete example: a 2024 meta-analysis on arousal and anger changed how the site’s statement about venting should be worded. A claim that nobody revisits does not become visibly wrong; it just stops being true. The design goal was to make that ageing visible and to make the build fail on it.

The decision was recorded in Majordomus at 12:26 UTC on 14 September 2026 and written up as ADR-0001, “Evidence ledger with dated review for every public claim”. It shipped in commit 4cbb6fb and release v0.4.0.

Three files, three reasons to change

File	Holds	Changes when	Checked by
data/references.toml	authors, year, title, container, DOI or URL	a citation is added or corrected	Crossref, offline and online
data/sources.toml	design, level A–E, population, appraisal in both languages, review dates	a source is re-read or re-graded	the ledger validator
data/claims.toml	statement, type, level, confidence, sources, scope, caveat, pages, review dates	the literature or the wording changes	the ledger validator

A fourth file, data/evidence_changelog.toml, records what changed, newest first. The ADR explains why appraisal is not a field in the bibliography: that “mixes facts Crossref can verify with judgements it cannot, and a single file would change for two unrelated reasons.”

A claim entry looks like this (abridged from the ledger):

[[claims]]
id = "expressive-writing-coherence"
claim_type = "empirical"
evidence_level = "B"
confidence = "MODERATE"
sources = ["pennebaker-beall-1986", "pennebaker-1997"]
used_in = ["research/what-is-catharsis/index.md", "research/crying-and-sharing/index.md",
           "advice/write-to-make-sense/index.md"]
last_reviewed = 2026-09-14
review_due = 2027-09-14

The omitted fields are the ones readers see: statement, scope and caveat, each an inline table with en and cs values. Here the caveat says that effects vary across studies, that short-term distress often rises after writing, and that “the 1986 abstract could not be re-verified”. That last sentence is typical of the ledger’s purpose. It does not hide a weak point in the evidence; it records it next to the claim.

Vocabulary: level and confidence are different things

The ledger uses two scales, and keeping them apart was deliberate.

Evidence level (A to E) describes the design of the best available support. Level A is reserved for syntheses of controlled studies; syntheses of correlational studies are graded by the design of the studies they pool, which is why some meta-analyses sit at D.
Confidence (HIGH, MODERATE, LOW, UNKNOWN) describes how certain the site is that the statement, as worded and scoped, is supported. A historical statement about Aristotle can be level E and HIGH confidence; a laboratory finding can be level B and MODERATE because the studies disagree.

The changelog has its own small vocabulary: added, revised, downgraded, upgraded, retracted and reviewed. A downgrade or a retraction is therefore a visible, dated event with a summary in both languages, not a silent edit to a number, and it lists the ids of the claims it touched.

Claim types set the review policy: empirical, theoretical, historical, clinical-boundary, technical and artistic. The artwork’s invented telemetry is recorded as an artistic claim, so the ledger is also where the site states which of its own numbers are not measurements.

What the validator enforces

python3 scripts/validate-claims.py is 255 lines of standard-library Python. Its header lists the failures, and each corresponds to a way the ledger could quietly lose its meaning:

unknown vocabulary (a level, confidence, type or design outside the lists) and missing translations;
dangling ids and paths: a claim citing a source that has no appraisal, or a used_in page that does not exist;
a claim graded stronger than its best source;
a review interval longer than policy allows;
a research note, theory essay or advice entry that no claim covers;
any entry whose review date has passed.

The intervals are data in the script, not prose in a document:

CLAIM_INTERVALS = {
    "clinical-boundary": 180, "empirical": 365, "theoretical": 730,
    "historical": 730, "technical": 730, "artistic": 730,
}
SOURCE_INTERVALS = {"classical": 730, "book": 730, "software": 730, "record": 730}
SOURCE_INTERVAL_DEFAULT = 365

Two design choices keep this honest. First, the build can fail on a date alone. The ADR names that as a consequence, not a bug: “moving a date requires a review, recorded in the changelog.” Second, the date is injectable. The script accepts --today, and the test suite uses it to show that the repository’s own ledger expires:

def test_repository_ledger_expires_without_review():
    result = run([PY, "scripts/validate-claims.py", "--today", "2031-01-01"])
    assert result.returncode == 1
    assert "(stale)" in result.stdout

The other tests in tests/python/test_evidence.py build a minimal ledger in a temporary directory and break it one way at a time: a level A claim on a level C source must fail with “stronger than its best source (C)”; a clinical-boundary claim with a one-year interval must fail with “the maximum is 180”; a research note that no claim covers must fail. A validator that has never been seen to fail is not evidence of anything, so every rule has a test that makes it fail.

A weekly check that does not wait for a push

A ledger that is only validated on push expires silently whenever nobody pushes. evidence-freshness.yml runs every Monday at 06:17 UTC, independent of commits:

- name: Claims, sources and glossary are consistent and within their review dates
  run: python3 scripts/validate-claims.py --warn-days 30 | tee claims.txt
- name: References still resolve in Crossref with matching titles and years
  if: always()
  run: python3 scripts/check-references.py --online | tee references.txt

When either step fails, the workflow opens an issue labelled evidence-update, or comments on the open one, with the error lines and instructions that end: “Dates move only after a review.” The workflow was run once by hand on the day it was added (run 34846806810, 38 seconds, passed). The same Crossref comparison, run by python3 scripts/check-references.py --online, is also a CI job, and it produced the first pipeline failure of the day; see Testing a static site.

Pages render grades; they do not restate them

The rule in the evidence component’s header is short: “Pages never restate a claim’s grade by hand; they render it from the ledger.” The essay template computes the page’s content path and asks for every claim whose used_in names it:

{%- set ledger = load_data(path="data/claims.toml") -%}
{%- set_global found = [] -%}
{%- for c in ledger.claims %}{% if path in c.used_in %}{% set_global found = [...found, c] %}{% endif %}{% endfor -%}

Each claim renders as a card with a kind label, a level badge and a confidence badge, the statement, scope and caveat in the page’s language, the review dates, and a button that opens the appraised sources. The button is an Alpine disclosure; without JavaScript the panel stays open, and a browser test checks both states. The evidence page lists the whole ledger, and the status page computes counts, the most recent review and the next reviews due at build time. Its template says: “Nothing here is typed by hand.”

Because the ledger is data, it is also exported. The JSON API includes claims.json, sources.json, glossary.json and evidence_changelog.json, and the Rust crate re-checks two invariants independently of the Python validator: every reference has an appraisal, and no claim is graded above its best source. The CLI can list claims due before a date with cargo run -p caas-cli, which is how another application could refuse to show a stale claim.

What it cost and what it does not do

At v0.5.0 the ledger held 50 claims, 58 appraised sources and 54 glossary terms, all written in both languages. The ADR is explicit about the price: “Every new substantive claim costs a ledger entry in both languages; content without one fails CI.” The parallel content agents described in Parallel agents, one working tree delivered their ledger entries (references, appraisals, claims, glossary terms and changelog lines) as proposal files, which had to validate before their pages could merge.

The validator cannot judge whether an appraisal is right. It can check that a level exists, that it does not exceed the sources, and that someone committed to a review date. Whether a study really supports a sentence remains reviewer-owned, and the project rule project.evidence-ledger says so in its verification section. The grading scale is coarse and partly a judgement. The ledger does not make the site’s claims correct. It makes their strength, scope and age inspectable, and it makes neglect fail the build instead of going unnoticed.

How the site was built: Zola, Tera components, Tailwind and a content API

2026-09-14T00:00:00+00:00

What the build has to deliver

Catharsis as a Service™ is two things at once: an artwork (a poster and a mock “emotional state transition endpoint”) and a bilingual evidence library about catharsis. The library makes psychological claims to a general audience, so the engineering had four requirements from the start:

Every page in English and Czech, with identical structure, so neither edition can quietly fall behind.
Every claim traceable to a dated, graded ledger entry and to sources checked against Crossref.
Readable without JavaScript, fast on a phone, and correct under the repository subpath that GitHub Pages serves (/catharsis-as-a-service/).
Nothing reported as done until it is live: deployment is complete only after the production URL has been tested.

All of it was built on 14 September 2026 in one working session, by an AI coding agent directed by one person. The case study shows the timeline; Supervised AI delivery describes how the work was directed. This article covers the result.

Source layout: content, data and strings

The repository separates four kinds of source, and each has one home:

Concern	Lives in	Read by
Pages and landing blocks	`content/**/index.md` and `index.cs.md`	Zola
Structured records	`data/*.toml`: references, source appraisals, claims, glossary, command registry	Zola templates, validators, API exporter
UI strings	`[translations]` in zola.toml	`trans()` in templates
Presentation	`templates/` and styles/app.css	Zola, Tailwind

The rule behind the split is recorded in the repository as project.zola-source-of-truth: “the next artifact must be a Markdown file pair, not a template edit”. Templates never branch on language; they call trans(key=…, lang=lang), and the translation validator fails the build when a key exists in one dictionary but not the other, or when two editions of a page have a different front-matter shape.

The data files are the part most static sites do not have. A research note cites registry ids such as kjaervik-bushman-2024; the bibliography entry, its appraisal and the claims that rely on it live in data/. The template resolves ids at build time:

{%- set registry = load_data(path="data/references.toml") -%}
...
{%- set ref = registry.references[id] %}

Keeping bibliographic facts in data rather than in page bodies means one correction reaches every page that cites the source, and a validator can compare the whole registry with Crossref in one pass. The ledger design is described in The evidence ledger as code.

Zola 0.23.6 and Tera v2 components

Zola was chosen because it is a single static binary with built-in multilingual content, taxonomies, feeds, a sitemap and an internal link checker. There is no Node or Ruby runtime between Markdown and HTML, and the version is pinned in .zola-version; CI downloads that release and checks its SHA-256 digest before using it.

Version 0.23 ships Tera v2, which replaces macros with components: named, typed template functions with a JSX-like call syntax. The first checkpoint in the Majordomus record for this project notes the port explicitly (“Tera v2 port: macros replaced by components”). A component declares its parameters and their types:

{% component library.tags(tags: array, lang: string) %}
{%- if tags | length > 0 %}
<ul class="tag-list" aria-label="{{ trans(key='nav_tags', lang=lang) }}">

and a page template calls it like an element:

{{ <library.cards pages={section.pages} lang={lang} /> }}

Typed parameters matter more than they look. A missing or misspelled argument fails at build time instead of rendering an empty string, which is the failure mode that makes template bugs invisible in static sites. The components are grouped by feature rather than by HTML element: library (cards, key points, references, figures), evidence (kind labels, level and confidence badges, claim cards), discover (glossary terms and related reading), models, landing, artifact and posters. The page you are reading is rendered by templates/essay.html, the same template the research and theory essays use; this section only sets kind = "technical" so the page carries the TECHNICAL label instead of EMPIRICAL.

Styling: Tailwind CSS 4 and Flowbite, self-hosted

Tailwind 4 is configured in CSS rather than in a JavaScript config file. The entry point declares exactly which files the class scanner reads, loads the Flowbite plugin and defines the design tokens:

@import 'tailwindcss' source(none);

@source '../templates';
@source '../static/js/app.js';
@source '../node_modules/flowbite/lib/esm/components';

@plugin 'flowbite/plugin';

@theme {
  --color-ink: #050505;
  --color-bone: #e9e4d8;
  --color-blood: #ef2929;

source(none) turns off automatic detection, so a class only ships if it appears in a template, in the one script that toggles classes, or in the Flowbite components the markup uses. The tokens (ink, bone, blood; Anton, Barlow Condensed and JetBrains Mono) come from the poster, so the library and the artwork read as one object.

No asset is loaded from a CDN. Fonts are self-hosted with their OFL licences, and a 17-line script, scripts/vendor.mjs, copies the pinned Alpine and Flowbite bundles from node_modules into static/js/vendor/. The site therefore makes no third-party requests, which is also what lets the evidence ledger state that there are no analytics, trackers or cookies.

Behaviour: Alpine and Flowbite, never owners of content

Two small libraries provide interaction, with a strict division of labour written into the architecture rule: “One owner per piece of DOM state: Alpine or Flowbite, never both.”

Flowbite supplies mechanics that are easy to get wrong for accessibility: the modal, the drawer, tooltips and the accordion.
Alpine.js supplies local state: the view switch on the artifact page, the copy button, claim panels that open their sources, the interactive models and the search page.

Scripts are deferred and ordered so that page scripts register their Alpine components before Alpine starts:

<script src="{{ get_url(path='js/vendor/flowbite.min.js', cachebust=true) }}" defer></script>
<script src="{{ get_url(path='js/app.js', cachebust=true) }}" defer></script>
<script src="{{ get_url(path='js/vendor/alpine.min.js', cachebust=true) }}" defer></script>

Every piece of content is in the HTML. A claim card’s sources panel, for example, is hidden by an Alpine binding, and a .no-js rule keeps it open when scripts do not run. The browser suite loads pages with JavaScript disabled and checks that this holds.

The subpath problem

GitHub Pages serves a project site under /<repository>/. A hard-coded /assets/poster.png works on a developer machine and breaks in production. The project handles this in three places:

base_url in zola.toml is copied from the GitHub API, and a comment says so. The deploy job compares it with the URL that GitHub’s configure-pages action returns and stops if they differ.
Templates build every URL through get_url, resize_image or a permalink, never by string concatenation.
npm run preview serves the production build under http://127.0.0.1:4173/catharsis-as-a-service/, and the browser suite runs against that address, so a base-path mistake fails before a push rather than after it.

The content API and the Rust client

The library is also published as data. python3 scripts/export-api.py writes api/v1/ next to the site: an index, research and advice collections per language, references, the claims ledger, source appraisals, the glossary, the evidence changelog, the command registry and a search index per language. The exporter’s docstring states the constraint that makes it testable: “Output is deterministic: no timestamps, stable ordering.” A Python test exports twice and compares the bytes.

The Rust workspace in crates/ is the consumer. caas-content deserializes the export into types and checks the invariants other applications depend on; caas-cli is a command-line client:

let library = Library::load("public/api/v1")?;
assert!(library.problems().is_empty());

for claim in library.claims_due_by("2027-03-31") {
    println!("{} {} {}", claim.id, claim.claim_type, claim.review_due);
}

The point of a second language is independence. The Python exporter and the Rust types are written separately, so a field renamed on one side fails the contract tests on the other. CI exports the API of the commit being tested and runs cargo test --workspace against it, next to cargo clippy with warnings treated as errors. The CLI is runnable as cargo run -p caas-cli; its claims list --due-by command is a practical way to see which claims need review before a given date. The contract is documented in docs/RUST-INTEGRATION.md.

The pipeline end to end

npm run build is the short path: vendor the bundles, compile Tailwind, run Zola, export the API. npm run validate is the path that matters, a single Bash script, scripts/validate.sh, whose steps stop at the first failure:

step toolchain zola_version
step assets npm run --silent assets
step javascript js_syntax
step i18n python3 scripts/validate-i18n.py
step content python3 scripts/validate-content.py
step references python3 scripts/check-references.py
step claims python3 scripts/validate-claims.py
step social python3 scripts/render-social.py --check
step zola-check zola check --skip-external-links
step zola-build zola build --base-url "$BASE_URL"
step api python3 scripts/export-api.py --out public --base-url "$BASE_URL"
step html python3 scripts/validate-html.py public "$BASE_URL"

The first release ran seven of these stages; the others were added as the library grew (the history of validate.sh on GitHub shows when). Stages are ordered from cheapest to most expensive, so a missing translation fails in a second, before a full build.

On GitHub, pages.yml chains four stages, each gated on the previous one:

CI, reused from ci.yml so that the gate before a deployment is the gate before a merge: conventional commits, Majordomus supervision, validate and build, Python tests, Rust, references against Crossref, and the browser suite on the subpath preview.
Deploy uploads the site artifact produced by the validate job. The deployment is not a second build, so the files that were checked are the files that ship.
Verify production waits until the live site serves the commit being deployed, then runs npm run smoke and npm run test:production against the real URL.
Release publishes a GitHub Release. The version is computed from the Conventional Commit subjects since the last tag, and the notes link the pipeline run; scripts/release.sh --dry-run prints the same computation locally without publishing.

The wait in step 3 relies on the footer, which renders the commit SHA from the build environment:

for attempt in $(seq 1 40); do
  if curl -fsS "${BASE_URL}?revision=${GITHUB_SHA}" | grep -q "$GITHUB_SHA"; then
    echo "production serves $GITHUB_SHA (attempt $attempt)"
    exit 0
  fi
  sleep 15
done

Without it, the tests could pass against the previous deployment while the CDN still served old files. The same metadata feeds the generated status page.

What the architecture cost, and what it bought

The split between content, data and templates costs indirection. A contributor adding a research note has to know that its claims go in data/claims.toml, its references in data/references.toml with an appraisal in data/sources.toml, and its commands in data/commands.toml. The validators turn that knowledge into error messages with file and line, which is the only reason the split is workable for an AI agent and for people alike.

What it bought is visible in the records. The first deployment pipeline, run 34834554169, passed all seven jobs and published v0.1.0. By the time of writing, five more releases had followed the same day, each after a live verification. The two pipeline failures recorded that day stopped where the design says they should: one before deployment, when the Crossref check rejected a reference title, and one after deployment but before release, when a browser test failed against production. Both are discussed in Testing a static site, and the timing is in Time accounting.

Interactive models and search in a static site

2026-09-14T00:00:00+00:00

Constraints

Everything on this site is static files on GitHub Pages. There is no server to run code, no database to query and, by design, no third-party service: the site makes no requests to other hosts, sets no cookies and loads no analytics. Content must also stay readable without JavaScript. That rules out the usual answers for three features the library needed:

interactive models that let a reader change an assumption and watch the consequence;
search across both language editions and the glossary;
a distinct link preview image for every page in every language.

Each was built so that the expensive or dynamic part happens either in the reader’s browser or once on a developer machine, and so that the build can check the result without repeating the work.

Five models as Alpine components

The interactive models page runs five simulations: the relief loop, venting and arousal over time, peak–end memory, synchrony in coupled oscillators, and allostatic load. Each is an Alpine.js component defined in static/js/models.js, a 344-line script loaded only on that page and registered before Alpine starts.

The division of responsibility is the same as elsewhere on the site (How the site was built): content lives in Markdown, behaviour in JavaScript. Each model’s question, explanation, how to read the chart, assumptions, limits, references and every UI label are front-matter fields in content/models/index.md and its Czech edition. The template passes the labels to the component as JSON:

<div class="model" data-requires-js data-model="synchrony" x-data="synchrony"
     data-labels="{{ l | json_encode }}">

so the script contains no language-specific text, and a reader without JavaScript still gets the full explanation, the assumptions and the limits, plus a note that the interactive part needs scripts.

Charts are plain SVG. The component exposes computed getters, and Alpine binds them to path and circle attributes:

const line = (values, max) =>
  values.map((value, index) => `${index ? 'L' : 'M'}${x(index, values.length).toFixed(1)},${y(value, max).toFixed(1)}`).join(' ');

There is no chart library. A 600 × 240 viewBox, three helper functions and a handful of CSS classes cover every chart on the page, so the visual language matches the rest of the site and the whole feature adds no dependency.

Determinism

Two of the models involve chance: whether the cause of distress changes in an episode, and the natural tempos of the oscillators. Math.random() would make every page load different, which is bad for readers comparing notes and worse for tests. The script uses a small seeded generator instead:

// mulberry32: a small deterministic generator so every reader sees the same run for the same inputs.
const random = (seed) => {
  let state = seed >>> 0;
  return () => {
    state = (state + 0x6d2b79f5) >>> 0;
    let t = state;
    t = Math.imul(t ^ (t >>> 15), t | 1);
    t ^= t + Math.imul(t ^ (t >>> 7), t | 61);
    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
  };
};

With a fixed seed and fixed time steps, the same slider positions always produce the same curves. The browser test “interactive models respond to their inputs and stay labelled as illustrative” can therefore change an input, assert that the output changes, and confirm that the models stay labelled; the page banner reads “Illustrative models · not predictions about any person”.

The Kuramoto model in a browser

The synchrony model is the most computational of the five. The Kuramoto model describes a population of oscillators, each with its own natural frequency, each pulled towards the phases of the others. Below a critical coupling strength the phases drift; above it, part of the population locks into a common rhythm. The site uses it as an analogy for collective rhythm, and says so in the ledger and on the page: “It is a mathematical model of synchronization in general, not a model of dancers, crowds or bonding.”

The implementation uses the mean-field form, which avoids comparing every pair of oscillators. First the component measures the order parameter, the length r and angle ψ of the average of all phases taken as unit vectors:

measure() {
  const n = this.phases.length;
  const re = this.phases.reduce((sum, phase) => sum + Math.cos(phase), 0) / n;
  const im = this.phases.reduce((sum, phase) => sum + Math.sin(phase), 0) / n;
  this.order = Math.hypot(re, im);
  this.angle = Math.atan2(im, re);
},

Then every oscillator advances by one Euler step of dθᵢ/dt = ωᵢ + K·r·sin(ψ − θᵢ):

this.phases = this.phases.map((phase, index) =>
  phase + DT * (this.tempos[index] + k * this.order * Math.sin(this.angle - phase)));

With at most 40 oscillators, a time step of 0.05 and two steps per animation frame, one frame costs a few hundred trigonometric calls, well within budget on a phone. The animation runs through requestAnimationFrame and stops when paused, and a “Step” button advances 40 steps at once without animating. The circle of dots and the arrow for r are SVG elements bound to the phases; a second SVG draws the recent history of r.

Small populations do not behave like the theory, and the page says so instead of hiding it: “With 4 to 40 oscillators, r fluctuates and rarely falls to 0 even without coupling; the sharp threshold of the mathematical theory holds for very large populations.” A simulation that looked cleaner than the mathematics allows would contradict the site’s own standard for data-like output. The model’s claims, including the coupling threshold and its use only as an analogy, are entries in the evidence ledger with Kuramoto (1984) and Strogatz (2000) as sources.

Search without a search server

The search page searches every page in the current language, including this one, and every glossary term. The index is produced at build time by python3 scripts/export-api.py, which writes api/v1/search.en.json and search.cs.json. Each entry has a URL, title, description, section, kicker, tags and up to 1,500 characters of body text with Markdown syntax, links and code blocks stripped. Glossary terms become entries whose URL points at the term’s anchor on the glossary page.

static/js/search.js fetches the index once, pre-normalizes every field and ranks entries as the reader types. The normalization carries most of the weight:

const normalize = (value) =>
  String(value ?? '')
    .normalize('NFD')
    .replace(/[̀-ͯ]/g, '')
    .toLowerCase();

const stem = (term) => (term.length >= STEM_FROM ? term.slice(0, Math.max(STEM_MIN, term.length - STEM_TRIM)) : term);

NFD splits “č” into “c” and a combining caron, and the regular expression removes the combining marks, so “přehodnocení” and “prehodnoceni” compare equal. Czech inflects heavily, and a query for “katarzi” should find pages that say “katarze”. A full stemmer for Czech would be a large dependency, so the script truncates: terms of six or more characters lose their last two, but never shrink below five. “Katarze” and “katarzi” both become “katar”. The browser test types exactly those two cases on the Czech search page.

Ranking is weighted substring matching. Every query term must match somewhere, or the entry scores zero; a match adds 10 points in the title, 5 in the kicker or tags, 3 in the description and 1 in the body text. Results are sorted by score, then by title, and capped at 50. The query is mirrored into ?q= with history.replaceState, so a search can be linked and the back button is not flooded.

The trade-offs are deliberate and visible. Truncation over-matches short common stems. Only the first 1,500 characters of a page’s body are searchable, which keeps the index small but misses late sections of long essays. There is no typo tolerance. In exchange, search adds no dependency and no service, the query never leaves the browser, and the whole index is an inspectable file in the public API.

A link shared in a chat or a social network shows a preview image. A single image for the whole site makes every shared page look the same, so each content page in each language gets its own 1200 × 630 card with its title, section and kicker.

Rendering those cards takes a real browser, since the design uses the site’s web fonts and Czech text. python3 scripts/render-social.py opens artwork/og.html in headless Chrome with the page’s fields as query parameters, takes a screenshot and converts it to a progressive JPEG with ImageMagick, rendering several pages in parallel. The output path mirrors the content path: content/research/x/index.cs.md becomes static/og/research/x/index.cs.jpg.

CI cannot sensibly do this on every push, and it does not need to. The script writes a manifest recording, for each image, a SHA-256 digest of everything that determines its pixels:

digest = hashlib.sha256(json.dumps({**fields, "template": template_hash}, sort_keys=True).encode()).hexdigest()

where fields are the title, section label, kicker, the red footer line and the language, and template_hash is the digest of og.html itself. With --check the script needs no browser at all. It recomputes the digests from the content files and fails when a page has no image, when an image’s recorded digest differs (a title was edited, or the template changed), or when an image belongs to no page. That check is the social stage of npm run validate. The Python tests give it a missing and a stale preview and assert both messages, and a browser test confirms that sampled pages in both languages serve distinct og:image URLs with image/jpeg responses.

One detail from the template shows why rendering was not left to a generic tool. The display font is a condensed face whose capitals sit tight against the line above, and Czech capitals carry accents above the cap height:

/* Czech capitals carry accents above the cap height; Anton needs extra leading so they clear the line above. */
:lang(cs) h1 { line-height: 1.24; padding-top: 0.08em; }

A similar adjustment appears on the site itself: a checkpoint from the first release records that the Czech display leading was raised to 1.2 “after measuring diacritic collisions”.

The hash scheme has limits, and they are worth stating. The digest covers the page fields and the template file, but not the font files the template loads or the Chrome version that renders it. Changing a font would leave every image marked current, and a re-render with --force is the remedy. Rendering is also tied to a developer machine with Chrome and ImageMagick installed; the pipeline verifies images, it does not produce them.

A common pattern

All three features follow the same pattern. Put the expensive or dynamic work where it is cheapest: in the reader’s browser for models and search, once on a workstation for images. Export what the work needs as plain, inspectable data: front matter, a JSON index, a hash manifest. Then make the build verify the result rather than trust it. The models are deterministic enough to test, the search index is part of a versioned API with its own tests, and a preview image cannot drift out of date without failing validation.

Parallel agents, one working tree: fan-out, proposals and explicit-path commits

2026-09-14T00:00:00+00:00

Why one tree and not branches

The usual answer to parallel work in Git is branches, one per worker, merged through pull requests. On 14 September 2026 the site was extended differently: several AI workers wrote into one working copy of main at the same time. The reason was practical. The work was mostly additive (new pages, new ledger entries), the pipeline releases every verified push to main, and the person directing the work wanted the result visible in one place as it grew.

One tree removes merge conflicts in the Git sense and replaces them with a different problem: two workers can write the same file at the same time, a worker can commit someone else’s half-finished change, and every repository-wide check sees everyone’s work in progress. This article describes how that was handled and what went wrong. It only reports what the records show: the Git history and its commit trailers, the GitHub Actions runs, the Majordomus ledger, and the files in the working copy. The case study shows the same period on a timeline.

Who was working

Three kinds of worker appear in the records.

The main session. Every commit from the poster to v0.5.0 carries the same Claude-Session trailer. This session started the supervised task that was still active at the time of writing, recorded at 14:54:07 UTC with the objective “Extend the site: more research and advice, per-page social previews, search and glossary cross-linking, second artifact Closure as a Service”.

Content agents dispatched by the main session. They do not appear in Git at all, because they were not allowed to commit. Their work appears as new directories in the working copy (research notes, theory essays, the interactive models page) and as proposal files in the main session’s private scratch space, one directory per agent. This article is itself the output of such an agent.

A peer session. Commits d2b3c47, 93fd07f and 2125f1e carry a different Claude-Session trailer from the main session’s. They were made on the same checkout, during the main session’s active task, while the tree held the main session’s and the content agents’ uncommitted work.

Fan-out: a brief that fixes ownership

A content agent receives a brief, and the brief carries the concurrency rules. The brief behind this section is typical. It lists the only files the agent may create (the section index and one directory per article, in both languages), names the files it must not touch (“templates, scripts, data/*.toml, zola.toml, tests or other content; other workers … edit those”), forbids Git operations and Majordomus write commands, and says where proposals go: a directory in the scratch space.

The rules come down to one principle: an agent writes only paths that no other worker writes. A research note lives in its own directory, so two agents writing two notes never touch the same file. This makes most of the fan-out conflict-free by construction, with no locking.

The exceptions are the files every page depends on:

data/references.toml, data/sources.toml, data/claims.toml, data/glossary.toml and data/evidence_changelog.toml, because every substantive claim needs a ledger entry (The evidence ledger as code);
data/commands.toml, because every command mentioned in prose needs a registry entry;
zola.toml, because UI strings live there.

Proposals and a single merger

For shared data, agents write proposals: TOML files with the same shape as the target, containing only their additions. The main session’s scratch space holds four such directories from the day’s fan-out, with names that match the work (expand, expand-glossary, expand-models, expand-theory), each with some of references.toml, sources.toml, claims.toml, glossary.toml and changelog.toml.

A short script in the same scratch space, merge_expand.py, applies them. Its docstring states the contract:

Merge agent proposal files (scratchpad/expand*/{references,sources,claims,glossary,changelog}.toml)
into <root>/data/*.toml by appending TOML text. Entries whose id already exists are skipped and reported.

Three details make the merge safe to run repeatedly:

Identifiers are checked before appending. A reference, appraisal, claim or term whose id already exists is skipped with a report line, so two agents that cite the same paper do not produce a duplicate key, which TOML would reject.
Each block must parse on its own (tomllib.loads(block) before it is written), so a malformed proposal fails before it reaches the shared file.
The changelog is prepended, because the ledger keeps its changelog newest first, and after merging the script parses all five data files again.

The merge appends text rather than re-serialising the data. That keeps comments, section headings and formatting in the hand-maintained files intact, and it keeps the diff for review to exactly the added entries. The main session then runs the validators on the merged tree, where problems the agent could not see on its own (an id that another agent used for a different paper, for example) come to light.

This is the one place where the work is serialised: many agents propose, one session merges. The evidence changelog in the working copy shows the result, for example an entry recording the claims and three references added for the interactive models page.

A peer session on the same tree

The peer session did not use proposals. It worked directly in the tree and committed its own changes, and the commits show how: by naming paths rather than committing everything that had changed.

d2b3c47 (15:14:22 UTC) changed five files: static/js/app.js, styles/app.css, templates/components/library.html, tests/site.spec.js and zola.toml. At the time of writing, four of those files (styles/app.css, the library component, the browser suite and zola.toml) carry further uncommitted changes from other workers, and those changes are not part of the commit.
2125f1e (15:38:00 UTC) extended the research note on measuring emotion and added eight references and six claims. The committed data/claims.toml contains 56 claims, the 50 from v0.5.0 plus its own six. When this article’s author counted the claims in the shared working copy a few minutes before that commit, the file already held 86; a count after the commit found 92, the same 86 plus the six. The other 36 entries stayed uncommitted.

The second case is the harder one. The peer session committed part of a file that other workers had also changed, and the result on main is a ledger consistent with the pages committed alongside it. The records do not say which Git mechanism was used to stage only those entries. They show the outcome: the commit’s contents match its message.

Both sessions ran under the same hooks: commit-msg lints the subject and pre-commit runs majordomus doctor. The Majordomus ledger records one active task in this period, the one the main session started. The supervision layer’s unit is the working tree and its task, not the session, so the peer session’s commits fell within the scope of a task it did not start. That task’s scope listed every top-level path in the repository, so the peer session’s files fell inside it; under a narrower scope, majordomus check would have reported them as drift.

What went wrong

The records show four problems in this phase. None was a lost change, and all four follow from sharing one tree.

Repository-wide validators fail on other workers’ unfinished files. When this article was drafted, running python3 scripts/validate-content.py on the shared tree reported FAIL on files this agent does not own. The translation validator reported content/about/index.cs.md: [extra] structure differs from en, a page another worker was editing at that moment. On a shared tree, a red validator does not tell a worker that its change is wrong. Content agents are therefore told to ignore errors in files they do not own, and to build a private copy of the repository when they need a clean render.

A commit can pass CI and still fail production. d2b3c47 passed all seven CI jobs and deployed. In run 34860959250, one production browser test failed and the release was skipped. The cause and fix are covered in Testing a static site. What matters here is how long the failure stayed open: the fix, 93fd07f, was committed 14 minutes 28 seconds after the failing job ended. Meanwhile the live site served a commit that had not been verified, while other workers kept writing into the tree.

Two commits eight seconds apart produce one pipeline run. 93fd07f and 2125f1e were committed at 15:37:52 and 15:38:00 UTC. The Pages workflow has a concurrency group with cancel-in-progress: true, and the run list shows one run, 34863500153, for the head commit. The run passed, and v0.6.0 released d2b3c47, the fix and the content change together. That is correct behaviour: the release notes list every commit since the last tag. It also means that a fix and an unrelated content change are verified, and released or not, together.

Earlier, before the fan-out: scope drift. In the single-session morning, three supervised tasks ended as partial because the declared scope no longer matched the files being changed; see Supervised AI delivery. With several workers on one tree, the main session’s task scope was set to cover every top-level path, which avoids those interruptions. It also means the supervision layer no longer separates one worker’s files from another’s.

What held, and what would have to change

The conventions that held are simple enough to write in a brief: own a directory, propose shared data, let one session merge, commit by path, validate your own files, render in a copy. At the time of writing, no work was lost and no commit mixed one worker’s changes with another’s: each commit after the fan-out contains the files and entries its message describes, and each went through the same pipeline as everything else.

Two things would have to change for this to scale beyond a handful of workers. The first is attribution. Neither Git nor the Majordomus ledger records which content agent produced which page, and the proposal directories that would show it are private scratch files that will not survive the session. The second is isolation. The shared-tree failures above, validators contaminated by unfinished work and fixes batched with unrelated changes, are exactly what separate worktrees and branches exist to prevent. The agent bootstrap file in the repository already describes a worktree convention for future work. Using it would give each worker a clean tree and a pipeline run of its own, at the price of merges.

Supervised AI delivery: briefs, scoped tasks, hooks, CI and live verification

2026-09-14T00:00:00+00:00

The claim this article makes, and the one it does not

This site was produced by an AI coding agent (Claude Code) directed by one person, with Majordomus as a supervisory layer. That sentence invites two readings. One is “the AI built a website”, which hides the work the person did and the constraints that made the output usable. The other is “the AI only typed”, which hides how much of the design was the agent’s. The records support neither. This article describes the loop that actually ran, using the Git history, the Majordomus ledger and the GitHub Actions runs as sources.

The case study presents the same records as an interactive timeline. The research note Method: building an evidence-gated artifact with Majordomus covers the first release in detail. Here the focus is the mechanism: how a brief becomes a verified release, and where each safeguard sits.

Instructions: from a one-line request to a completion gate

The method note classifies the person’s messages during the first phases into five kinds: a one-line request to create the project; art direction for the poster, with a choice between offered approaches; three structured execution briefs of increasing scope; short directives during the work (“deploy”, “validate in a browser”, “release automatically”, “use Majordomus and enforce it”); and editorial direction, such as replacing promotional copy with sourced psychology.

The briefs are the interesting part. They did not specify templates or test cases. They specified outcomes, constraints and, crucially, what counts as done: numbered requirements and an explicit completion gate. That shape matters for an agent, because an agent will otherwise report completion when the code is written. A gate of the form “deployed, verified against the live URL, released” turns completion into something a machine can check.

Scope also changed during the work, and the records say so. A checkpoint written at 11:09 UTC notes: “User scope expanded mid-task: landing explainer of catharsis, research essays with citations, advice library, tags, illustrations, per-page previews, Majordomus method page with real timeline, Rust integration, command registry with links and test coverage, pytest, JS coverage, docs.” The agent’s response was not to absorb the expansion into the running task but to close the task, hand over, and start a new one whose scope covered the new paths.

Scoped tasks

Every change ran as a Majordomus task started with majordomus start and a list of paths the work may touch. The ledger, which majordomus history prints, records six tasks on 14 September 2026:

Task	Started (UTC)	Closed (UTC)	Outcome	Why it ended
Ship the site with CI/CD	10:26:23	10:29:16	partial	files moved outside the declared scope; restarted
Ship the site (scope `.`)	10:29:16	10:39:39	partial	`.` was not read as the whole repository; restarted
Ship the site	10:39:39	10:47:25	completed	v0.1.0 deployed, verified and released
Boilerplate, previews, structured data	10:50:57	11:10:10	partial	new top-level paths planned; restarted with a wider scope
Evidence-based library, Rust, registry	11:10:11	13:08:34	completed	v0.5.0 deployed, verified and released
Extend the site	14:54:07	—	active	open at the time of writing

Three of six tasks ended as partial, and each ended for the same kind of reason: the declared scope no longer matched the work. That is the intended behaviour. majordomus watch and majordomus check report drift between the task and the files that changed; the handover written at 10:29 UTC states the problem plainly (“The first task scope omitted the pre-existing root files that were moved into artwork/ and static/ … so watch reports scope drift”) and the next action (“Restart the task with a scope covering the whole repository”).

One caveat about sources. The Majordomus ledger, checkpoints, handovers and decision log live under .ai/local/, which by policy is never committed. The quotations from them in this article are from the working copy on which the site was built; the public, independently checkable records are the Git history, the Actions runs and the releases linked throughout.

Restarting is cheap because state is externalised. Before closing a task, the agent writes a handover with majordomus handover, which has three required headings: Objective, Current State and Next Action. Progress notes go through majordomus checkpoint. The next task begins from those records rather than from the agent’s memory of the conversation, which is what makes the restart take seconds. The three partial tasks cost 2 minutes 53 seconds, 10 minutes 23 seconds and 19 minutes 13 seconds of recorded task time, and none of them pushed anything.

Decisions are records, not chat

Choices with lasting consequences were recorded with majordomus decision. The ledger holds four decision events that day: structured data built as Tera data and serialised with json_encode; link previews localized per language with every edition in the sitemap; Python validators tested with pytest in a project virtual environment; and the evidence ledger design. Each entry in the decision record has the same fields:

Why: Hand-written JSON-LD inside HTML templates gets HTML-escaped and breaks on quotes;
     building the graph as template maps keeps escaping correct by construction
Rejected: hand-written JSON-LD strings; a client-side script injecting JSON-LD
Evidence: templates/partials/structured-data.html; scripts/validate-html.py parses every JSON-LD block

The largest decision became an architecture decision record, ADR-0001, “Evidence ledger with dated review for every public claim”. Its status is proposed. The handover written when v0.5.0 shipped lists acceptance as the first next action for a person: “Accept or amend ADR-0001.” The agent can propose architecture; accepting it is left to the human.

Hooks: the checks nobody has to remember

Three Git hooks live in .githooks/, and npm ci points Git at them through the prepare script in package.json:

commit-msg runs scripts/lint-commits.sh on the message. Conventional Commits are not a style preference here: the release version is computed from them.
pre-commit requires Majordomus and runs its health check:

majordomus doctor || exit $?

pre-push runs the finish contract in check mode, so a push is refused while the active task would not be allowed to finish:

majordomus finish --check || exit $?

majordomus doctor earned its place before the first commit: it reported that the README did not link the agent bootstrap file, and the defect was fixed locally. majordomus finish is the contract that makes “completed” mean something. The task that shipped v0.1.0 was closed with the live smoke test as its verification command; the contract result in the ledger records each rule it evaluated (scope-integrity: pass, blocker-resolution: pass and so on).

Hooks can be skipped locally, so CI repeats the check. The Majordomus supervision job installs a pinned Majordomus release, verifies its SHA-256 through the installer, and runs the health check on the pushed tree. A repository ruleset names that job, together with commit linting, validation and the end-to-end suite, as required checks for main. The ruleset allows the repository administrator to bypass it, and the checkpoint after the first push records that the bypass was used: direct pushes to main are how this pipeline starts. The protection that does not depend on the ruleset is the order of the pipeline itself, which will not deploy a commit whose CI failed.

CI, deployment and live verification

The pipeline is described in How the site was built; what matters for supervision is its order. pages.yml runs CI, then deploys the exact artifact CI validated, then verifies production, then releases. The project rule project.verified-deployment states the reporting standard: “A deployment is complete only when the Pages workflow’s verify job has passed against the live URL returned by GitHub, and a release is cut only after that job.”

This is a rule for the agent’s reports as much as for the pipeline. An agent that watches a green build and says “deployed” is wrong in a way that is hard to notice. Here the agent watched the run with gh run list, and the checkpoint written after the first push says what it was waiting for: “Pages run 34834554169 triggered (ci -> deploy -> verify -> release); watching it now.” Its next steps were to investigate any failing job, run the smoke test and the browser suite against the live URL, confirm the release tag, and only then finish the task with a live verification command.

The records also show the rule holding when something failed. Run 34860959250, for commit d2b3c47, passed all seven CI jobs and deployed, but one of 84 browser tests failed against production. The Release job was skipped, and under the rule above that commit did not count as deployed. It was released only in v0.6.0 at 15:43:36 UTC, after a fix had passed verification in run 34863500153.

Releases

A release is the last job and has no manual step. scripts/release.sh finds the last v* tag, reads commit subjects since then, bumps minor for feat, patch otherwise (breaking changes bump minor while the version is 0.x), and publishes notes that name the live URL and the pipeline run. Tags are protected by a ruleset that forbids deleting, moving or updating them, and main forbids force pushes. Six releases were published on 14 September 2026 by the time of writing, from v0.1.0 at 10:45:42 UTC to v0.6.0 at 15:43:36 UTC.

Who did what

Attributing work in a human–agent session is easy to get wrong, so this list stays with what the records show.

The person created the direction and the boundaries: the project itself, the art direction, the briefs and their completion gates, the mid-task scope expansions recorded in checkpoints, the instruction to make supervision mandatory, and editorial standards for the psychology. The person owns the decisions still open in the handover: accepting ADR-0001 and reviewing Czech glossary terms that the agent flagged as non-standard. The time the person spent reading, thinking and reviewing is not recorded anywhere, which matters for Time accounting.

The agent wrote the code, templates, content drafts in both languages, validators, tests, workflows and documentation; chose the implementation (Tera components, the ledger schema, the test strategy); configured the repository rulesets; ran every command; recorded checkpoints, handovers and decisions; and proposed the ADR. Commits carry a Co-Authored-By trailer naming the model and a Claude-Session trailer naming the session.

Majordomus wrote nothing. It refused, recorded and reported: scope drift, a malformed scope, a missing bootstrap link, and the finish contract. The pipeline refused twice that day: once before deployment (a reference title mismatch against Crossref) and once before release (a failing production test).

What this does not show

The records show a sequence of checked steps, not that supervision made the work faster or better than an unsupervised session would have been. There was no second session to compare with. The agent also made errors that the supervision layer did not catch and the pipeline did; those are listed in Testing a static site. What the records do support is narrower and still useful: every “done” in this repository is attached to a command, an exit code and a timestamp that a reader can check.

Testing a static site: validators, negative tests, browsers and contracts

2026-09-14T00:00:00+00:00

What can go wrong in a static site

The site in this repository has no database and no application server, and still has many ways to be wrong in production without any error message:

the Czech edition of a page is missing, or has a different front-matter shape than the English one;
a UI string exists in one dictionary and not the other, so a template renders an empty label;
a claim on a page has passed its review date;
a reference cites a DOI whose title no longer matches;
an in-page anchor or an asset URL is broken, often only under the /catharsis-as-a-service/ subpath;
structured data does not parse, or describes a page as a medical resource;
a command mentioned in documentation does not exist, or exists but no test runs it;
the deployment succeeded, but the CDN still serves the previous build.

Each item on that list has a check in the repository. This article walks through the checks from the fastest to the slowest, then through the failures they recorded. The pipeline that runs them is described in How the site was built.

Stage 1: one ordered validation script

npm run validate runs scripts/validate.sh. Every step is wrapped in the same function, which records PASS or FAIL, prints a summary table and stops at the first failure:

step() {
  local name="$1"
  shift
  printf '\n── %s: %s\n' "$name" "$*"
  if "$@"; then
    RESULTS+=("$name|PASS")
  else
    RESULTS+=("$name|FAIL")
    report
    printf '\nRESULT %s FAIL\n' "$(dots RESULT)"
    exit 1
  fi
}

The stages are, in order: toolchain (the Zola version must equal .zola-version), assets, JavaScript syntax, translations, content standards, references, the evidence ledger, social previews, the Zola check, the Zola build, the API export, and the HTML checks on the generated site. The order is deliberate. The translation check takes well under a second and catches the most common authoring error, so it runs before a build that takes much longer.

The script has no options to skip a stage. The project rule on deployment says “Validation is never bypassed to make CI pass: no || true around a mandatory check, no --no-verify”, and the CI job runs the same entry point as a developer does.

Stage 2: Python validators that are tested to fail

The validators are small standard-library Python scripts, each with one responsibility:

Validator	Catches
validate-i18n.py	missing translations of content files, front-matter shape mismatches, UI keys used in templates but not defined
`python3 scripts/validate-content.py`	missing research-note fields, description length, prohibited phrasing, unregistered or unlinked commands
`python3 scripts/check-references.py --online`	unknown citations, articles without a DOI; online, title and year mismatches with Crossref
`python3 scripts/validate-claims.py`	the evidence ledger rules described in The evidence ledger as code
`python3 scripts/render-social.py` with `--check`	a page without a preview image, a stale image, an orphaned image
validate-html.py	duplicate ids, missing alt text, heading jumps, broken first-party references and anchors, invalid JSON-LD, medical schema types, incomplete preview metadata

A validator is only useful if it fails when it should, and a validator that passes on the repository proves nothing about that. The test module’s docstring states the rule: “Validators must pass on the repository and fail on the defects they exist to catch.” Almost every validator therefore has two kinds of test in tests/python: one that runs it on the real repository and expects success, and several that build a minimal broken fixture in a temporary directory and assert the specific message:

def test_content_rejects_wrong_registry_anchor(tmp_path):
    root = _content_fixture(tmp_path, "Run [`npm run validate`](../commands/#npm-test).\n")
    result = run([PY, str(ROOT / "scripts/validate-content.py"), "--root", str(root)])
    assert result.returncode == 1
    assert "must link to the registry anchor #npm-run-validate" in result.stdout

The fixture approach requires every validator to accept a --root argument or a CAAS_ROOT environment variable, a small design constraint that falls out of wanting negative tests. The ledger validator also accepts --today, which lets a test show that the repository’s own claims expire in 2031 without a review.

npm run test:python runs the suite in a project virtual environment. At v0.5.0 it had 40 tests.

Stage 3: the command registry is executed

Documentation that lists commands tends to rot, so every command the site, the README or the docs mention is recorded in data/commands.toml with a verified_by list naming the CI jobs or pytest functions that run it. The content validator refuses a command in prose that does not link to its entry. tests/python/test_commands.py closes the loop in the other direction: it parses the workflow files and the test modules, and fails if a verified_by name does not exist.

for reference in entry["verified_by"]:
    kind, _, name = reference.partition(":")
    if kind == "ci" and name not in jobs:
        problems.append(f"{entry['id']}: CI job '{name}' does not exist (known: {sorted(jobs)})")
    elif kind == "pytest" and name not in tests:
        problems.append(f"{entry['id']}: pytest '{name}' does not exist")

The same module runs commands for real: the commit linter on a good and a bad message, the release script in dry-run mode, a production build, the development server (polled until it answers), --help for every Majordomus subcommand, and a read of the pipeline history through the GitHub CLI. Tests that need an external tool skip locally when the tool is missing, but CI sets CAAS_REQUIRE_TOOLS=1, and then a missing tool fails instead of skipping. Without that switch, a registry entry could be “verified” by a test that never ran.

Every link to a command in this article points at that registry. You are reading the result of the check that enforces it.

Stage 4: Playwright on the production subpath

npm test runs the browser suite in tests/site.spec.js. With no BASE_URL, Playwright starts the preview server itself, so the suite runs against the production build served under the same subpath as GitHub Pages:

const LOCAL_URL = 'http://127.0.0.1:4173/catharsis-as-a-service/';
const baseURL = process.env.BASE_URL || LOCAL_URL;

Several test blocks are loops over locales, viewports and pages, so the number of executed tests is larger than the number of blocks. The suite checks what only a browser can see: pages render at every configured viewport without horizontal overflow; every page is self-canonical with correct hreflang alternates; the Flowbite modal traps focus and the drawer closes; claim panels open their sources with JavaScript and stay open without it; the glossary is anchored; the detection simulation on the methods page stays labelled as a simulation; print styles turn the artifact into a paper document; library pages are Article and never MedicalWebPage.

A browser test is also the only reliable test of “readable without JavaScript”. One block creates a context with JavaScript disabled and checks that content, navigation and claim sources remain available.

The suite grew with the site: 30 tests at v0.1.0, 72 passing at v0.5.0 according to the handover written when it shipped, and 84 in the production run for commit d2b3c47.

Stage 5: the same suite against production

After deployment, the Verify production job waits until the live site’s footer contains the SHA of the commit being deployed, then runs npm run smoke and npm run test:production. The smoke script requests every entry point, feed, preview image and first-party asset and compares status codes and lang attributes; the browser suite then runs unchanged against the live URL.

Running the same tests twice is not redundant. The preview is a Python HTTP server on a Linux runner; production is GitHub Pages behind a CDN, with its own headers, caching and redirects. The third failure below passed on the preview and failed on production.

Stage 6: Rust contract tests

The JSON API is tested from a second language. crates/caas-content/tests/contract.rs runs the exporter into a temporary directory, or reads CAAS_API_DIR when CI sets it to the export of the commit under test, and asserts that the typed library loads and has no integrity problems:

#[test]
fn the_exported_library_has_no_integrity_problems() {
    let problems = library().problems();
    assert!(
        problems.is_empty(),
        "integrity problems:\n{}",
        problems.join("\n")
    );
}

The CLI crate has its own tests that run the caas binary. cargo test --workspace, cargo clippy with warnings as errors and cargo fmt run in the CI job Rust. At v0.5.0 the workspace had 21 Rust tests.

What the checks caught on the first day

The records show three failures that matter, and each stopped at a different layer.

A flaky print test, caught locally. A checkpoint written at 11:09 UTC records “E2E was 36/37: print test read computed style once”. A preview probe showed that the print styles did apply, so the defect was in the test, which read the computed style a single time instead of waiting for it. The test was changed to poll. Nothing was pushed with the failure.

A Crossref title mismatch, caught in CI before deployment. Commit 51dde0b added the research library. In run 34842607207 the References (Crossref) job failed at 12:17:36 UTC with:

error: lieberman-2007: title differs from Crossref (0.42): 'putting feelings into words'

The reference was correct; the comparison was not. The error shows that Crossref returned only the main title, while the registry held the full title with its subtitle, and the validator compared the two as they were. Deploy, verify and release were skipped. The fix, 94d734c, compares the registered title and its main part against Crossref’s title with and without the subtitle and keeps the best match. It was committed at 12:20:54 UTC, and v0.3.0 was published at 12:24:28, 6 minutes 52 seconds after the failing job ended.

A browser difference, caught in production before release. Commit d2b3c47 added in-place previews for the concept screens. All seven CI jobs passed, including the full browser suite on the preview, and the site deployed. In run 34860959250, Verify production then failed one of 84 tests, “concept screen preview: modified clicks keep the link to the original PNG”, with a timeout while waiting for a new tab after a Ctrl or Meta click; Playwright’s retry failed the same way. The Release job was skipped. The follow-up commit, 93fd07f, explains the cause: “on the Linux runner a Ctrl+click tab stayed on about:blank”. It changes the test to check what the page controls (modified and middle clicks are not cancelled and do not open the dialog) rather than how a particular browser opens tabs. The next pipeline, run 34863500153, passed all ten jobs, and v0.6.0 was published at 15:43:36 UTC, 20 minutes 12 seconds after the failing verification ended.

The third case is worth dwelling on. The site was briefly live with a commit that had not passed verification, because deployment has to happen before production can be tested. The rule the project follows is about claims, not about bytes on a CDN: without a passing verify job there is no release, and nothing is reported as deployed.

What these tests do not cover

None of these checks judge whether a sentence about psychology is right, whether an appraisal is fair, or whether the Czech reads naturally; those remain review tasks, and the handover from v0.5.0 lists a terminology review as open. The suite does not measure performance budgets, and accessibility is checked through specific behaviours (focus trapping, labels, alt text, heading order) rather than a full audit. The pipeline also has a structural blind spot: a defect that appears only on GitHub Pages is found after the site is live. The mitigation is that verification is automatic, it runs within minutes, and nothing is released until it passes.

Time accounting: what the records measure, what an estimate suggests, and why they cannot be compared directly

2026-09-14T00:00:00+00:00

The question and the trap

People who hear that a site like this was built in a day ask how much time that saved. It is a fair question, and an easy one to answer badly. The bad answer takes the hours on the clock, compares them with a guess of how long “a team” would take, divides, and publishes the ratio. That ratio compares a measurement with an opinion, mixes elapsed time with effort, and ignores that the two products might not be equally good.

This article separates three things and keeps them separate:

Measurements from records: the Git history, the GitHub API (repository, releases, workflow runs and their jobs) and the Majordomus ledger. Every number is linked or reproducible.
An estimate, explicitly labelled, of the effort the same scope would take a human team, built bottom-up so that every assumption can be disputed line by line.
The reasons the two cannot be combined into one factor, and what a real measurement would require.

The research note Method: building an evidence-gated artifact with Majordomus covers the first release in detail, and the case study shows the full timeline interactively. This article covers the whole day as recorded at 15:47 UTC.

Part A: what the records measure

Milestones

All times are UTC on 14 September 2026.

Time	Event	Source
09:45	Working session begins (poster rebuild)	local file timestamps, as reported in the method note; not independently verifiable
09:54:06	Repository created on GitHub	GitHub API `created_at`
10:26:23	Majordomus initialised; first supervised task	Majordomus ledger
10:45:42	v0.1.0: bilingual site, CI/CD, live verification	release list
11:13:28	v0.2.0: boilerplate practices, previews, structured data	release list
12:24:28	v0.3.0: research library, advice, command registry, Rust API	release list
12:58:21	v0.4.0: evidence ledger, methods, glossary, status	release list
13:07:10	v0.5.0: ledger in the API, Rust client, documentation	release list
13:08:34	Task closed as completed; no supervised task is recorded until 14:54:07	Majordomus ledger
15:43:36	v0.6.0: screen previews, a fix, expanded research note	release list

Derived durations:

repository creation to first verified release: 51 min 36 s;
repository creation to v0.5.0: 3 h 13 min 04 s;
repository creation to v0.6.0: 5 h 49 min 30 s, a window that includes 1 h 45 min 33 s with no supervised task recorded.

Ten commits, six releases, four recorded decisions and one proposed ADR fall in that window.

Supervised task time

The Majordomus ledger, readable with majordomus history, records when each task started and finished. Up to v0.5.0, five tasks account for 2 h 38 min 38 s. The three tasks that ended as partial and were restarted (scope errors, described in Supervised AI delivery) account for 32 min 29 s of that. Task time is when the agent was working under a declared task. It does not show how much of that time the person spent reading, deciding or typing.

Pipeline runs

Each push to main starts the Pages workflow. The table lists every run that day, with wall-clock time from creation to completion, the sum of the durations of the jobs that ran (jobs run in parallel, so this is larger), and the time from the run’s creation to the published release. The data can be listed with gh run list.

Run	Commit	Result	Wall clock	Job time	Push to release
34834554169	9bc6897	success	2:43	3:02	2:40 (v0.1.0)
34836842423	54a31a2	success	3:06	3:24	3:02 (v0.2.0)
34842607207	51dde0b	failure: references	1:29	3:03	none
34842946706	94d734c	success	3:25	4:57	3:21 (v0.3.0)
34845982889	4cbb6fb	success	4:17	6:01	4:14 (v0.4.0)
34846771284	f2890f4	success	5:02	6:46	4:59 (v0.5.0)
34860959250	d2b3c47	failure: verify production	8:01	8:44	none
34863500153	2125f1e	success	5:07	6:15	5:03 (v0.6.0)
Total		6 passed, 2 failed	33:10	42:12

The pipeline grew from seven jobs to ten when the Python, Rust and Crossref jobs were added with the research library, and the browser suite grew from 30 to 85 tests, so later runs are longer. The weekly evidence workflow was also run once by hand (38 seconds).

Failures and recovery

Failure	Detected	Fix committed	Verified release	Detection to release
Crossref title comparison (details)	12:17:36	12:20:54	12:24:28 (v0.3.0)	6 min 52 s
Modified-click test on production	15:23:24	15:37:52	15:43:36 (v0.6.0)	20 min 12 s

Neither failure reached a release. The second left an unverified commit on the live site for about twenty minutes, because production can only be tested after deployment.

What the measurements do not include

Planning, thinking and prompt writing before the first recorded file.
The person’s time. No record shows when the person was reading, deciding or away; the only bound is the window itself.
Time after the records were read: the review items the v0.5.0 handover leaves open (accepting ADR-0001, reviewing flagged Czech glossary terms), and any rework they cause.
Model usage and its cost, which are not part of the repository’s records.

Part B: an estimate for a human team

This section is an estimate, not a measurement. It asks how many person-hours an experienced team (a web developer, a technical writer with a psychology background, a Czech editor, and part-time DevOps and Rust engineers) would need to produce the scope of v0.5.0, to the same automated checks, starting from a finished poster. v0.5.0 is used because its scope can be measured from the tagged tree and it closed a supervised task; later work is excluded from both sides.

The scope column was measured from the repository at commit f2890f4. Word counts exclude front matter. The hour ranges are the author’s judgement, and the assumptions say where each comes from. Readers who disagree with a rate can recompute that row.

#	Work item	Scope at v0.5.0 (measured)	Assumption behind the range	Low (h)	High (h)
1	English content with sources read	32 English files, about 19,300 words; 10 research notes, 10 advice entries	150–300 words per hour including reading abstracts and checking wording against sources	65	130
2	Czech edition	32 Czech files, about 17,100 words	250–375 words per hour for professional adaptation with psychological terminology	45	70
3	Source appraisals	58 references, each graded with a bilingual appraisal	0.5–1 h per source: locate abstract, classify design, write both languages	29	58
4	Claims ledger	50 claims with scope, caveat and dates in two languages	15–30 min per claim once sources are appraised	13	25
5	Glossary	54 bilingual terms with cross-references	9–18 min per term	8	16
6	Design system, templates and components	2,053 template lines, 2,472 CSS lines, two locales	a senior front-end developer at roughly 50–100 finished lines per hour including responsive and accessibility work	40	80
7	Interaction	Alpine and Flowbite integration, 381 lines of site JavaScript, no-JS fallbacks	small code, high testing and accessibility cost	12	24
8	Metadata and SEO	JSON-LD, localized previews, feeds, sitemap alternates	known patterns, per-locale verification	8	16
9	Validators and API exporter	1,601 lines of Python and shell in the scripts directory	35–65 lines per hour for tested tooling	24	48
10	Test suites	40 pytest tests with fixtures; 23 Playwright blocks executing 72 tests	fixtures and browser flakiness dominate	24	48
11	Rust crate and CLI	1,182 lines, 21 tests	typed contract plus CLI and contract tests	16	32
12	CI/CD and repository settings	three workflows and a composite action (352 lines), release script, rulesets, Pages	includes debugging live deployment	12	24
13	Supervision layer, rules, ADR, documentation	6 project rules, 1 ADR, 5 documents, README	writing and wiring, not tool development	10	20
14	Artwork production pipeline	poster render script, icon set, localized preview cards	excludes the artistic concept	8	16
	Subtotal			314	607
	Coordination, review and integration		10–20% of the subtotal for a team of four to five people	31	121
	Estimated total			≈ 345	≈ 730

At 40 hours a week that is roughly 9 to 18 person-weeks. The range is wide on purpose. The largest items, content and appraisals, depend on how carefully sources are read, and that is also where a human expert might produce a different and possibly better product.

Several judgements pull the estimate in opposite directions:

Towards lower: an experienced team with a reusable Zola starter, an existing validator library or a translation memory would be faster on rows 6, 8, 9 and 2. Some items might not be built at all without an agent doing them cheaply (a second-language contract test, an executed command registry), and a team would be right to skip them.
Towards higher: the ranges assume little rework. The agent’s own records show rework (three task restarts, two pipeline failures, a flaky print test), and a human team would have its own. Bilingual review cycles, stakeholder feedback and waiting between people are not in the table.

Part C: why these numbers cannot be divided

It is tempting to set 3 h 13 min against 345–730 person-hours and state a factor. This article does not, for five reasons.

1. Different units. Elapsed time is wall-clock time during which an agent, and sometimes several at once, worked while a person directed and reviewed. Person-hours are effort summed across people. Parallel human work would shorten the calendar time of the estimate without changing its hours, and parallel agents do the same to the measurement.

2. No control condition. There was one project, built once. There is no human team that built the same scope, no second agent session without supervision, and no repeat. A single case can show that something is possible. It cannot show how large an effect is.

3. The quality bar is not demonstrably equal. The estimate assumes the human team meets the same automated checks, and automated checks are not the whole of quality. The ledger records that source designs were checked against published abstracts and that two sources without a retrievable abstract are marked unverified. A human expert reading full texts would take longer and might grade differently. The Czech terminology review and the ADR acceptance are still open. Where the agent’s product is weaker, the comparison flatters it; where it is more thorough than a team would bother to be (every command executed in a test), the comparison penalises it.

4. Human direction time is unrecorded. The person wrote briefs, expanded scope, chose between options and reviewed output. None of that is timed. If the person attended the whole window, their effort is bounded by it; if not, it is smaller. Either way it belongs in the comparison and is missing.

5. The estimator is not independent. The estimate was written after the fact by an AI agent of the same kind that did the work, with knowledge of the result. Anchoring on the known outcome is a plausible bias in either direction. Line counts and word counts are also weak proxies for effort: a 17-line vendoring script and a 255-line validator do not cost in proportion to their length.

What the records do support is narrower. The scope of v0.5.0, measured above, reached a verified release about three hours after the repository was created. Every release in the window passed the same automated checks, and the two failures were caught by those checks and recovered within minutes. Against any plausible reading of the estimate table, that is a large difference in elapsed time. How large, in a sense that would survive scrutiny, is not something one uncontrolled case can say.

What a real measurement would need

A defensible comparison of supervised AI delivery with conventional delivery would need at least:

A fixed, pre-registered scope and acceptance gate. The same brief, and the same validation script and test suite as the definition of done, agreed before any work starts.
Comparison arms. A human team, an agent directed by a person without supervision tooling, and an agent with supervision, ideally on several projects of different kinds, with arms assigned so that the same person does not direct two arms of the same project.
Effort, not just elapsed time. Time tracking of human attention (direction, review, waiting) in every arm, and agent compute time and cost alongside.
Blind quality assessment. Independent experts rating samples of content (accuracy of claims against sources, appropriateness of grades, quality of the Czech) without knowing which arm produced them, plus an accessibility audit.
Outcomes after release. Defects found in the weeks after release, review dates missed, and the cost of the first substantial change, since maintainability is where hurried work usually charges its bill.

Until then, the honest summary is the pair of numbers with their labels: a measured three hours and thirteen minutes to a verified v0.5.0, and an estimated 345 to 730 person-hours for a human team to build the same scope, which are not the same kind of number.

Evidence ledger

2026-09-14T00:00:00+00:00

Why a ledger

A site that argues that relief is not resolution has to hold its own statements to the same standard. It is easy to write “research shows” and much harder to say which research, of what design, with which limits, and when someone last checked. The ledger answers those questions in one place.

Every substantive public claim on the site has an entry: the four findings on the landing page, the answers to frequently asked questions, the key points of each research note, the recommendation in each advice entry, the statements made by the artwork and the technical statements about the site itself. Each entry records the claim’s wording, its type, its evidence level, a confidence judgement, its sources, its scope, its most important caveat, the pages where it appears and the dates of its last and next review.

Bibliographic facts and judgement are kept apart. The reference registry holds titles, authors and DOIs checked against Crossref. The ledger holds the appraisal of each source, including its design, population and main limitation, and that appraisal is dated because judgement can go out of date even when a citation does not.

How a claim is graded

Grading happens in two steps.

First, each source is graded by design. A meta-analysis of randomised trials is level A; a single controlled experiment is level B; a longitudinal or uncontrolled pre-post study is level C; correlational, qualitative, narrative-review and theoretical work is level D; classical texts, project records and artistic material are level E. A meta-analysis that pools correlational studies is graded by what it pools, not by the word “meta-analysis” in its title.

Second, each claim receives an evidence level no stronger than its strongest source, and a separate confidence rating. The two answer different questions. The level describes the kind of study behind the claim. Confidence describes how well the claim, as worded, is supported. A meta-analysis with heterogeneous results can support a level A claim held at moderate confidence. A well-replicated correlation can be level D and still justify high confidence, provided the claim speaks of association and not of cause.

When a source’s abstract could not be retrieved, the ledger says so, states no sample sizes for it and lowers confidence in claims that rely on it. When wording on a page went further than its sources, the ledger records the stricter version and the changelog notes the revision.

Relief, resolution and the limits of evidence

Most of the research cited here measures short-term outcomes in specific samples: anger minutes after a provocation, closeness after a dance, mood after a film. That evidence is suited to questions about relief. It is much less suited to questions about resolution, which unfold over months in circumstances no laboratory reproduces. Where the ledger says there is insufficient evidence, that is a finding about the literature, not a verdict on the experience.

Art and evidence on the same site

The site is both an artwork and a library, and the two use different kinds of statements. The ledger marks them explicitly. A claim labelled ARTISTIC belongs to the artwork: the telemetry values, the status line and the field problem_solved: false express a thesis and are never measurements of anyone. A claim labelled EMPIRICAL reports what studies observed, with sources and limits. Theoretical, historical, technical and clinical-boundary claims sit between them and are labelled as well.

Keeping the labels visible protects both sides. The artwork can use the language of dashboards without its numbers being quoted as data, and the research notes can be read without the artwork’s rhetoric lending them false authority.

How to report a problem

If a claim seems outdated, overstated or wrongly sourced, use the report link on this page. A useful report names the claim, explains what is wrong and, where possible, cites a newer or better source. Reports are checked against the original publications. Confirmed problems lead to a revised, downgraded or retracted claim, and the change is recorded in the evidence changelog with its date and reason.

The content of this site is general education, not medical advice. If distress persists, disrupts daily life or includes thoughts of harming yourself, contact a qualified professional or local emergency services.

Should I talk about a traumatic event right away?

2026-09-14T00:00:00+00:00

What the debriefing review actually found

A Cochrane systematic review of eleven randomised trials examined single-session individual psychological debriefing delivered within a month of a traumatic event. It found no evidence that the intervention prevents PTSD or reduces distress, found some evidence of increased PTSD risk at one year in one trial, and advised that compulsory debriefing should cease.

Two limits belong with that result. The trials were generally of poor quality. And the review evaluated one specific format: a single, structured, individual session, often offered to everyone exposed to an event whether or not they had symptoms. It does not show that talking after trauma is harmful in general, and it did not evaluate other early interventions.

The practical reading is narrow and worth stating plainly. Being required to recount an event, to a stranger, on a schedule set by someone else, has not been shown to help and may not be neutral. That says nothing about a conversation you choose, with someone you trust, when you want it.

What the trial evidence supports instead

The picture changes once symptoms are in the frame. Early multiple-session interventions showed no clinically important benefit over usual care for trauma-exposed people who were not screened for symptoms. For people who did have traumatic stress symptoms, trauma-focused CBT, cognitive therapy without exposure and EMDR showed clinically important benefits. For chronic PTSD, trauma-focused CBT and EMDR reduced clinician-assessed symptoms more than waitlist or usual care.

These are group-level results with real limits. The Cochrane review rated the evidence for every comparison as very low quality, drop-out was higher in the active treatments, and long-term follow-up is limited. Choosing among these treatments requires clinical assessment, which no page can do.

The shape that emerges is a distinction between offering something to everyone who was present and offering something to people who are struggling. Screening, waiting, and then treating those with symptoms is supported better than treating in advance.

Why “just get it out” is not the mechanism

Emotional processing theory holds that lasting change in fear requires activation of the fear memory together with information incompatible with it; activation alone is not sufficient. The theory accounts for why retelling can help in one setting and do nothing in another: the retelling is a vehicle for the new information, not the active ingredient on its own. Later research has revised parts of the theory, and the effectiveness of treatments rests on trials rather than on the theory itself.

This is the same distinction the rest of the site draws. Recounting an event can produce release. Whether anything changes depends on what the recounting is attached to.

In practice

There is no deadline. Not talking in the first days is not a failure, and talking is not a treatment you owe anyone. What the evidence does support is watching what happens over the following weeks, and treating persistence as the signal rather than intensity in the first hours.

If distress persists, disrupts daily life, or includes thoughts of harming yourself, contact a qualified professional; immediate danger warrants contacting local emergency services.

Can emotions be measured?

2026-09-14T00:00:00+00:00

What an instrument actually records

Nothing in a laboratory records an emotion. A camera records the geometry of a face over time, and a coder or an algorithm labels the movements. Electrodes record the interval between heartbeats, or the conductance of the skin on a fingertip. A questionnaire records what a person is willing and able to say about their state. The emotion is a construct inferred from one or more of those recordings, and the quality of the inference is the whole question.

Mauss and Robinson reviewed the field and concluded that there is no gold-standard measure: experiential, physiological and behavioural measures converge only modestly and cannot be treated as interchangeable. That is a narrative review rather than a pooled estimate, so it summarises a literature rather than quantifying it, but the conclusion has not been overturned by newer multimodal methods.

Faces

Barrett and colleagues reviewed studies of how people move their faces during emotion and how observers read those movements. Typical configurations do co-occur with the expected emotions more often than chance, but weakly and inconsistently: the same configuration accompanies different states, the same state is shown in many ways, and the pattern varies across people, situations and cultures. The review is associated with one theoretical position, and some researchers argue the expression–emotion link is stronger than it concludes.

Durán and Fernández-Dols put numbers on it in a meta-analysis. For six classic basic emotions, average co-occurrence effect sizes were .13 for full expressions coded with FACS and .23 once partial expressions were included, and reported intensity correlated with expressed intensity at .30 on average. Heterogeneity between samples was found, and co-occurrence was highest for disgust and lowest for surprise, so these averages do not describe every emotion equally.

Bodies

Siegel and colleagues meta-analysed 202 studies of autonomic responses and found that they did not clearly distinguish one emotion category from another, with substantial variation within categories. Kreibig’s earlier review of 134 publications is a partial counterpoint, reporting considerable specificity once emotion subtypes are considered. The two syntheses disagree, so the defensible statement is that no consistent autonomic fingerprint has been established — not that physiology is uninformative.

Coherence across systems is likewise partial. In Mauss and colleagues’ film study, inducing amusement and sadness, experience and facial behaviour were highly associated while physiology was associated only modestly. Each kind of measure carries variance the others do not. That is one laboratory study, and coherence may differ by emotion and intensity, as the gap between amusement and sadness already shows.

What follows for anyone reading a number

Heart rate variability can index cardiac vagal tone, and electrodermal activity is a usable signal, but both are easily misconstrued without careful design, analysis and reporting — which is what Laborde and colleagues and Boucsein’s methods guidance exist to enforce. That guidance concerns how to measure well. It is not evidence that any signal corresponds to a specific emotion.

So the workable answer: aspects of emotional responding can be measured, one system at a time, with known error. A single reading from a single sensor supports a conclusion about one person in one moment only with low confidence, and a product claiming otherwise is claiming more than the literature supports.

Is catharsis the same as venting?

2026-09-14T00:00:00+00:00

Two words that everyday language merges

In ordinary speech, “cathartic” and “letting it out” are near synonyms, and that merging hides a distinction worth keeping. Catharsis, in the sense this library examines, names an experience: the sense, after expressing or exerting something, that a tension has been released. Venting names a behaviour with a target and a purpose: expressing anger at, or about, the person or situation that caused it, in the belief that expressing it reduces it.

One is a feeling reported from the inside. The other is an action whose consequences can be measured from the outside. They can occur together, and they often do, which is why the two words drifted into each other. But they answer different questions, and the evidence for each is different.

Why the distinction matters for reading evidence

The literature on venting is fairly clear about what venting does to anger. Kjærvik and Bushman’s meta-analysis of 154 studies (184 samples, 10,189 participants) grouped anger-management activities by whether they raise or lower physiological arousal. The arousal-increasing activities, hitting a bag among them, showed no reduction in anger and aggression overall, with a pooled effect of g = -0.02 (95% interval -0.13 to 0.09). The studies in that group were heterogeneous, so the pooled near-zero average does not mean every such activity in every context behaves the same way, and it says nothing about the other benefits exertion has.

That result is about anger and aggression. It is not a finding about dancing, about music, about crying, or about the release a crowd can produce. Those have their own literatures, and they were not tested here. Stretching an anger finding to cover them would be the same error in reverse: treating a behaviour’s outcome as though it settled a category of experience.

There is one point the distinction does not dissolve. In the venting studies, people who reported feeling released were not thereby less angry on the measured outcomes. The release was real as an experience and, because it felt good, it made the behaviour more attractive to repeat. What it was not is evidence that the anger, or the provocation behind it, had changed.

That generalises beyond anger. Whatever the activity, a felt release is a report about how a state feels now. Whether the conditions that produced the state have changed is a separate question that has to be asked separately, and often answered by something other than the feeling.

How the site uses the words

This site keeps catharsis for the felt release and venting for the anger behaviour, and it tries never to let a claim about one travel silently into the other. Where a finding concerns venting, it is labelled as such. Where a finding concerns crying, exercise, writing or collective movement, it is reported on its own evidence, including where that evidence is thin.

The distinction is conceptual rather than experimental: nobody ran a study to establish it. Its value is that it keeps two honest statements from being confused into a false one.

Does a crowd change anything after the night ends?

2026-09-14T00:00:00+00:00

What the large synthesis found

Pizarro and colleagues pooled 50 samples covering participants in demonstrations, rituals, ceremonies and celebrations (N = 182,738). Collective effervescence during these gatherings was associated with self-transcendent emotions, group identity, social integration, wellbeing and collective efficacy. That is a large body of people and a consistent pattern, and it is worth taking seriously.

It is also mostly correlational. The pooled studies do not show that gatherings cause these outcomes: people who already feel connected to a group are more likely to attend, more likely to report intensity, and more likely to report wellbeing for reasons that have nothing to do with the night itself. The association is real; the arrow is open.

Why the experiments cannot close the gap

The experimental work — tapping and walking in synchrony, group dance, listening to self-chosen music — is where causal direction is established, because the conditions were assigned. But those studies measured liking, cooperation, pain thresholds, closeness and reward responses during the activity or shortly after it. Nobody followed participants for a week.

So on the specific question of whether collective rhythm reduces the difficulties people bring into it, there is insufficient evidence. That is not the same as evidence of absence. Longer studies could find lasting benefits; they have largely not been run, and the honest position is that the question stands open rather than answered negatively.

The one mechanism with a longer horizon

There is a plausible route by which a night out could matter later, and it is social rather than physiological. Cohen and Wills reviewed the literature on social support and found that the perceived availability of support matched to the needs a stressor creates is associated with protection against its adverse effects, while social integration has a broader association with wellbeing.

Read carefully, this points away from the dancefloor and towards the people on it. What would plausibly carry over is not the effervescence but a relationship that is still there on Tuesday. The evidence here is largely observational and cannot fully separate cause from effect, so this is a direction worth acting on, not a demonstrated mechanism.

Relief and resolution, again

A crowd reliably produces relief: the felt state changes, sometimes dramatically. Resolution means the conditions producing the strain have changed. Relief has value of its own and does not need to be justified by anything further. But relief can be repeated indefinitely — the next event is already on sale — while the circumstances stay exactly as they were. That argument is the project’s synthesis of separate research lines, not a finding from a study that tested the loop directly.

The practical version is unglamorous. If a night out leaves you closer to two people you can call, something may have changed. If it leaves you with a good memory and the same Monday, that is still worth having, and it is not the same thing.

Does crying help?

2026-09-14T00:00:00+00:00

What people report

Bylsma, Vingerhoets and Rottenberg (2008) collected retrospective accounts of the most recent crying episode from students in 35 countries. Most participants said their mood improved, a substantial group reported no change, and some felt worse. What separated them was context. A reported benefit was more likely when the person had social support while crying, and when the episode led to a resolution or a new understanding of what triggered it. It was less likely when the person felt ashamed of crying or tried to hold it back.

This is cross-sectional recall data: people describing an episode after the fact, with no comparison condition and no measurement at the time. It documents what crying is associated with in memory, not what crying causes.

What the laboratory finds

The immediate picture looks different. In studies where participants watch sad films, mood is often worse right after crying than in people who did not cry. Gračanin, Bylsma and Vingerhoets (2014) reviewed this apparent contradiction and proposed that any calming effect may emerge with a delay, and that much of the benefit comes from the social responses crying elicits rather than from the crying itself. Retrospective reports may also be shaped by how an episode ended rather than by its average course.

Neither result says crying is useless. Together they say that crying is not a valve that empties a fixed quantity of feeling.

Talking, and the gap between felt and measured

The same gap appears for talking about a distressing experience. Zech and Rimé (2005) ran two experiments with follow-up up to two months in non-clinical participants: those who talked about a negative emotional experience rated the conversation as helpful, yet the emotional impact of the memory was not reduced compared with control conditions.

The natural reading — that sharing helps when it supports cognitive work on the experience rather than re-expression alone — is an interpretation consistent with this research, not something those experiments tested.

Why we do it anyway

Rimé (2009) reviewed a long research programme showing that people share most of their emotional experiences with others, across ages and cultures, and that sharing has interpersonal functions: it draws attention, empathy and support. The evidence comes largely from one research programme and leans heavily on self-report, so the size and generality of these functions are less firmly established than their existence.

That may be the most useful frame for the question. Crying and talking reliably do something social: they signal distress and recruit other people. Whether they also reduce the distress itself depends on what those other people do, and on whether the situation that caused the crying changes.

So the honest answer is conditional. Cry if you need to. Whether it helps is decided less by the tears than by who is there, what is understood afterwards, and what changes.

Does exercise improve mood?

2026-09-14T00:00:00+00:00

An acute lift, and how it was established

Reed and Ones pooled studies of single bouts of aerobic exercise and found that one session is associated with a short-term increase in positive activated affect: the energised end of mood, feeling alert and awake, rather than the calm end. A meta-analysis is the right design for asking whether an effect turns up consistently across many small samples. It is a weak design for asking how long the effect lasts in any one person, or which people it fails for.

Two limits belong with the finding. The effect is acute and fades, so it describes the hours after a session and not the weeks around it. And this site does not state the pooled effect size or the number of studies behind it, because the source abstract could not be re-verified — a gap in our checking, not a hidden number.

Organised dance is a different question

Koch and colleagues meta-analysed 41 controlled intervention studies of dance movement therapy and dance, and found a medium overall effect on health-related psychological outcomes (d = 0.60). That is a stronger design than a single trial, because it compares programmes against control conditions across many settings.

Three qualifications travel with the number. Heterogeneity between studies was high, which means the true effect is unlikely to be the same size everywhere. Effects were larger when outcomes came from observational measures rather than self-report, and that pattern may indicate bias in how outcomes were rated. And the trials studied structured programmes, many of them in healthcare or preventive settings, with a facilitator and a schedule. They are not evidence about a night of dancing for its own sake, which nobody in these studies was doing.

What the lift is not

An hour of running can change how a difficulty feels by the end of it. It does not, on its own, change the difficulty. This site keeps the two apart: relief changes the felt state, resolution changes the conditions that produce it. Relief has value of its own, and treating it as worthless is its own mistake.

The reason to keep them apart is that relief can be repeated. Hayes and colleagues describe experiential avoidance, the pattern in which a person works to escape or alter unwanted internal experience and, in doing so, keeps arranging their life around it. A run that reliably makes the afternoon bearable can become the reason the conversation is never had. That is a synthesis drawn from separate lines of research, not a finding from a study that tested the loop directly, and it may not hold for particular people or particular activities.

What would settle it

Studies that follow people for months, measure both mood and the circumstances producing the strain, and compare exercise against another active strategy would settle more than the acute literature can. That work is thin, so for the question of whether exercise changes what a person is dealing with there is insufficient evidence either way. For the question of whether it lifts mood for a few hours, the answer is a fairly confident yes.

Does this site measure my emotions?

2026-09-14T00:00:00+00:00

What the numbers in the artwork are

The artwork borrows the vocabulary of an observability dashboard: input state, synchronisation, root cause, a chart that climbs and falls. None of it is data. Every metric, status value and chart in the artwork, and the relief figures elsewhere on the site, are artistic or conceptual values chosen to express a relationship. They were designed to resemble telemetry, which is exactly why they carry the ARTISTIC label wherever they appear, and why they must never be quoted as findings.

The simulation on the methods page works the same way: it runs on invented inputs and says so on its face. Nothing on this site takes a reading from a visitor, and nothing is inferred about one.

What the site is technically

It is a static site. Pages are generated ahead of time from Markdown and templates and served as files; there is no application collecting anything at the other end. It contains no analytics scripts, no trackers and no cookies.

That claim is the kind that should not be taken on trust, so it is made checkable. The source code is public. You can read the templates and the built HTML, search them for third-party script tags, and open your browser’s network panel while loading a page to see which hosts are contacted. Verification beats assurance, and this is the form of assurance that can be turned into verification.

There is one honest limit. The hosting provider serves the files, and a provider may keep its own server logs — request times, addresses, user agents — as part of operating a web server. That is outside this project’s control and not something it can promise about on the provider’s behalf.

Why a site about emotion refuses to measure

Rosalind Picard framed affective computing as a research field in 1997, and it has produced serious work. It has also produced systems sold on the promise of inferring emotional states from faces, voices and physiological signals. Stark and Hoey’s argument is that such systems are built on contested models of emotion and on proxy data, and that these choices — which model, which proxy — shape the ethical and social implications of what gets built. It is a conceptual argument about the field, not an audit of particular products.

Two reasons follow for this project. The first is validity: the syntheses on measurement do not support inferring a specific emotional state from any single recorded signal, so a dashboard here would be presenting error as precision. The second is consent: someone reading an essay about catharsis has not agreed to be analysed, and an artwork that quietly did so would be enacting the thing it criticises.

So the artwork keeps the interface and drops the measurement. The gauges move because a designer decided they should, and the page says so.

Does writing about it help?

2026-09-14T00:00:00+00:00

What was studied

The paradigm is simple and has been repeated for decades: people write for a short period on several consecutive days about a distressing experience, covering both what happened and how they felt about it, while a control group writes about something neutral (Pennebaker & Beall, 1986). Writing about both the facts and the feelings has been associated with health benefits measured weeks or months later. The studies are mostly in student and other non-clinical samples, and the abstract of the 1986 paper could not be re-verified for this site’s reference registry.

The effects vary considerably across studies, so “writing helps” is an average, not a prediction about one person on one week.

Why it seems to work

The interesting part is the mechanism, because it is not discharge. Pennebaker (1997) linked the benefits to changes in how people construct an account of the experience across sessions rather than to the intensity of emotion expressed in any one of them. What appears to matter is that a jumbled experience becomes something that can be told: with an order, a cause and a place in the rest of a life.

That reframes writing as a form of sense-making rather than a form of release. It also predicts something the folk theory does not: that the sessions where the most is felt will not necessarily be the sessions that help most.

Why it often feels worse first

Short-term distress commonly rises after writing about something painful. That is a normal feature of the procedure rather than a sign of harm, but it does mean the feeling immediately afterwards is a poor guide to whether the exercise is doing anything, in either direction.

It also sets a limit. Expressive writing is not a treatment for severe or recent trauma, and using it that way, without support, is not what the evidence describes.

Writing versus going over it again

Writing is not automatically better than thinking, and the difference is in the kind of thinking. Watkins (2008) argued that repetitive thought is not uniformly harmful: abstract, evaluative forms — why does this always happen to me — tend to be unconstructive, while concrete, process-focused forms — what exactly happened, and what could I do next — can support problem solving and recovery.

Writing tends to push toward the concrete, because a sentence has to name particulars that a circling thought can leave vague. But a page can rehearse an abstract complaint as easily as a mind can. The specific advice to switch to concrete next steps has not been tested here as a stand-alone intervention; it is drawn from the distinction between the two kinds of repetitive thought.

A practical reading

If you write, write about what happened and about how it felt, in enough detail that someone else could follow it, over a few short sessions rather than one long one. Expect to feel worse before anything else, and judge the exercise by whether the account is clearer a week later, not by how much came out on the page.

How do you grade the evidence?

2026-09-14T00:00:00+00:00

Two steps, two different questions

Grading happens in two steps, and they answer different questions.

First, each source is graded by design. A meta-analysis or systematic review of controlled studies is level A. One or more randomised or controlled experiments, usually with short-term outcomes and specific samples, is level B. Prospective, longitudinal or quasi-experimental studies, including uncontrolled pre-post comparisons, are level C. Cross-sectional, correlational or qualitative studies, narrative reviews and theoretical work are level D. Classical and historical texts, expert opinion, project records, policy statements and artistic material are level E.

One consequence is worth spelling out: a synthesis is graded by what it pools, not by the word “meta-analysis” in its title. A meta-analysis of correlational studies is graded by the design of those studies.

Second, each claim receives an evidence level no stronger than its strongest source, and a separate confidence rating. The level describes the kind of study behind the claim. Confidence describes how well the claim, as worded, is supported.

The four confidence values

HIGH — the claim, as worded, is well supported and further research is unlikely to reverse it.
MODERATE — the claim is supported, but results are limited, heterogeneous, disputed or not fully re-verified.
LOW — the claim rests mainly on theory or indirect evidence and could change with new research.
UNKNOWN — the available evidence is insufficient to judge the claim.

Because the two ratings are independent, they come apart in useful ways. A meta-analysis with heterogeneous results can support a level A claim held at moderate confidence. A well-replicated correlation can be level D and still justify high confidence, provided the claim speaks of association and not of cause. That proviso is where most of the editorial work happens: the same data supports a strong claim about association and a weak one about cause.

What a ledger entry contains, and what expires

Each entry records the claim’s wording in both languages, its type, its evidence level, its confidence, its sources, its scope, its most important caveat, the pages where it appears, and the dates of its last and next review. Bibliographic facts are kept separately in the reference registry, checked against Crossref, because a citation and a judgement about a source go out of date on different schedules.

Review intervals depend on the claim type. Clinical-boundary claims are reviewed at least every 180 days, empirical claims every 365 days, and theoretical, historical, technical and artistic claims every 730 days. An automated check runs every week and on every change; a claim or source past its review date fails the build until someone re-examines it and records the outcome in the evidence changelog.

That is the part that makes the rest more than a promise. A stale claim does not sit quietly on a page — it stops the site from being published.

Art is labelled, not graded upward

The site is an artwork as well as a library, so the ledger labels statements by type. A claim marked ARTISTIC belongs to the artwork and expresses a thesis; it is never a measurement of anyone. A claim marked EMPIRICAL reports what studies observed, with sources and limits. Keeping the labels visible protects both sides: the artwork can use the language of dashboards without its numbers being read as data, and the research notes are not lent false authority by the artwork’s rhetoric.

If a claim looks outdated, overstated or wrongly sourced, the ledger page carries a report link.

How was this site built, and how long did it take?

2026-09-14T00:00:00+00:00

What the records say

The site is a bilingual static publication generated with Zola, deployed to GitHub Pages and verified against the live URL on every change. It was produced in a single working session by an AI coding agent (Claude Code) directed by one person, with Majordomus as the supervisory layer between the agent and the repository.

All the times below are UTC on 14 September 2026 and come from the Git history, the GitHub API, the Majordomus ledger and the pipeline records. The working session began at 09:45 with the poster rebuild. The GitHub repository was created at 09:54:06. The Majordomus layer was initialised and the first supervised task started at 10:26:23. The site was committed at 10:42:45. The Pages pipeline ran 7 jobs in 2 minutes 43 seconds, all passed, and release v0.1.0 was published at 10:45:42 after live verification.

The derived durations are: repository creation to first verified release, 51 min 36 s; Majordomus initialisation to first verified release, 19 min 19 s; push to verified release, 2 min 57 s. At that release the site was checked by 7 validation stages and 30 browser tests.

What the person supplied

The work was driven by about fifteen messages from one person, of five kinds: a one-line request to create a project on GitHub; art direction for the poster, followed by a choice between three offered approaches; three structured execution briefs of increasing scope; short directives during work such as deploy, validate in a browser and release automatically; and editorial direction, including replacing promotional copy with serious, sourced psychology.

None of those messages specified implementation details such as template syntax or test cases. Those were the agent’s decisions, and the supervision layer’s job was to keep them checkable.

What supervision caught

Majordomus did not write code. It scoped and recorded the work, and three defects became local findings rather than failed pipeline runs. At 10:29:16 scope drift was reported after files were moved, and the task was restarted. At 10:39:39 a malformed scope was reported by the pre-push check, and the task was restarted before any push. Before the first commit, the health check that runs in the pre-commit hook reported that the README did not link the agent bootstrap file, and the defect was fixed before it reached the repository.

One mechanism matters more than the rest: a task cannot be closed as completed unless its verification command passes. Here that command was the live smoke test, so “completed” meant “deployed and checked”, not “written”. The task was closed at 10:47:25 with that test passing in 7 seconds.

Why no speed multiple is claimed

There was one session and no comparable session without supervision, so there is no control condition and no speed-up factor can be stated. That is the honest limit, and it is recorded in the ledger alongside the timeline itself.

What the records do show is where time was not lost. No push was rejected by CI, no deployment had to be rolled back, and the first pipeline run passed all seven jobs. Three further limits apply: the timeline starts at the first recorded file, so planning before it is excluded; durations depend on the tools, hardware and network latency of this session; and the session also relied on conventional safeguards unrelated to Majordomus, whose effects are not separated out.

Is this medical advice?

2026-09-14T00:00:00+00:00

The short version

No. The research notes and advice on this site are general education based on published research. They are not medical advice, they cannot assess an individual situation, and they do not replace assessment or treatment by a qualified professional. That is a statement about the purpose of the site, not an empirical finding, and it applies to every page without exception.

What general education can and cannot do

The library reports what studies observed, in whom, with which design, and what the results cannot establish. Almost all of it is group-level: an average difference between conditions, an association across a sample, a pooled effect across trials. A group average describes a distribution, not the next person to read it. People inside those samples differed widely from one another, and a person outside them may differ more.

So a finding can tell you what tended to happen to people studied under particular conditions. It cannot tell you what is happening to you, what caused it, or what would help. Those are judgements that depend on history, circumstances, physical health and the things nobody wrote down, and they are made by a professional who can ask questions and follow up.

This is why the advice entries are written as practices to try rather than as prescriptions, why each carries explicit limits, and why none of them promises an outcome. Where the evidence is thin, the pages say so, because “insufficient evidence” is a finding about the literature rather than a verdict on your experience.

What the site does not do

Nothing here measures, profiles or assesses you. The artwork’s telemetry, its status line and its charts are artistic values chosen to express a relationship; the interactive model on the methods page runs on invented inputs and says so. No page infers anything about a visitor, and no number displayed anywhere is a reading of a person.

The site also does not recommend, rank or endorse any particular treatment, provider, medication or service. Where research on a treatment is cited, it is cited as evidence about a group of trials, with the quality of those trials stated.

When to involve someone qualified

Distress that persists, disrupts daily life, or includes thoughts of harming yourself warrants contact with a qualified professional. Immediate danger warrants contacting local emergency services. This is standard safety guidance rather than a finding, and the details matter locally: the thresholds at which services intervene, and the services that exist at all, differ by country.

Reading is not a substitute for that step, and it is not a test you have to pass before taking it. If you are unsure whether what you are experiencing is serious enough, that uncertainty is itself a good reason to ask someone qualified rather than to keep reading.

Is venting anger healthy?

2026-09-14T00:00:00+00:00

What the experiments found

Bushman, Baumeister and Stack (1999) gave participants a message claiming that hitting a punching bag relieves anger, a message claiming the opposite, or a neutral one, then provoked them. Those who had read the pro-catharsis message were more inclined to choose the bag, and those who vented were subsequently more aggressive toward the provocateur, not less. Believing in catharsis increased the wish to vent without reducing later aggression.

Bushman (2002) compared three responses to a provocation: hitting a bag while thinking about the person who caused the anger, hitting a bag while thinking about fitness, and sitting quietly. The first group reported the most anger and behaved most aggressively; the group that did nothing behaved least aggressively. The proposed mechanism is rehearsal. Exertion raises arousal, and if attention stays on the grievance, that arousal is experienced as anger.

What the meta-analysis adds

Single experiments have particular procedures, so Kjærvik and Bushman (2024) asked the same question across the literature: 154 studies, 184 independent samples, 10,189 participants, sorted by whether an anger-management activity raises or lowers physiological arousal.

Arousal-raising activities, including hitting a bag, jogging and cycling, showed no reduction in anger and aggression overall (g = -0.02, 95% interval -0.13 to 0.09), and those studies were heterogeneous. Arousal-lowering activities, such as slow breathing, relaxation and mindfulness, reduced anger and aggression (g = -0.63, 95% interval -0.82 to -0.43), and the result held across students and non-students, offenders and non-offenders, people with and without intellectual disabilities, and group, individual, digital, laboratory and field formats.

A null pooled average is not evidence of harm, and it concerns anger and aggression only, not fitness, mood or social connection, where exercise has other documented benefits.

Rehearsing versus reinterpreting

The problem is less the expression than what attention does during it. Ray and colleagues (2008) found that, relative to reappraisal, ruminating on an anger-provoking event produced a stronger experience of anger, more cognitive perseveration and greater sympathetic activation. Denson (2013) describes angry rumination as a risk factor for aggression that temporarily reduces self-control. Both are short-term laboratory findings, and the self-control account comes from a theoretical model rather than a direct test.

Read together with the meta-analysis, the picture is consistent: what keeps anger active is going over the provocation, and venting is a way of going over it with the body involved.

When anger is more than a bad evening

None of the above is a treatment. For adults whose anger is established by standardised measures, a meta-analysis of 23 controlled studies found medium to large treatment effects, with cognitive therapies supported for driving anger, suppressed anger and trait anger, and relaxation for state anger (Del Vecchio & O’Leary, 2004). Those studies run to 2002 and included few treatment-seeking participants, and they describe professional treatment rather than self-help.

If anger is frequent, intense, or is damaging relationships, work or safety, that is a reason to talk to a clinician rather than to work through a list.

If relief is temporary, is it worthless?

2026-09-14T00:00:00+00:00

Two different changes

Relief is a change in how a state feels. Resolution is a change in the conditions that produce the state. From the inside they are hard to tell apart, because in both cases the person feels better afterwards. A long run, a good cry, an evening of loud music and a difficult conversation can all end in relief. Only some of them also end in resolution.

Saying so is not a verdict against relief. Recovering from an activated state is valuable on its own terms, and a person who is calmer is usually better placed to deal with whatever is waiting. The claim is only that the feeling does not report on the situation.

Why relief tends to repeat itself

Learning theory supplies the first half of the reason. A behaviour that removes an aversive state is negatively reinforced: it becomes more likely to happen again in similar circumstances. That is adaptive when the behaviour also addresses the source. When it does not, each repetition is still rewarded, and the source is still there.

Hayes and colleagues (1996) described experiential avoidance, the attempt to escape or suppress unwanted internal experience even at long-term cost, and argued that it is a functional process shared across many psychological difficulties. Applying that account to a night out or a collective release is an inference rather than a tested result, and not every cathartic experience is avoidance. It does suggest a question worth asking: is the state being worked through, or postponed?

Why relief is remembered generously

The second half of the reason is memory. Fredrickson and Kahneman (1993) found that retrospective evaluations of an experience are dominated by its most intense moment and by how it ends, while its duration is largely neglected. In the best-known demonstration, most participants chose to repeat a longer episode of cold-water discomfort that ended slightly less painfully over a shorter one that did not (Kahneman et al., 1993).

Applied to a restorative evening, the rule predicts that a strong peak and a good ending will make the whole episode remembered as more helpful than its lasting effect justifies, which makes it attractive to repeat. That application is an extrapolation from short laboratory episodes, and the pattern is not universal: Rode and colleagues (2007) found, across three studies of remembered meals, consistent duration neglect but no peak, primacy or recency effects. One domain of pleasant experience, not a test of painful ones, but enough to show that the rule has limits.

The useful question afterwards

None of this recommends giving up what brings relief. It recommends a second question alongside “did it help?” — namely, “what exactly changed?”

If the answer is how you feel, that is relief, and it counts. If the answer also includes what you now understand, what you have decided, or what you will do differently, something has begun to resolve. The two can coexist in the same evening; they just need to be counted separately.

No study cited here tested this whole loop directly. It is a synthesis of separate lines of research, and for particular people or particular activities it may not hold.

What helps instead of venting?

2026-09-14T00:00:00+00:00

Changing how the situation is read

The most studied alternative is reappraisal: changing the interpretation of a situation rather than trying to push the feeling down. In Webb and colleagues’ synthesis (2012), instructed cognitive change had a small-to-medium effect on emotional outcomes, while attempts to suppress the experience of an emotion were not effective. Separately, people who habitually reappraise report more positive and less negative emotion than people who habitually suppress (Gross & John, 2003), though that comparison is correlational and cannot establish direction.

The average conceals variation. Effects differ by emotion, by the specific tactic and by the person, and the laboratory inductions behind them are short.

Stepping back from the memory

Self-distancing means reflecting on an upsetting experience as an observer rather than reliving it from inside. Kross, Ayduk and Mischel (2005) found across two experiments that a self-distanced perspective on an anger-eliciting memory reduced negative affect compared with immersed recall.

A later synthesis by Moran and Eyal (2022) pooled 230 effects and found that a distant perspective attenuated emotional experience (g = 0.52), with smaller effects for self-conscious emotions than for basic ones; addressing oneself in other than the first person enhanced distancing and went with less distress under social stress. Mechanisms remain unclear and terminology differs across studies, so the size of the benefit for a given person and emotion is not settled. Outcomes in these studies are short-term, and recent or severe trauma calls for professional support rather than a self-help exercise.

Putting the feeling into words

Lieberman and colleagues (2007) had participants label the emotion shown in photographs of faces. Labelling was associated with reduced amygdala response and increased activity in the right ventrolateral prefrontal cortex, a pattern the authors read as affect labelling acting as implicit regulation.

The task used standardised images, not personal events, and it did not follow recovery over time. The everyday advice to name what you are feeling extrapolates from the mechanism rather than resting on a test of the advice itself.

Where these strategies stop

Reappraisal is not adaptive everywhere. Troy and colleagues (2013) found higher reappraisal ability associated with lower depression under uncontrollable stress but with greater depression under controllable stress, and Ford and Troy (2019) argue that people often cannot reappraise successfully and that reappraising to feel better is not always functional. One correlational study and a conceptual review do not overturn the experimental evidence that reappraisal reduces emotional responses; they place a condition on it. When a situation can be changed, changing it is not a second-best option.

And these are self-help strategies with small averages, not treatments. For adults with anger established by standardised measures, a meta-analysis of 23 controlled studies found medium to large effects for structured treatment, favouring cognitive therapies for driving anger, suppressed anger and trait anger and relaxation for state anger (Del Vecchio & O’Leary, 2004). Those studies run to 2002 and describe professional care.

A pooled average also does not predict what will happen for one person on one evening. The practical value of these strategies is that they change the relationship to an emotion instead of rehearsing it.

What is catharsis?

2026-09-14T00:00:00+00:00

Where the word comes from

Aristotle’s Poetics describes tragedy as achieving, through pity and fear, the katharsis of such emotions. The passage is brief, Aristotle does not define the term further, and commentators still disagree about whether he meant purgation, purification or clarification. No reading is settled, which is worth knowing before the word is used as though its meaning were obvious.

The clinical sense arrived much later. Breuer and Freud’s cathartic method sought relief from symptoms through the discharge of emotion, which they called abreaction. Freud later abandoned the method, in part because the relief it produced was often temporary.

What it means on this site

On this site, catharsis means a subjectively felt release of emotional tension after expression, exertion or immersion. Whether the cause of the tension changes is treated as a separate empirical question, not as part of the definition.

This is a working definition adopted by the project, not a consensus definition in psychology. Other authors use the word for purgation, for clarification, or as a synonym for venting, so findings stated in their terms do not map one to one onto this one.

Why the definition is doing work

Keeping the release apart from its effects is what makes the rest of the library possible. A concert, a good cry, an argument or a long run can all produce release. Each has been studied separately, and the answers differ: some of these leave a person calmer afterwards, some leave them more activated, and none of them, on their own, has been shown to change the conditions that produced the tension.

If catharsis is defined as “expression that resolves”, then every case where relief fades looks like a failure of the person. If it is defined as felt release, the fading is simply information about what release does.

How do I know when to seek professional help?

2026-09-14T00:00:00+00:00

The plain answer first

If you are in immediate danger, contact local emergency services now. That is not a threshold to deliberate over.

Short of that, distress that persists, disrupts daily life, or includes thoughts of harming yourself warrants contact with a qualified professional. This is standard safety guidance rather than a research finding, and it is deliberately general, because the thresholds at which services intervene and the services that exist at all differ by country. Looking up what is available where you live, before you need it, is worth the ten minutes.

Why duration beats intensity as a signal

People tend to judge whether something is serious by how bad the worst hour felt. That is an unreliable instrument. Very intense distress can pass within a day, and distress that never reaches a peak can quietly reorganise a life over months.

Two better questions are how long it has gone on and what it has cost. If sleep, eating, work, or contact with other people have changed and stayed changed, that is information independent of how the feeling ranks on a scale. So is finding that things which used to provide relief no longer do.

None of this is a test, and none of it produces a verdict. It is a way of noticing that you already have enough grounds to ask, rather than waiting for a feeling dramatic enough to justify asking.

What the research adds about repetitive thinking

Among the emotion-regulation strategies studied, rumination shows the largest association with symptoms of anxiety, depression, eating and substance-related disorders, and it predicts the onset of depression. The pooled data are largely correlational and cannot show that rumination causes symptoms, so this is a reason to take repetitive brooding seriously as a marker, not evidence that stopping it would resolve anything on its own.

The trauma literature points the same way. Across 75 studies, posttraumatic stress symptoms are associated with rumination (r = .52), experiential avoidance (r = .48) and thought suppression (r = .29). These are associations rather than causes, and the few longitudinal studies suggested somewhat smaller associations.

Read together, they suggest a pattern worth watching in yourself: going over the same ground repeatedly without moving, and steering around the situation, the place or the person rather than through it. Neither pattern assesses anyone, and neither means something is wrong. They are reasons to raise it with someone who can ask follow-up questions.

What asking actually involves

A first conversation with a general practitioner, a psychologist or a local service is usually a description of what has changed and for how long. You do not need to have a name for it, and you do not need to have tried self-help first. If you are unsure whether it counts as serious enough, that uncertainty is a reason to ask rather than a reason to wait.

The content of this site is general education, not medical advice. It cannot assess your situation, and nothing on it replaces that conversation.

Why an artwork and not just a library?

2026-09-14T00:00:00+00:00

What the artwork says that a paragraph cannot

The library can report that relief and resolution come apart. The artwork puts a reader inside the gap. It presents catharsis in the form the contemporary world actually delivers it in — a service, with a status line, a dashboard, a completed transaction — and then leaves one field unchanged.

That field, problem_solved: false, is the whole argument in one line. It is an artistic statement of the project’s thesis that relief is not resolution: a transaction can succeed completely while the request stays unresolved. Every indicator can go green and the situation that produced the tension can still be exactly where it was.

Said as a sentence, that is a claim to be weighed. Delivered as a successful response with one stubborn field, it is closer to a recognition. A reader who has ever left a good cry, a shouting match or a long run feeling lighter and finding nothing changed does not need the sentence explained.

The numbers are art, and must not be quoted

Every metric, status value and chart in the artwork, and the relief figures on the site, are artistic or conceptual values chosen to express a relationship. None of them is a measurement of anyone. They are deliberately designed to resemble telemetry, which is precisely why the warning has to be explicit: they must not be quoted as data.

This runs in both directions. The artwork’s numbers do not support any empirical statement, and the empirical statements on the site do not borrow authority from the artwork’s air of instrumentation. Nothing measures, profiles or assesses a visitor; the interactive model on the methods page runs on invented inputs and says so.

problem_solved: false is an interpretive device as well. It is not a finding that collective release never helps. Whether release helps, when and for how long, is an empirical question handled in the research notes and graded separately from anything the artwork asserts.

Keeping them apart on separate sites would have been safer and would have lost the point. The reason a false promise about catharsis is worth writing about is that it arrives wrapped in the aesthetics of measurement: apps, scores, dashboards, a sense that an inner state has been handled. Criticising that register from outside it is weaker than occupying it.

So the site does both, and marks which is which on every page. Claims belonging to the artwork carry the ARTISTIC label in the evidence ledger; claims reporting what studies observed carry EMPIRICAL, with sources, scope and limits. The labels are visible rather than implied, and they are enforced by the same checks that grade everything else.

The result is a publication that can use the language of dashboards without its numbers being read as data, and a library that can be read without the artwork’s rhetoric lending it false authority. The artwork makes the question felt. The library answers it, and says how confidently.

Why does dancing with other people feel so good?

2026-09-14T00:00:00+00:00

Moving in time with someone changes how you regard them

The cleanest evidence comes from experiments that strip dancing down to almost nothing. Hove and Risen had participants tap in or out of time with an experimenter; Wiltermuth and Heath had groups walk or sing in step. Synchrony increased liking for a partner and increased later cooperation, including cooperation that cost the participant something personally. Because the synchrony was assigned rather than chosen, these are experiments, not correlations: the direction of the effect is not in doubt.

What they do not settle is size and reach. Tapping a finger to a metronome is minimal compared with a room of people dancing, the groups were small, and the outcomes were measured shortly afterwards. Rennung and Göritz pooled 60 such experiments and found a medium effect of interpersonal synchrony on prosocial attitudes and behaviour — with the effect moderated by experimenter effects and by intentionality, which means part of it depends on how the studies were run. That is a real caution, not a footnote.

Synchrony and effort contribute separately

Tarr and colleagues ran a group-dance experiment with a 2 × 2 design, varying synchrony and physical exertion independently. Each raised pain thresholds, an indirect proxy for endorphin activity, and each increased in-group bonding. The design matters: it shows the two ingredients are not the same thing, and that a hard, unsynchronised session and an easy, synchronised one each do something.

The proxy matters too. Endorphin release was not measured directly — a raised pain threshold is consistent with it and does not single it out. This is one experiment, and one experiment needs replication before the mechanism is treated as established.

The music is doing its own work

Salimpoor and colleagues studied listeners with music they had chosen themselves, using PET and fMRI. Peak emotional responses — the shiver at a particular moment — were accompanied by striatal dopamine release, with activity in the caudate more involved in the anticipation and the nucleus accumbens in the peak itself. Anticipation and arrival register differently, which is an appealing match for how a build-up and a drop are constructed.

The match is an analogy, not a finding. Electronic dance music was not studied, nor were crowds; the listeners were alone in a scanner with headphones. A small imaging study establishes that something is going on in reward-related circuitry during musical peaks. It does not license a story about what a dancefloor does.

What the whole picture leaves open

Put together, the evidence says the pleasure of dancing with others is not mysterious: synchrony, exertion and music each contribute something that can be measured. Every one of those measurements was taken during the activity or minutes to hours after it. Whether any of it reaches the next morning is a separate question, and the studies cited here were not designed to answer it.

Methods: signals, inference and the limits of detecting a state change

2026-09-14T00:00:00+00:00

Why this page exists

The artwork on this site presents collective catharsis as a service, complete with telemetry: input state, synchronization, discharge, relief. Its endpoint answers 200 OK with problem_solved: false. The telemetry is conceptual. No number on the poster or in the dashboard was recorded from a person, and Research note 05 explains why the artwork refuses to measure its audience.

That refusal raises a fair question. If someone did want to know whether a catharsis-like change of state had taken place, in a concert crowd, in a therapy session or in a laboratory, what could real measurement support? This page answers as carefully as the evidence allows. It describes how signals from the body and from self-report could be turned into a statement about a change of state, where each step introduces error, and what no combination of signals can settle.

Two commitments shape everything below. First, measurement and interpretation are kept apart: a sensor records skin conductance, and a person or a model concludes something about arousal. Second, nothing on this site measures anyone. The site collects no physiological data, runs no analytics and sets no trackers. The interactive part of this page is a simulation that works only with values you invent.

The detection model: signal, feature, inference, context

Any claim of the form “this person went through a release” rests on a chain of four steps, shown in the chain below.

A signal is what an instrument records: voltage from electrodes, light reflected through the skin of the wrist, acceleration, sound pressure. A signal has units and a sampling rate, but no psychological meaning.

A feature is a quantity computed from the signal: mean heart rate over a minute, the number of skin conductance responses above a threshold, the correlation between two people’s movements. Every feature embodies choices about the time window, the filter, the threshold and the way artefacts are removed. Those choices are rarely visible in the final number, yet they change it.

An inference maps features onto a description of a person: elevated arousal, a probable phase of an episode, a likely emotional category. This is where measurement becomes interpretation. Whether a trained classifier or a researcher looking at a plot performs it makes no difference to its logical status.

Context is everything known about the situation that is not in the signal: what the person was doing, where, with whom, what preceded the recording, and what the person says about it. Context often decides between interpretations that the features alone cannot separate.

Error enters at every step, and it compounds. A movement artefact in the signal becomes a spurious spike in a feature; the spike becomes an inferred “response”; and a missing context record allows that response to be interpreted as fear rather than as someone bumping into the sensor. Later steps cannot recover information that earlier steps lost. They can, however, make a weak signal look authoritative by giving it a confident label.

Baseline: nothing means anything without one

A heart rate of 95 beats per minute is unremarkable for one person who is walking and notable for another who is sitting still. Physiological features acquire meaning only relative to a baseline: the range of values typical for that person under comparable conditions.

Useful baselines are within-person. They are estimated from repeated recordings of the same individual, ideally at similar times of day, in similar postures and at similar levels of activity. Many ordinary factors move physiological signals without any emotional change: the circadian rhythm, standing rather than sitting, a recent meal, caffeine and nicotine, alcohol, medication, lost sleep, illness, ambient temperature and humidity, and the simple anticipation of being measured. Skin conductance, for example, responds to temperature and exertion as well as to psychological events, and methodological handbooks on electrodermal activity treat recording conditions, electrode placement and individual differences as central to its interpretation (Boucsein, 2012).

Heart rate variability is widely used in psychophysiology because it indexes cardiac vagal tone, the parasympathetic contribution to the regulation of the heart, and because it is inexpensive, non-invasive and easy to record. Laborde, Mosley and Thayer (2017) nonetheless warn that this ease of access should not obscure how easily HRV findings are misconstrued, and they set out recommendations for planning experiments, analysing data and reporting results so that conclusions are sound and comparable across laboratories.

Habituation adds another complication. Responses to a repeated stimulus tend to weaken, so the tenth loud chorus of an evening may produce a smaller skin conductance response than the first even when the listener describes both moments as equally intense. A falling response over time is therefore consistent with a fading experience, but also with an unchanged experience and a habituating body.

Group norms mislead for a related reason. A population average describes what is typical across people, not what is typical for this person. Resting heart rate, skin conductance level and heart rate variability differ with age, fitness, body size, health and medication. Applying a population threshold to an individual turns ordinary differences between people into apparent states, and it systematically mislabels those furthest from the average.

Multimodality: more signals, not more certainty

If one signal is ambiguous, it is tempting to combine several. Multimodality does help in some respects. Sensors with independent sources of artefact can cross-check each other: an accelerometer can show that a rise in heart rate coincided with running, and a microphone can show that faster breathing coincided with singing. Combining modalities reduces some errors, especially technical ones.

It does not remove the central ambiguity, for two reasons.

The first is that confounds are often shared across signals. Dancing in a warm room raises heart rate, lowers heart rate variability, increases skin conductance, speeds up breathing and produces movement, all at once. Five signals that agree are then five views of one cause, not five independent pieces of evidence about an emotion.

The second is that the physiological patterns themselves may not be specific to emotion categories. Siegel and colleagues (2018) meta-analysed 202 studies that measured autonomic nervous system reactivity during laboratory inductions of emotion in non-clinical adult samples. Autonomic variables did change, but the pattern of effect sizes did not clearly distinguish one emotion category from another. Variation within categories was considerable, and correcting for publication bias reduced the estimated effects further. The authors concluded that the findings were more consistent with emotion categories as variable populations of instances than with fixed autonomic “fingerprints”.

The question is genuinely debated. An earlier review by Kreibig (2010), covering 134 experimental publications, concluded that autonomic responding showed considerable specificity when subtypes of distinct emotions were considered, and stressed that precise terminology for the states studied and a careful choice of physiological measures matter for such conclusions. The two reviews differ in method and in the questions they ask. For the purpose of this page, what they have in common matters more: both describe patterns estimated across many participants, under controlled inductions of defined states. Neither offers a pattern that could be applied to one person in an uncontrolled setting to establish what that person feels.

Facial movement, often added as “one more modality”, has the same limitation. Barrett and colleagues (2019) found that people do smile, frown or scowl in the expected emotional situations more often than chance would predict, but that the way emotions are expressed varies substantially across cultures, situations and individuals, that similar facial configurations express more than one emotion category, and that a configuration such as a scowl often communicates something other than an emotional state.

Behind all of this lies the problem of reverse inference. Poldrack (2006) described it for neuroimaging: concluding that a mental process was engaged because a region associated with that process was active. Such inferences are not deductively valid, although they can add some evidence, and their value is limited by how selective the observed activity is. The same logic applies to bodily signals. Arousal rises during fear, but also during joy, effort and heat. The less selective a signal, the less its presence says about any single state, and adding more unselective signals does not make any of them selective.

The detection sandbox makes this chain inspectable. It labels each part it shows as observed, derived, inferred or unknown, ranks competing readings of the same synthetic signals with the arithmetic beside them, and shows how the setting, which no sensor records, changes which reading leads.

Self-report: indispensable and imperfect

Self-report is the only access to how an experience feels from the inside. Physiological data can show that something changed; only the person can say whether the change felt like release, dread, boredom or nothing in particular, and whether its valence was pleasant or unpleasant. Any serious model of catharsis therefore needs self-report, and it needs to treat self-report as a measurement with errors of its own.

Robinson and Clore (2002) proposed an accessibility model of emotional self-report built on a distinction between emotion itself, which is episodic, experiential and tied to context, and beliefs about emotion, which are semantic, conceptual and detached from context. The distinction helps explain the discrepancies that often appear between reports of feelings a person is experiencing now and reports of feelings they are not currently experiencing, such as how they felt last week or how they usually feel at concerts. When the experience is no longer present, a report can draw more on what the person believes such situations feel like.

Retrospective reports are also shaped by the structure of the episode. Fredrickson and Kahneman (1993) asked participants to rate film clips continuously and then to evaluate each clip as a whole. The duration of a clip had only a small effect on the overall evaluations, which behaved like a weighted average of “snapshots” of the experience, as if duration did not matter. In a related experiment, Kahneman and colleagues (1993) exposed participants to a short trial of painfully cold water and a longer trial that added 30 seconds of slightly less cold, still unpleasant water. When asked which trial to repeat, a significant majority chose the longer one. Retrospective evaluations of such episodes are often dominated by the worst and the final moments. A night that ends on a high will tend to be remembered as more restorative than its minute-by-minute course would suggest.

Ecological momentary assessment (EMA) addresses part of this problem. It samples people’s current experience and behaviour repeatedly, in real time and in their everyday environments, often at random moments, using tools that range from written diaries to electronic devices and physiological sensors. Its aim is to minimise recall bias and maximise ecological validity (Shiffman, Stone & Hufford, 2008). EMA does not remove every source of error. Being asked repeatedly can change what people notice, response options limit what can be reported, and people may answer in ways they consider acceptable, especially when they know someone else will see the data. That last concern, social desirability, is sharpest in exactly the settings where emotional monitoring is most often proposed: workplaces, schools and health services.

The practical consequence is that self-report and physiological data are not a truth and its proxy. They are two imperfect measurements of different aspects of an episode, and disagreement between them is information, not noise to be averaged away.

The temporal model

Catharsis, as the artwork describes it, is a sequence rather than a state. The temporal model in the table below divides a catharsis-like episode into seven phases: baseline, activation, synchronization, peak, discharge, recovery and post-state. For each phase it lists what could in principle be recorded and why that observation does not establish the interpretation.

The model is a heuristic for organising observations, not an empirical finding. It borrows its framing from research on emotion dynamics, which studies the trajectories, patterns and regularities with which the experiential, physiological and behavioural elements of emotion fluctuate over time, the processes behind those fluctuations and their consequences for well-being (Kuppens & Verduyn, 2017). That framing is useful because it treats an emotional episode as a set of time series rather than as a single label. It does not imply that every episode passes through these seven phases, in this order, or at all.

Three cautions apply to the whole table. Phases are identified after the fact: a peak is a peak only once values have fallen. The physiological, behavioural and reported versions of a phase can unfold on different timescales and need not line up. And synchronization applies only to shared settings; a person crying alone may show activation, a peak and discharge with no synchronization at all.

Synchronization deserves a closer look because it is the phase most specific to collective catharsis. Across three experiments, people who had acted in synchrony with others cooperated more in subsequent group economic exercises, even when cooperation required personal sacrifice, and positive emotions did not need to be generated for this effect to appear (Wiltermuth & Heath, 2009). In group dance, synchrony and exertion each independently raised pain thresholds, used as a proxy for endorphin activity, and increased in-group bonding (Tarr et al., 2015). These results support treating shared movement as a meaningful social variable. The first of them also shows why it cannot serve as a readout of feeling: the social effect of synchrony did not depend on positive emotion. Research note 03 discusses this work in more detail.

Relief detected is not cause resolved

The artwork’s endpoint answers problem_solved: false. This is not a joke at the expense of relief. It is the most precise statement the telemetry could make.

Suppose every step above worked well: a good baseline, clean signals, sensible features, rich context and a momentary report of relief. The strongest supportable conclusion would be that a pattern consistent with activation followed by a return towards baseline was observed, and that the person reported feeling relieved. That is a statement about a change of state. It says nothing about whether the conditions that produced the distress, such as a conflict, a loss, debt, loneliness or an unsafe home, have changed. Those conditions are not in the signal; they are not measured. Relief and resolution are different quantities, and Research note 04 describes the mechanisms, from negative reinforcement to memory, that make them easy to confuse.

The evidence on anger shows why a visible discharge should not be mistaken for a solution. Kjærvik and Bushman (2024) meta-analysed 154 studies, with 184 independent samples and 10,189 participants, that tested anger management activities. Activities that decrease arousal, such as deep breathing, mindfulness and meditation, reduced anger and aggression (Hedges’ g = −0.63). Activities that increase arousal, such as hitting a bag, jogging or cycling, were ineffective overall (g = −0.02), and their results were heterogeneous. The authors concluded that the findings do not support venting anger or going for a run as effective ways of managing it. An arousal-increasing activity can produce a vivid peak and discharge in a physiological trace while, on this evidence, doing little on average to reduce anger. The advice entry Let anger cool turns this into practice.

An honest model therefore has two outputs where a dashboard has one. The first, a change of state, can be inferred with a stated level of confidence. The second, whether the cause is resolved, is either reported by the person and unverified, or unknown.

Measurement versus interpretation

Consider an invented but realistic heart-rate trace. From a resting value of about 65 beats per minute it climbs over several minutes to around 140, stays there for a while, then falls quickly and settles near 75 over the following quarter of an hour. Skin conductance responses become frequent during the climb and sparse afterwards.

Recorded at a concert, the trace is consistent with dancing, excitement and a crowd moving through a build-up and a release. Recorded during a panic attack, it is consistent with a surge of fear, rapid breathing and gradual calming as the attack subsides. Recorded during an interval run, it is consistent with effort followed by a cool-down, with no emotional episode at all.

The measurement is identical in all three cases. What differs is the interpretation, and what licenses the interpretation is not the trace but the context: the location, an accelerometer showing running or jumping, the person’s own account, their history. Without context the only defensible output describes the measurement, “heart rate rose by roughly 75 beats per minute above baseline and returned close to it”, together with the statement that the cause of the rise is not established.

The three cases share something else. In each of them, a system that labelled the fall in heart rate as “relief” would be using a word that belongs to experience to describe a quantity that belongs to the heart. At the concert, the label might happen to match what the person says. After the panic attack it would be misleading, because the fear may return. After the run it would be a category error.

What each signal can and cannot tell us

The detection matrix below lists ten kinds of evidence that a system for detecting catharsis-like changes might draw on, from heart rate to context records. For each it separates what is recorded, what the recorded pattern is consistent with, what it cannot establish on its own, and the common confounds that produce the same pattern for unrelated reasons.

Two features of the matrix are deliberate. Every “consistent with” cell names something broader than a specific emotion, because the evidence reviewed above does not support tying these signals to specific emotions in individuals. And the last two rows, momentary self-report and context records, are not physiological at all. They are included because the physiological rows are largely uninterpretable without them, and because they have limits of their own: a report is an account, not a verification, and a log of time and place is not an account of experience.

Speaking about uncertainty

A system that describes people should use words whose strength matches its evidence. The vocabulary below is used consistently across this site, and the simulation on this page uses it for its output.

The terms form a rough ladder. “Not measured” and “unknown” mark the absence of any basis for a claim. “Insufficient evidence” marks data that exist but cannot distinguish between interpretations. “Consistent with” and “associated with” mark compatibility and statistical relationship without implying cause or applicability to an individual. “Suggests” marks convergence after alternatives have been considered. “Inferred with low confidence” labels a model output for what it is. “Contradicted by” records a conflict between sources, which should prompt scrutiny of the model before any conclusion about the person.

What the vocabulary leaves out matters as much. It contains no term that would let a signal settle what someone feels, and no term for a model having access to someone’s experience, because no method described on this page could support either claim.

False positives and false negatives

Every classifier makes two kinds of mistake. A false positive reports a state that is not present; a false negative misses a state that is. Both are described with examples below.

It is tempting to summarise a system with a single accuracy figure. That hides two facts: the two errors rarely cost the same, and their costs depend on who uses the output and for what. Consider two labels a monitoring system could attach to a person after an intense evening.

A false “recovered” label, which is a false negative for continuing distress, can close a conversation that should have continued. A friend, a clinician or a support service that trusts the label may not ask the question that would have revealed that the person is not all right. The cost falls on someone who needed help and did not receive it.

A false “distressed” label, a false positive, can trigger unwanted attention, concern or intervention. In a workplace or a school it can attach a psychological description to someone who never disclosed anything, with consequences for how they are treated. The cost again falls on the person described, and it grows when the institution holds power over them.

Lowering a decision threshold to catch more real distress produces more false alarms; raising it to avoid false alarms misses more distress. No setting eliminates both. Choosing the threshold is therefore a decision about whose errors are acceptable. It should be made openly, by people accountable for the consequences and with the people affected involved, rather than tuned quietly for a benchmark.

Base rates matter as well. When the state being looked for is rare in the monitored population, even a system with good sensitivity and specificity will produce many false positives relative to true ones. This follows from arithmetic, not from any particular study, and it is one more reason to be cautious about screening whole crowds, classrooms or workforces.

The context engine

The sections above keep returning to context. It is worth being specific about what a component combining signals with context would require, because the list itself shows why this page describes such a component as conceptual.

At a minimum it would need the situation: where the person is and what is happening around them. It would need the activity: whether they are sitting, dancing, running, speaking or singing. It would need recent history, such as sleep, substances, illness and what happened earlier that day, and a personal baseline built from the person’s own repeated recordings. Above all it would need the person’s own account, collected at the time, in their own words, with the freedom not to answer. And to say anything about resolution rather than relief, it would need information about the conditions the person returns to, which no sensor records and which people have every right to keep to themselves.

Each of these inputs is difficult to obtain reliably, intrusive to collect, or both, and several can come only from the person. Once they are available, most of the interpretive work is being done by the person’s account and the situation rather than by the sensors, and the physiological signals become supporting evidence instead of the source of the conclusion. That is the appropriate relationship, and it is why a context engine is not a shortcut to knowing what people feel.

On this site the context engine is conceptual: a way of reasoning about what honest inference would require. It is not implemented, and no data flow exists that could feed it.

Ethics and privacy

Affective computing, a term introduced by Picard (1997), is the study and design of systems that recognise, interpret and respond to human emotion. Stark and Hoey (2021) developed a taxonomy of the conceptual models of emotion and the proxy data used in digital analysis of emotional expression, and argued that the paradigms computer scientists have developed or adapted should not simply be trusted to produce ground truth about human emotions. How emotion is conceptualised, sensed, measured and turned into data shapes the ethical and social implications of the resulting systems.

Several concerns follow directly from the methods described above.

Consent. Physiological signals are produced continuously and involuntarily. People cannot decline to have a heart rate in the way they can decline to post a message. Meaningful consent to emotional inference has to be specific, informed and revocable, and it is hard to obtain in crowds and public spaces.

Purpose limitation. Data collected for one purpose, such as fitness tracking, can support inferences about stress, mood or health that the person never agreed to. A responsible system collects only what a declared purpose requires and does not repurpose it.

Undisclosed states. The central risk of emotional inference is not only that it is often wrong, but that it claims access to states people did not choose to disclose. Even a low-confidence inference, once written into a record, can be treated as a fact about the person.

Power asymmetries. The same inference means different things depending on who holds it. Between friends it might prompt a kind question. In an employer’s dashboard, a school’s monitoring system or an insurer’s model it can shape evaluation, discipline or price, while the person described has little ability to see, contest or correct it. Use in workplaces and schools deserves particular scrutiny, because participation there is rarely voluntary in practice.

Validity as an ethical issue. Given the evidence on autonomic and facial patterns, deploying emotion inference with confident labels is not merely a technical overreach. It exposes people to consequences based on claims the methods cannot support.

This site’s own stance follows from these points. It collects no physiological data and no personal data. It uses no analytics, trackers or cookies. The simulation on this page runs entirely in your browser on values you invent, and nothing you enter leaves the page.

Technical architecture and implementation status

The status table below lists the components of this project and labels each one plainly: implemented, simulation, conceptual or not planned. The labels are part of the argument. A project that criticises confident claims about people should not make confident claims about itself.

What is implemented is publishing infrastructure: the static site, the reference registry checked against Crossref, the evidence ledger with review dates, a versioned content API, and a Rust client library and command-line tool for reading published content. None of these components receives data about readers. The browser simulation is a teaching device that applies the reasoning on this page to invented numbers. The context engine exists only as the description above. Physiological signal ingestion, individual emotion inference or profiling, and analytics or tracking of visitors are not planned, and their absence is a design decision rather than a missing feature.

Limits of this page

This page is educational. It is not a clinical guideline, a validated measurement protocol or an engineering specification, and it should not be used to assess anyone’s emotional or mental state. The signals and confounds it lists are illustrative rather than exhaustive, the temporal model is a heuristic rather than a finding, and the worked examples use invented values. Summaries of studies are deliberately brief; readers who need detail should consult the original sources in the reference list, and the evidence overview shows how claims on this site are graded and reviewed.

If you are worried about your own distress or someone else’s, a conversation remains the most informative source available, and for persistent or severe difficulties so does a qualified professional. No sensor replaces either.

Interactive models: what relief, arousal, memory and synchrony do over time

2026-09-14T00:00:00+00:00

Why models, and why they are not predictions

Many of the arguments in this library are about time. Relief is felt now, and the difficulty that produced it may return tomorrow. Anger fades at different speeds depending on what a person does. An evening is remembered for its best or worst moments rather than for its length. A crowd falls into step gradually, and stress that is not recovered from builds up over weeks. Prose describes these processes one sentence at a time; a model lets you watch them unfold and change the conditions.

A model, in the sense used here, is a small set of explicit rules applied repeatedly to numbers. Its value is transparency. Every assumption is written down, so you can see why a curve rises or flattens and which choice produced the effect. If you disagree with an assumption, the disagreement is precise: you can name the rule you think is wrong.

That transparency comes at a price. The models on this page are deliberately simple, and their numbers are invented. They are not fitted to data, they have not been validated against anyone’s experience, and they make no forecast about what will happen to you or to anyone else. When a curve on this page resembles something you have lived through, the resemblance shows that the rules can produce such a pattern. It does not show that those rules are why it happened to you.

How to use these models responsibly

Treat each model as a thought experiment with a calculator attached. Three habits help.

Change one thing at a time. Each model has two or three inputs, and the clearest lessons come from holding the others still. Look for thresholds, where a small change in one input changes the pattern, and for ranges where nothing much happens.

Read the assumptions before the chart. The list under each model says what is fixed, what is simplified and what is left out. A striking result often depends on one of those assumptions, and the page states the limits so that you do not have to guess.

Keep the direction of evidence separate from the numbers. Where a model draws on research, the cited studies support a direction of effect, such as calmer activities reducing anger more than vigorous ones. The size of each effect on screen, its timing and the exact shape of each curve are choices made for illustration. The claims this page makes are recorded in the evidence ledger with their sources and their strength.

Background: the relief loop

Negative reinforcement is the process by which a behaviour becomes more likely because it removes something unpleasant. It does not require the behaviour to address the source of the unpleasant state; it requires only that relief follows. Experiential avoidance, the attempt to escape unwanted inner experience, has been proposed as a process common to many psychological difficulties (Hayes et al., 1996).

The model represents learning with the error-correcting update of the Rescorla–Wagner model (Rescorla & Wagner, 1972), in which each learning episode closes part of the gap between what has been learned and the most that can be learned. Here the gap is closed by 15% per episode, scaled by the amount of relief, so strong relief builds the urge quickly at first and more slowly as it approaches its limit. Because distress returns to the same level before every episode, the chart can show a growing habit alongside an unchanged problem. That is the pattern Research note 04 describes as the difference between relief and resolution. When the cause changes, the baseline falls and the urge, no longer reinforced, fades.

Background: venting and arousal over time

The hydraulic image of anger, as pressure that must be let out, predicts that hitting something should leave a person calmer. The experimental evidence points the other way. Venting while ruminating on a provocation increased anger and aggression compared with doing nothing (Bushman, 2002), and in a meta-analysis of anger management activities, those that lowered arousal reduced anger and aggression while those that raised arousal were ineffective overall (Kjærvik & Bushman, 2024).

The model shows these findings as three trajectories over half an hour. It is not a measured account of how anger decays, and the time constants are invented. What it keeps is the ordering: lowering arousal ends above waiting, and raising arousal ends no better than waiting. The separate arousal curve is a reminder that a vigorous activity can feel like discharge while leaving the body activated. Research note 02 sets out the studies and their limits.

Background: peak and end in remembered discomfort

Retrospective evaluations do not add up an experience moment by moment. Participants who rated film clips gave overall evaluations that were barely affected by how long the clips lasted (Fredrickson & Kahneman, 1993), a finding known as duration neglect. In the cold-water experiment, extending an unpleasant trial with a slightly milder ending made it the one most participants chose to repeat (Kahneman et al., 1993). The peak–end rule summarises the pattern: remembered discomfort tracks the worst and the final moments.

The model makes the arithmetic visible. You can build an episode whose total discomfort increases while its remembered version improves. The relevance to catharsis is direct: an intense night that ends on a high may be remembered as more restorative than its course would justify, and that memory can shape the choice to seek it again.

Background: synchrony in coupled oscillators

Synchronization appears in systems as different as flashing fireflies, cardiac pacemaker cells and oscillating chemical reactions. Kuramoto’s model captures what these systems share: many rhythms with different natural tempos, each nudged by the others. Strogatz (2000) reviewed a quarter of a century of work on the model, which shows a transition: below a critical coupling strength the population stays incoherent, and above it some oscillators spontaneously synchronize while others continue to drift.

The model is included as an analogy for interpersonal synchrony, the experience of moving in time with others. The analogy is useful for one point in particular: synchrony is a property of the group that emerges from interaction, without any conductor imposing it. Its limits are just as important. The oscillators have no feelings and no bonds. The human findings, that people who acted in synchrony cooperated more afterwards (Wiltermuth & Heath, 2009) and that synchronised dance raised pain thresholds and in-group bonding (Tarr et al., 2015), come from experiments with people and do not depend on this mathematics. Research note 03 discusses them.

Background: allostatic load and recovery

McEwen (1998) described stress mediators as protective in the short term and damaging when they are activated too often or not switched off efficiently. Allostatic load names that cumulative cost. The model reduces the idea to input and recovery. With frequent stressors and slow recovery, load keeps rising for weeks; with the same stressors and faster recovery, it settles at a lower plateau. The lesson is qualitative: how quickly a person recovers can matter as much as how often stress arrives, and a single restorative event lowers the load only briefly if the rate of arrival does not change. Research note 09 describes the research on stress and social support.

What models cannot capture

Every model on this page leaves out the things that matter most in a life.

Individual differences are absent. The models apply the same rules to everyone, whereas people differ in temperament, history, health, sensitivity to stress and ways of coping. A group-level finding describes an average across participants and does not say how any one person will respond.

Context is absent. A relief behaviour at a party, in a crisis or after a loss is the same number in the model and a different event in life. Who is present, what is at stake and what options a person has are not variables in any of these simulations.

Meaning is absent. The models count units of distress, anger, discomfort, phase and load. They cannot represent what an experience means to the person living it, how people make sense of what happened to them, or why an episode that looks the same on a chart can be a turning point for one person and a repetition for another.

Change through action is absent. In the relief loop, the cause changes by a random draw. In life, causes change because people talk, decide, ask for help, leave, repair or are supported by others. None of that is modelled, and it is precisely the part a simulation cannot do for anyone.

For these reasons the models are not tools for assessing yourself or anyone else. If distress, anger or stress is persistent or severe, a conversation with someone you trust and, where needed, with a qualified professional is more informative than any curve.

Where to read further

The methods page explains why signals and models support inferences with stated uncertainty rather than conclusions about a person, and the evidence overview shows how every claim on this page is graded and reviewed.

The research notes give the empirical background for each model: relief is not resolution, the venting hypothesis, collective synchrony and stress and social buffering.

The theory pages develop the ideas behind them in more depth: learning, avoidance and relief, allostasis and stress, and collective emotion and ritual.

Clinical perspectives: emotional processing, exposure and the lesson of debriefing

2026-09-14T00:00:00+00:00

Why the clinic is the decisive test

In everyday life, the claim that release helps is hard to evaluate: people choose when to release, what else they do, and how they remember it. Clinical research tests the claim under the conditions where it matters most, with people in significant distress, with structured interventions, and with outcomes measured over months. It also has the strongest obligation to find out when an intervention does harm.

Emotional processing and exposure

Foa and Kozak (1986) proposed a theory of emotional processing of fear that became the foundation for exposure-based treatments. Fear, in this account, is represented in memory as a structure linking stimuli, responses and their meanings. Two conditions are needed for that structure to change. First, the fear memory must be activated, which is why avoidance maintains fear. Second, the person must encounter information that is incompatible with the elements of the structure, for example that a feared outcome does not occur or that anxiety declines without escape.

The second condition is the crucial difference from catharsis. Activation of the emotion is necessary but not sufficient. Re-experiencing fear without corrective information leaves the structure intact and can even strengthen it. Prolonged exposure therapy for post-traumatic stress disorder was developed on this principle: the traumatic memory is revisited repeatedly and systematically, and the processing that follows each exposure is treated as the active ingredient.

The debriefing lesson

For several decades it was common to offer people exposed to traumatic events a single session soon afterwards in which they described the event and expressed their emotional reactions. The practice, known as psychological debriefing, was intuitively compatible with the idea that early emotional release prevents later disorder.

Rose, Bisson, Churchill and Wessely (2002) conducted a Cochrane systematic review of randomised controlled trials of single-session individual debriefing. They found no evidence that it prevented post-traumatic stress disorder or reduced general psychological distress. In the trials with longer follow-up, there was some evidence of an increased risk of post-traumatic stress disorder among people who had received debriefing. The authors concluded that compulsory debriefing of trauma victims should cease.

The review does not show that talking after trauma is harmful in general, and it concerns a specific single-session format. Its lesson for the catharsis question is nevertheless direct: an intervention built on early emotional expression, delivered in good faith to large numbers of people, did not deliver the protection it was expected to provide.

Which regulation strategies relate to symptoms

Aldao, Nolen-Hoeksema and Schweizer (2010) meta-analysed studies relating emotion regulation strategies to symptoms of anxiety, depression, eating disorders and substance-related disorders. Rumination showed the largest association with psychopathology, and avoidance and suppression were also associated with more symptoms. Reappraisal, acceptance and problem solving were associated with fewer symptoms, though these associations were generally smaller. The data are largely correlational, so they describe patterns rather than proving that a strategy causes symptoms.

Webb, Miles and Sheeran (2012) meta-analysed experiments in which people were instructed to use a particular strategy, organised by Gross’s process model of emotion regulation (Gross, 1998). Strategies based on cognitive change, particularly reappraisal, had a small-to-medium effect on emotional outcomes. Suppressing the outward expression of emotion reduced expression, but attempts to suppress the experience of emotion were not effective. Distraction was modestly effective, whereas instructions to concentrate on one’s feelings were not. The pattern is consistent with the rest of this library: changing how a situation is understood changes the emotion more reliably than either holding it in or dwelling on it.

Structured attention without discharge

Mindfulness-based interventions train sustained, non-judgemental attention to present experience, including unpleasant emotion, without trying to change or expel it. Hofmann, Sawyer, Witt and Oh (2010) meta-analysed studies of mindfulness-based therapy and found moderate reductions in symptoms of anxiety and depression overall, with larger effects in people diagnosed with anxiety and mood disorders. The approach is neither suppression nor release. Its relevance here is as a contrast: a structured way of relating to emotion that does not rely on discharge.

Clinical implications

Four implications follow from this evidence.

When emotional expression is therapeutic, it is usually structured, repeated and combined with new learning, as in exposure-based and cognitive therapies.
Early, one-off emotional release after trauma is not a validated method of prevention.
Rumination and avoidance deserve attention because they are among the strongest correlates of psychological symptoms.
Persistent, disabling distress warrants assessment by a qualified professional. Nothing in this library replaces that assessment.

The artifact’s problem_solved: false is, in clinical terms, a reminder that symptom relief in the moment and recovery over time are different outcomes, measured differently, and achieved by different means.

Collective rhythm: synchrony, exertion and social bonding

2026-09-14T00:00:00+00:00

Why crowds move together

A dancing crowd entrains to a shared beat. People adjust their movements to the music and to each other until hundreds of bodies move at nearly the same tempo. Research on interpersonal synchrony asks what that coordination does to the people involved.

Synchrony, liking and cooperation

Hove and Risen (2009) asked participants to tap along to a visual metronome while an experimenter tapped beside them. When the two tapped in synchrony, participants later rated the experimenter as more likeable than when their tapping was out of time, even though nothing else in the interaction differed.

Wiltermuth and Heath (2009) had small groups walk in step, sing or move together, or do the same activities out of synchrony. Groups that had acted in synchrony cooperated more in subsequent economic games, including games in which cooperation required individuals to give up some personal gain. The authors propose that synchrony strengthens the sense of being part of a group.

Dance, exertion and pain thresholds

Tarr, Launay, Cohen and Dunbar (2015) manipulated two variables in group dance: whether participants moved in synchrony and how physically demanding the movements were. Both synchrony and exertion independently raised pain thresholds, and synchrony also increased reported closeness to the rest of the group.

Pain threshold is used in this research as an indirect marker of endogenous opioid activity, because endorphins have analgesic effects. Opioid release itself was not measured. The findings support the interpretation that group dance works partly as a bonding mechanism, while leaving the physiological pathway to be confirmed with direct measures.

Music, anticipation and reward

Salimpoor and colleagues (2011) combined positron emission tomography and functional MRI while listeners heard music they had chosen because it reliably moved them. Moments of peak emotional response were accompanied by dopamine release in the striatum. Activity in the caudate was associated with anticipating those moments and activity in the nucleus accumbens with experiencing them. The structure of a build-up and a drop in electronic dance music is an obvious cultural parallel, although that genre was not the subject of the study.

Janata, Tomic and Haberman (2012) studied groove, the pleasurable urge to move in time with music, and linked it to sensorimotor coupling between perception of rhythm and bodily movement.

Collective effervescence: gatherings and demonstrations

The studies above manipulate synchrony in small groups. Large gatherings, such as demonstrations, rituals, ceremonies and celebrations, are harder to study experimentally, and research on them relies mainly on measuring what participants report. Pizarro and colleagues (2022) reviewed one construct from this literature, collective effervescence: a state of intense shared emotional activation and sense of unison that emerges during collective behaviour. The first part of their article traces the concept from Durkheim through Collins, Moscovici and Social Identity Theory to research on self-transcendence, and proposes a unified description.

The second part is a meta-analytic integration of 50 studies with a combined N of 182,738. Collective effervescence was strongly associated with individual emotions, particularly self-transcendent emotions, and with communal-sharing constructs such as group identity and identity fusion; the authors treat these as evidence of construct validity. Among more distal variables, it was associated with collective positive emotions, longer-term social integration such as ingroup commitment, social values and beliefs, and empowerment-related variables including well-being, collective efficacy and collective self-esteem. In demonstrations specifically, collective effervescence was associated with group identity (pooled r = 0.52), collective efficacy (r = 0.37), negative emotions (r = 0.14), self-transcendent emotions (r = 0.58) and morality-related beliefs (r = 0.43).

The limits follow from the form of the evidence. The pooled estimates are correlations, so they describe how strongly measures go together, not whether the experience causes the outcomes; people who already identify strongly with a cause may both experience more effervescence and report more efficacy. The construct is measured with questionnaire scales, so both sides of most associations depend on self-report, which can inflate correlations. Participants in a demonstration or ritual have chosen to be there, so the samples do not represent people in general. The authors examined moderators such as study design, the scale used and the type of gathering, which indicates that results varied across these features. The review supports the claim that intense shared emotion at gatherings goes together with belonging, meaning and a sense of collective agency. It does not show that attending a gathering produces those outcomes, or that it changes individual difficulties.

What these findings explain

Taken together, the studies describe why shared rhythmic activity feels good and brings people closer. Coordinated movement strengthens affiliation, exertion and synchrony are associated with analgesic and bonding effects, and musical peaks engage reward circuitry. They map closely onto the synchronization and discharge stages of the artifact’s process.

What they do not show

The outcomes measured are liking, cooperation, pain thresholds, reported closeness and reward responses during or shortly after the activity. None of the studies measured whether personal difficulties that participants brought with them were reduced afterwards. Collective rhythm is a genuine social and physiological resource. The artifact’s question is not whether it works, but what people expect it to accomplish.

Crying, talking and sharing: when emotional expression helps

2026-09-14T00:00:00+00:00

The folk theory of release

Few beliefs about emotion are as widespread as the idea that a good cry, or getting something off one’s chest, brings relief. The belief is a direct descendant of the hydraulic model described in Research note 01: emotion accumulates and must be discharged. Research on crying and on talking about emotional experiences allows that belief to be examined directly, and the results are more interesting than a simple yes or no.

Crying: retrospective benefit, immediate cost

Bylsma, Vingerhoets and Rottenberg (2008) analysed reports from several thousand participants across many countries who described their most recent crying episode. Most people reported that their mood improved after crying, but a substantial proportion reported no change, and some reported feeling worse. Benefit depended on the circumstances. Participants were more likely to report improvement when they received social support while crying and when the episode led to a resolution or a new understanding of the event that triggered it. They were less likely to benefit when they felt ashamed of crying or when the crying was accompanied by symptoms of mood disorder.

Laboratory research tells a different story in the short term. In studies where participants watch sad films, mood is often worse immediately after crying than in participants who did not cry. Gračanin, Bylsma and Vingerhoets (2014) reviewed this apparent contradiction and proposed that crying may act as a self-soothing behaviour whose calming effects, possibly mediated by parasympathetic activation, emerge with a delay, and that much of its benefit comes from the social responses it elicits. Retrospective reports may also be shaped by how the episode ended rather than by its average course.

The conclusion is not that crying is useless. It is that crying is not an automatic valve. Its effects depend on what happens around it and whether the situation that caused it changes.

Rimé (2009) reviewed a long research programme on the social sharing of emotion. Across ages, cultures and types of emotion, people talk about the large majority of their emotional experiences with others, usually within hours or days and often repeatedly. More intense emotions are shared more often.

Sharing has clear social functions. It elicits attention, empathy and support, and it strengthens the relationship between the person who shares and the person who listens. People also believe that sharing helps them recover. Experimental work, however, separates the perceived benefit from the measured one. In the studies reported by Zech and Rimé (2005), participants who talked about a negative emotional experience rated the conversation as helpful, yet the emotional impact of the memory, measured later, was not reduced compared with participants who did not talk about it. What did appear to matter was whether the conversation supported cognitive work on the experience, such as reappraisal and meaning-making, rather than a re-expression of the emotion alone.

This mirrors the evidence on expressive writing. Benefits have been linked to changes in how people construct an account of the experience over time, not to the intensity of emotion expressed in any single session (Pennebaker, 1997).

Naming a feeling

Lieberman and colleagues (2007) showed participants photographs of emotional facial expressions and asked them either to choose a word describing the emotion, to choose a name matching the person’s gender, or simply to look. Labelling the emotion was associated with reduced activity in the amygdala compared with the other conditions and with increased activity in the right ventrolateral prefrontal cortex; the two effects were inversely related. The authors interpret affect labelling as a form of implicit emotion regulation.

The study used standardised images rather than personal experiences, so it cannot show how labelling affects recovery from real events. It does illustrate a mechanism through which putting feelings into words can change their course without any discharge at all.

What this means for catharsis

Taken together, these findings suggest a revision of the folk theory. Emotional expression tends to help when it brings in one or more of three things: support from other people, understanding of what happened and what it means, or regulation through processes such as labelling and reappraisal. Expression that consists only of re-experiencing the emotion, without any of these, produces a sense of release more reliably than it produces recovery.

For the artifact this is a useful refinement. A crowd supplies support of a kind: shared rhythm, closeness and a sense of belonging (Research note 03). What it rarely supplies is understanding of the specific situation a person brought with them. That part usually happens elsewhere, in conversation, in writing, or in treatment.

Exercise and mood: what acute and regular physical activity change, and why exertion is not venting

2026-09-14T00:00:00+00:00

Two questions that are easy to merge

“Exercise helps with emotions” is one of the most repeated pieces of wellness advice, and one of the best supported. It also conceals two different questions and one tempting mistake.

The first question concerns acute effects: what happens to affect during a session of physical activity and in the hours afterwards. The second concerns regular activity: whether people who exercise over weeks or months show lower levels of depression, anxiety or distress than comparable people who do not. The evidence for these questions comes from different designs and supports different kinds of statements.

The tempting mistake is to read the acute lift in mood as a discharge, as if exertion burned off a quantity of anger or tension. That reading is the hydraulic model of emotion in athletic dress, and the evidence on anger does not support it (Research note 02). This note takes the questions in turn and then returns to the mistake.

Acute effects: a lift in positive, energetic affect

Reed and Ones (2006) meta-analysed studies of single sessions of aerobic exercise and positive activated affect, the family of feelings described as energetic, alert or enthusiastic. The abstract of that meta-analysis could not be retrieved for this library, so its study counts and effect sizes are not reported here; the library records it as support for a short-term increase in positive activated affect, held at moderate confidence (Research note 04).

Laboratory sessions are one thing; everyday life is another. Liao, Shonkoff and Dunton (2015) reviewed studies of the acute relations between affect and physical activity in free-living, naturalistic settings in non-clinical populations, typically using repeated momentary reports during the day. They identified 14 eligible studies. Twelve tested affective states after physical activity and yielded consistent evidence that activity predicted higher positive affect over the next few hours, with some evidence of more energetic feelings. The evidence that physical activity reduced negative affect in the next few hours was inconsistent and inconclusive. Six studies examined the reverse direction and found that positive affective states were associated with more physical activity over the next few hours, while negative states had no significant association.

Three points follow from this pattern.

The best-supported acute change is an increase in positive, energetic feeling, not a decrease in distress. This matters for anyone who exercises specifically to get rid of a bad feeling. The evidence from daily life is stronger for “feeling more energetic afterwards” than for “feeling less upset afterwards”.

Direction runs both ways. People who already feel good are more likely to be active, so everyday associations between activity and mood are partly the result of mood shaping activity. The designs reviewed are observational and cannot rule out that both are influenced by something else, such as a free afternoon or good company.

The effect is time-limited. “Over the next few hours” is the scale at which these studies measured. An acute lift is real and useful, and it is a change in state, not in the circumstances that produced a difficult emotion.

Intensity: why “harder” is not “better” for how exercise feels

If exertion discharged tension, harder exertion should feel better. The research on intensity points the other way. Ekkekakis, Parfitt and Petruzzello (2011) reviewed 33 articles published between 1999 and 2009 on the relation between exercise intensity and affective responses. Pleasure was reduced mainly above the ventilatory or lactate threshold, the point beyond which the body can no longer meet the demand aerobically. Below that threshold, most people experienced pleasant changes; close to it, individual variability was large; above it, changes were homogeneously negative. When people selected the intensity themselves, rather than having it imposed, they appeared to tolerate higher intensity levels.

The authors’ main concern is exercise adherence: people are more likely to repeat what felt pleasant. For emotion regulation the review has a direct implication. The pleasant affect that people seek from exercise is most reliably found at moderate, self-chosen intensities. Pushing to exhaustion in order to “get it out” moves exertion into the range where the experience during exercise tends to become unpleasant.

Regular activity: depression, anxiety and distress

The second question has been studied with randomised trials, and it has been synthesised many times. Four syntheses illustrate what is known and how certain it is.

Rebar and colleagues (2015) conducted a meta-meta-analysis of the effects of physical activity on depression and anxiety in non-clinical adult populations. They identified eight meta-analytic outcomes of randomised trials, based on 92 studies with 4,310 participants for depression and 306 study effects with 10,755 participants for anxiety. Physical activity reduced depression by a medium effect (standardised mean difference, SMD = -0.50, 95% CI -0.93 to -0.06) and anxiety by a small effect (SMD = -0.38, 95% CI -0.66 to -0.11). Neither effect showed significant heterogeneity across meta-analyses. Because meta-analyses can include overlapping trials, a meta-meta-analysis counts some evidence more than once, and the wide confidence interval for depression is a reminder that the size of the effect is uncertain.

Schuch and colleagues (2016) focused on people with depression, including major depressive disorder. They included 25 randomised trials comparing exercise with control conditions, 9 of them in major depressive disorder. After adjusting for publication bias, exercise had a large effect on depressive symptoms (SMD = 1.11, 95% CI 0.79 to 1.43). Larger effects were found for aerobic exercise at moderate and vigorous intensities. Most of their adjusted analyses suggested that publication bias had led earlier meta-analyses to underestimate the effect, which is the unusual direction.

Noetel and colleagues (2024) conducted a systematic review and network meta-analysis of 218 randomised trials with 14,170 participants who met clinical cut-offs for major depression. Compared with active controls such as usual care or a placebo tablet, moderate reductions in depression were found for walking or jogging (Hedges’ g = -0.62), yoga (g = -0.55), strength training (g = -0.49), mixed aerobic exercise (g = -0.43) and tai chi or qigong (g = -0.42). The effects were proportional to the intensity prescribed. The results appeared robust to publication bias, but only one study met the Cochrane criteria for low risk of bias, and confidence in the estimates was rated low for walking or jogging and very low for the other treatments. The authors note that expectancy effects could be reduced by blinding participants and staff, which is difficult in exercise trials.

Singh and colleagues (2023) took the widest view, an umbrella review of 97 systematic reviews covering 1,039 trials and 128,119 adult participants, including healthy adults, people with mental health disorders and people with chronic diseases. Physical activity had medium effects on depression (median effect size -0.43), anxiety (median -0.42) and psychological distress (-0.60) compared with usual care. Higher intensity was associated with greater improvements, and effectiveness diminished with longer interventions. Most of the included reviews (77 of 97) had a critically low quality score on the instrument used to assess them.

Taken together, these syntheses consistently associate regular physical activity with lower depressive and anxiety symptoms in randomised comparisons. The size of the effect depends on population, comparison and analysis, and the certainty is limited by trial quality and by the impossibility of blinding people to whether they exercise. The appropriate wording is that exercise is supported as a way of reducing symptoms, with moderate confidence in the direction and less in any particular number.

Intensity again: a tension to hold

Two findings above appear to pull in different directions. For how exercise feels during a session, pleasure declines above the ventilatory threshold. For regular exercise and symptoms, higher intensity was associated with larger improvements. These are different outcomes on different time scales, and they are compatible: an unpleasant session can still contribute to a later reduction in symptoms. What they do not support is the idea that intensity is valuable because it discharges emotion. The clinical trials measured symptoms over weeks, not release within a session.

The same tension has a practical side. Unpleasant exercise is less likely to be repeated, and the benefits in trials came from repetition. Self-selected intensity, which Ekkekakis and colleagues found to foster tolerance of higher intensity, is one way to reconcile the two.

Why exertion is not venting

Kjærvik and Bushman (2024) meta-analysed 154 studies of anger management activities and classified them by their effect on physiological arousal. Arousal-decreasing activities, such as deep breathing, mindfulness and meditation, reduced anger and aggression (g = -0.63). Arousal-increasing activities, including hitting a bag, jogging and cycling, were ineffective overall (g = -0.02), with heterogeneous results. The authors concluded that the findings do not support venting anger or going for a run as effective anger management activities. The meta-analysis and its limits are described in Research note 02.

This can look like a contradiction of everything above. It is not, once outcomes are separated. The exercise literature measures positive affect, depressive and anxiety symptoms, and distress. The anger literature measures anger and aggression, usually shortly after a provocation. An activity can lift energetic positive affect and still leave anger, measured at the same time, where it was.

Two further findings sharpen the distinction. First, exertion raises arousal, and arousal is not tied to a particular emotion. In Bushman’s (2002) experiment, participants who hit a punching bag while thinking about the person who provoked them reported more anger than both other groups and behaved most aggressively; in aggression, those who hit the bag while thinking about getting fit fell between that group and a group that sat quietly. What people attend to while their arousal is high appears to shape what that arousal becomes. Second, the acute effect that daily-life studies support is an increase in positive affect, not a reliable decrease in negative affect. Exercise is therefore better understood as adding something (energy, a sense of capability, often company and a change of setting) than as removing a stored quantity of anger.

The practical reading is straightforward. When angry, lowering arousal first is the approach with meta-analytic support. Physical activity is valuable for mood, and it is most compatible with anger reduction when it does not become a rehearsal of the grievance, and when its intensity is not used as a way of “getting it out”.

Relevance to catharsis

Exercise is a clear case of a restorative experience that feels like release. People say they “ran it off” or “sweated it out”. The evidence supports part of that experience: a real, short-lived lift in positive, energetic affect and, with regular practice, lower symptoms of depression and anxiety. It does not support the discharge metaphor. The benefits appear with moderate, repeated activity, not with a single exhausting burst, and exertion without a change in attention did not reduce anger.

That is the distinction this library draws between relief and resolution (Research note 04). Exercise can make a state easier to bear, and over time it can change symptom levels. It does not by itself change what produced the state in the first place.

Practical implications, with limits

Use physical activity for mood on its own terms. A walk, a ride or a session at a comfortable, self-chosen intensity is a reasonable way to lift energy for the next few hours.
Do not use exertion to discharge anger. Lower arousal first; if you move, keep attention on the activity rather than on the grievance.
Think in weeks, not sessions. The trial evidence on symptoms concerns regular activity; one hard session is not the unit of benefit.
Exercise is not a substitute for care. Trials in depression are encouraging, and exercise is often recommended alongside other treatments, but persistent low mood, anxiety or hopelessness warrants assessment by a qualified professional.

These implications follow from the evidence but have not been tested as a package. People with medical conditions should consider advice on appropriate activity from a clinician.

Limits of the evidence

Acute evidence from daily life is observational and runs in both directions. The acute meta-analysis on positive activated affect could not be re-verified from an abstract. The trial evidence on regular activity comes from syntheses with overlapping studies, frequent risk of bias, unblinded participants and, in the umbrella review, mostly low-quality reviews. The anger meta-analysis found heterogeneity among arousal-increasing activities, so specific activities in specific contexts may differ from the average. The evidence ledger records the grade and review date of each claim.

Expressive writing: what four decades of trials show, and why relief is not the measure

2026-09-14T00:00:00+00:00

A paradigm built to test disclosure

Most claims about the value of “getting it out” are hard to test, because people choose when and how to express emotion. Expressive writing is an exception. In the paradigm introduced by Pennebaker and Beall (1986), participants are randomly assigned to write, typically for 15 to 20 minutes on several days, either about their deepest thoughts and feelings concerning a difficult experience or about a neutral topic. Because assignment is random and the task is standardised, differences in later outcomes can be attributed to the writing.

In the original study, students who wrote about both the facts of a traumatic event and their emotions about it made fewer visits to the health centre in the following months than controls, although they reported more distress immediately after writing; writing about only emotions or only facts did not produce the same pattern. The abstract of that study could not be re-verified for this library, so no numbers are reported for it here. Pennebaker (1997) later reviewed the growing literature and highlighted evidence that changes in the use of words related to causation and insight across sessions predicted benefit, which pointed towards cognitive processing rather than release alone. Both are discussed in Research note 01 and the advice on writing to make sense.

This note looks at what happened once the paradigm had been tested hundreds of times, which is where the most informative evidence lies. The broader term used in some syntheses is experimental disclosure, which includes writing and talking under the same randomised design.

The first syntheses

Smyth (1998) conducted a research synthesis of the written emotional expression task in healthy participants. Writing was found to lead to significantly improved health outcomes in four outcome types: reported physical health, psychological well-being, physiological functioning and general functioning. Health behaviours were not influenced. Writing also increased immediate, pre- to post-writing distress, and that increase was unrelated to health outcomes. The relation between writing and health was moderated by several variables, including whether participants were college students, gender, the duration of the manipulation, the publication status of the study and the specific writing instructions.

Two findings from this early synthesis remain central. The first is that the benefit appeared across different kinds of outcome, not only in one measure. The second is the dissociation between immediate distress and later outcome. Writing usually makes people feel worse right afterwards, and how much worse did not predict whether they benefited. That already separates the experience of the session from its measured effect.

The largest meta-analysis: small, positive, variable

Frattaroli (2006) noted that the two earlier meta-analyses had been small and had used fixed-effects methods, which limit generalisation beyond the included studies, and that many studies had been completed since. She collected 146 randomised studies of experimental disclosure and used random-effects analyses. Experimental disclosure was effective, with a positive and significant average effect size of r = .075. A number of moderators were identified.

An r of .075 is small by any convention for effect sizes. It is not nothing: a cheap, brief, self-administered task with a reliably positive average effect is of practical interest. But it is far from the picture of a powerful release that the popular idea of catharsis suggests. The average also conceals variation, which is what the moderator analyses are for. The abstract does not list which moderators were significant, and this library does not itemise them from secondary accounts.

Later syntheses in specific populations and outcomes

The broad average leaves open whether writing helps with the problems people most often hope it will help with. Three later meta-analyses address narrower questions.

Depressive symptoms. Reinhold, Bürkner and Holling (2018) meta-analysed 39 randomised controlled trials with 64 intervention-control comparisons in physically healthy adults with varying degrees of stress but without posttraumatic stress disorder. Expressive writing did not yield significant long-term effects on depressive symptoms. Effects were larger when the number of sessions was higher and when the writing topic was more specific. The authors conclude that the results did not support brief, self-directed expressive writing as an intervention that decreases depressive symptoms in this population, and that longer, more directed interventions with additional therapeutic support might produce different results.

Adolescents. Travagin, Margola and Revenson (2015) analysed 21 independent studies in young people aged 10 to 18. The overall mean effect was positive but relatively small (g = 0.127), with significant effects between 0.107 and 0.246 in the domains of emotional distress, problem behaviour, social adjustment and school participation. Few significant effects were found for moderators such as participant characteristics, instructions or design. Adolescents with high baseline emotional problems showed larger effects on school performance, and higher-dosage protocols, with more and to some extent more widely spaced sessions, showed larger effects on somatic complaints. The authors highlight the importance of modifying the traditional protocol to enhance its efficacy and reduce potential detrimental effects, and judge the evidence promising but not decisive.

Posttraumatic stress. Pavlacic and colleagues (2019) took a different analytic approach, pooling change within the expressive writing conditions only rather than comparing writing with control groups. They found a small overall effect for posttraumatic stress and negligible to small effects for posttraumatic growth and quality of life; studies that required a diagnosis of posttraumatic stress disorder showed a medium to large effect. Because there is no comparison with a control condition, this analysis cannot separate the effect of writing from natural recovery, regression to the mean or repeated measurement. It is graded lower in this library for that reason, and its larger effect in people diagnosed with PTSD should not be read as evidence that writing treats PTSD.

What moderates the effect

Across these syntheses, a few themes recur, although each rests on moderator analyses, which are observational comparisons between studies rather than experimental tests.

Dose. More sessions were associated with larger effects on depressive symptoms, and higher-dosage protocols with larger effects on somatic complaints in adolescents. The duration of the manipulation was already a moderator in the first synthesis.
Instructions and focus. Specific writing topics were associated with larger effects than open ones in the depression meta-analysis, and writing instructions were a moderator in the first synthesis. This fits the early finding that writing about both facts and feelings differed from writing about feelings alone.
Who writes. Student samples, gender and baseline problems appeared as moderators in different syntheses. The pattern does not add up to a simple rule about who benefits.

Sloan and Marx (2004) reviewed the theories proposed to explain the paradigm and concluded that, although it may produce beneficial effects for some individuals, it remained unclear why it works and for whom it is best suited. Later syntheses have narrowed but not closed that gap.

Relief and benefit are different measurements

The expressive writing literature offers one of the clearest demonstrations in psychology that how a procedure feels and what it does can come apart.

First, the session itself tends to be unpleasant. Smyth found that writing increased immediate distress, and that increase did not predict health outcomes. People who felt worse after writing were not thereby less likely to benefit, and people who felt better were not thereby more likely to.

Second, when benefit appears, it appears later, over weeks or months, and in outcomes such as health centre visits, well-being and functioning, not in the mood at the end of the session.

Third, the features associated with benefit are those of sense-making rather than discharge: writing about facts and feelings together, specific topics, repeated sessions, and changes in causal and insight language. Pouring out feeling without building an account is not what the successful protocols asked people to do.

This has a direct consequence for anyone who uses writing informally. The sense of relief at the end of a writing session, or its absence, is weak evidence about whether the practice is helping. A more informative check is whether, over days, the account of the experience has become more coherent, whether the same fragments still intrude in the same form, and whether anything has changed in understanding or in what one does (Research note 04).

When writing may resemble rumination

If writing helps through sense-making, then writing that circles the same feelings without development may resemble rumination, the repetitive, passive focus on distress that is associated with more symptoms (Research note 15). None of the syntheses above directly tested this possibility, so it is a reasoned hypothesis, not a finding. It is consistent with the evidence on perspective: reflecting on an upsetting experience from a self-distanced perspective reduced emotional reactivity and rumination compared with immersed recall (Kross, Ayduk & Mischel, 2005). A written account that steps back from the experience may differ from one that re-immerses the writer in it.

The adolescent meta-analysis’s reference to potential detrimental effects is a further reminder that writing about painful material is not risk-free for everyone, and that protocols can be adapted to reduce harm.

What the evidence supports, and what it does not

The evidence supports expressive writing as a low-cost practice with a small average benefit across many randomised studies, with larger effects associated with more sessions and more specific topics, and with a short-term increase in distress that is expected and not in itself a warning sign. It does not support expressive writing as a treatment for depression in the form of brief, self-directed sessions, as a treatment for posttraumatic stress disorder, or as a reliable way of discharging emotion. It does not predict who will benefit.

In the language of this library, expressive writing is best understood as a structured opportunity for meaning-making whose average effect is modest and whose value for a particular person is unknown in advance.

Practical implications, with limits

Write about facts and feelings together, over several sessions. The protocols with benefits used repeated sessions and asked for both.
Choose a specific topic. Specific topics were associated with larger effects on depressive symptoms than open ones.
Expect to feel unsettled afterwards. Immediate distress was unrelated to later benefit; neither relief nor discomfort at the end of a session tells you much.
Look for change in the account, not in the intensity of feeling. If the same material repeats unchanged, a different approach, such as stepping back or moving to concrete steps, may help more (Advice).
Do not use writing in place of treatment. Persistent low mood, trauma-related symptoms or writing that leaves you overwhelmed rather than temporarily unsettled warrant speaking with a qualified professional.

These implications follow from the syntheses but have not been tested as a package.

Limits of the evidence

The meta-analyses pool studies with different populations, instructions, doses and outcomes; the average effect is small and heterogeneous. Moderator findings compare studies rather than testing conditions experimentally. The largest meta-analysis’s moderators are not itemised in its abstract. Many trials used student samples and self-reported outcomes, and publication status was itself a moderator in an early synthesis. The posttraumatic stress meta-analysis lacks control comparisons. The original 1986 study could not be re-verified from an abstract. The evidence ledger records the grade and review date of each claim.

Grief and mourning rituals: what rituals do, and where grief becomes a clinical question

2026-09-14T00:00:00+00:00

Why grief belongs in a library about catharsis

Grief is the situation in which the idea of emotional release seems most natural. Popular advice tells bereaved people to “let it out”, to “work through” the loss, and warns that grief held in will return later in a worse form. Mourning customs across cultures also involve intense, often collective, expression: weeping, lament, music, gathering. It is therefore reasonable to ask what the evidence says about grief and about the rituals that surround it, and whether their value lies in discharge or somewhere else.

This note is written with a particular caution. Many readers of a page about grief are grieving. Nothing here can say how a particular person’s grief is going, whether it is “normal”, or what they should feel. The evidence describes groups of bereaved people under particular conditions. The last sections describe when grief becomes a clinical question and when to contact a professional; those are general signs, not an assessment.

What grief usually looks like

Stroebe, Schut and Stroebe (2007) reviewed the relation between bereavement and physical and mental health. Two findings sit side by side in their review. Bereavement is associated with an excess risk of mortality, particularly in the early weeks and months after a loss, and with decrements in physical health, visible in symptoms, illnesses and use of medical services. At the same time, grief is not a disease, and most people adjust without professional psychological intervention. For a few people, mental disorders or complications in the grieving process follow, and the authors conclude that intervention should be targeted at people at high risk and those with complicated grief or bereavement-related depression and stress disorders.

The second finding runs against a common expectation. Bonanno and colleagues (2002) had an unusual opportunity: prospective data on 205 people collected several years before the death of their spouse, followed by assessments 6 and 18 months after the loss. This design allowed them to see grief trajectories against each person’s own pre-loss state instead of reconstructing it from memory. They identified five core patterns: common grief, chronic grief, chronic depression, improvement during bereavement, and resilience. Common grief, the pattern of marked distress that gradually subsides, was relatively infrequent. The resilient pattern was the most frequent. Chronic grief was associated with pre-loss dependency, and resilience with pre-loss acceptance of death and belief in a just world.

Two cautions keep this finding in proportion. The sample was older adults who lost a spouse, and trajectories after the sudden death of a child, a violent death or a death in a pandemic may differ. And “resilience” in this study describes a pattern of measured symptoms and functioning; it does not mean that those people did not miss their partner or were untouched. The practical point is narrower: the absence of intense visible distress is not in itself a sign that grief is being suppressed or delayed.

From “grief work” to oscillation

For much of the twentieth century, the dominant idea in bereavement theory was the grief work hypothesis: that bereaved people must confront the loss and work through their feelings about it, and that avoiding this work leads to worse outcomes. The idea is close to the hydraulic model of emotion that this library examines elsewhere (Research note 01).

Stroebe and Schut (1999) criticised the hypothesis on several grounds: imprecise definition, failure to represent the dynamic character of grieving, lack of empirical evidence and validation across cultures and historical periods, and a narrow focus on intrapersonal processes. In its place they proposed the dual process model. It identifies two kinds of stressors. Loss-oriented stressors concern the loss itself: yearning, reminders, the relationship with the person who died. Restoration-oriented stressors concern the secondary consequences: new roles, practical tasks, a changed identity and daily life. Adaptive coping, in the model, is an oscillation between confronting and avoiding each kind of stressor, and it includes what the authors call dosage: taking respite from both as an integral part of coping.

The model is a theoretical proposal, and the paper discusses rather than pools the supporting evidence. Its importance for the catharsis question is conceptual. It treats time away from grief, including distraction and ordinary activity, as part of healthy grieving, not as avoidance to be overcome. It also treats continuous confrontation with the loss as potentially problematic, a point that connects grief to the research on rumination (Research note 15).

Continuing bonds

A second change in bereavement theory concerns the relationship with the person who died. Older models tended to treat detachment as the goal of mourning. Later work described continuing bonds: ongoing inner relationships with the deceased, expressed in memories, conversations, keeping possessions, or a sense of their presence.

Stroebe, Schut and Boerner (2010) observed that the scientific literature lacked clarity about whether continuing or relinquishing bonds is adaptive, and that it was unclear which kinds of bonds, or which underlying processes, relate to better or worse outcomes. Empirical research had rarely been theory-driven. They proposed an integrative model grounded in attachment theory, using the dual process model to link ways of coping with the bond to outcomes and mental representations theory to describe cognitions about the deceased. The model is meant to guide research; it does not itself establish when a continuing bond helps.

The honest summary is that a continuing bond is common and is not in itself a sign of a problem, and that the evidence does not support a single rule, in either direction, about how attached a bereaved person should stay.

What rituals do: experimental evidence

Norton and Gino (2014) tested the psychological effects of mourning rituals in three experiments on different kinds of loss, which the title summarises as losses of loved ones, lovers and lotteries. Participants were either directed to reflect on rituals they had performed after past losses or assigned to complete novel rituals after experiencing a loss. In both cases they reported lower levels of grief. Increased feelings of control after the rituals mediated the link between rituals and reduced grief. The benefit was not confined to people who professed a belief in the effectiveness of rituals; those who did not believe in them also showed it.

The authors note that the specific rituals people perform after losses vary widely by culture and religion, and among their own participants, and they interpret their results as suggesting a common psychological mechanism: regained feelings of control.

Several features limit what this can show. The outcomes were self-reported grief measured shortly after the manipulation. Two of the three kinds of loss (a relationship and a lottery) are not bereavement, and the bereavement data rely on recalled losses and recalled rituals. The experiments do not show what rituals do for grief over months, whether they change the risk of prolonged grief, or whether a funeral works the same way as a private ritual invented for a study. They are consistent with a modest, short-term benefit that does not depend on belief.

The observational picture

Field data on real funerals give a more cautious picture. Mitima-Verloop, Mooren and Boelen (2021) surveyed bereaved people in the Netherlands six months and three years after a loss (552 and 289 respondents). Participants generally considered the funeral and the rituals they used to be helpful. Yet no significant association was found between how they evaluated the funeral, or how much they used grief rituals, and their grief reactions.

The two findings can both be true. Something can be experienced as helpful without that experience predicting the course of grief months or years later; this library meets the same gap between felt benefit and measured change in other domains (Research note 04). A correlational design also cannot separate the effect of rituals from the reasons people use them: those most distressed might use rituals more, which would mask a benefit. The honest conclusion is that the evidence for rituals improving the longer-term course of grief is insufficient, while the evidence that people value them is consistent.

Why rituals might help: a framework

Hobson, Schroeder, Risen, Xygalatas and Inzlicht (2018) reviewed research on rituals across social psychology, cognitive science, anthropology, behavioural economics and neuroscience and proposed a framework with three regulatory functions: regulation of emotions, of performance goal states and of social connection. They distinguish bottom-up processes, which arise from the physical features of rituals such as fixed sequences and repetition, from top-down processes, which arise from what a ritual means to the people performing it.

Applied to grief, the framework suggests several routes, none of which is discharge. A structured sequence of actions may restore a sense of control at a time when control has been lost, which is what Norton and Gino measured. Shared rituals gather people and make support visible. And rituals give meaning to a transition: they mark that something has ended and that the relationship has changed form. The framework generates predictions; it does not show which route matters for a given loss.

Rituals are not discharge

The contrast with the hydraulic view is worth stating directly. The mechanism supported in the experiments is a sense of control, not the emptying of emotion. Many mourning rituals involve strong expression, but they are also highly structured: they have a beginning, a sequence and an end, they are shared, and they place the loss in a larger frame of meaning. These features resemble the processes that other research in this library links with benefit, such as social support and meaning-making, more than they resemble venting.

This also means that there is no evidence-based reason to press a bereaved person to express more grief than they do, or to participate in a ritual they find alien. The experimental benefit appeared with rituals people chose or recalled, and with novel rituals that were brief and simple.

The clinical boundary: prolonged grief

For a minority of bereaved people, intense grief persists and becomes disabling. Prigerson and colleagues (2009) interviewed 291 bereaved respondents three times, between 0 and 24 months after the loss, to test criteria for prolonged grief disorder. The proposed criteria required yearning for the deceased together with at least five of nine further symptoms (such as feeling emotionally numb or that life is meaningless, bitterness over the loss, difficulty accepting the loss, identity confusion, avoidance of the reality of the loss, or difficulty moving on) experienced at least daily or to a disabling degree, present at least six months after the death and associated with functional impairment. The authors report that these criteria appeared able to identify people at heightened risk of enduring distress and dysfunction. Prolonged grief disorder has since been introduced in ICD-11 (Johannsen et al., 2019) and in DSM-5-TR, where the adult duration threshold is 12 months since the death rather than six, and current criteria should be taken from clinicians and current manuals rather than from this 2009 proposal.

How common is it? Lundorff and colleagues (2017) meta-analysed 14 studies of non-psychiatric adult populations after non-violent bereavement and found a pooled prevalence of 9.8% (95% CI 6.8 to 14.0), higher with higher mean age. The authors caution that the studies were methodologically heterogeneous and of limited representativeness, and that the finding should be interpreted cautiously.

These criteria are described here so that readers understand where the clinical boundary lies, not so that anyone can apply them to themselves or others. Intense grief in the first months is not a disorder, and a list of symptoms read on a website is no substitute for an assessment by a qualified professional who knows the person and the circumstances.

What helps whom

Currier, Neimeyer and Berman (2008) meta-analysed 61 controlled studies of psychotherapeutic interventions for bereaved persons. Overall, interventions had a small effect at post-treatment and no statistically significant benefit at follow-up. Interventions that exclusively targeted people displaying marked difficulties adapting to loss had outcomes comparable with psychotherapies for other difficulties. The authors suggest that the discouraging results for studies that did not screen for distress could be explained by a tendency among untreated controls to improve naturally over time.

Johannsen and colleagues (2019) meta-analysed 31 randomised controlled trials of psychological interventions for grief in adults. The pooled effect on grief symptoms was g = 0.41 after the intervention (0.31 when adjusted for possible publication bias) and g = 0.45 at follow-up across 18 trials. Effects were larger for individually delivered interventions, for participants at least six months after the loss, and for participants included because of high baseline symptoms. The trials were heterogeneous and of varying quality.

Read together, these reviews are consistent with the dual process view. Offering routine therapy to everyone who is bereaved adds little on average, because most people improve without it. Structured help for people whose grief remains intense and disabling is supported by trial evidence. This mirrors the lesson of psychological debriefing after trauma, where a well-intentioned intervention for everyone did not deliver the expected protection (Research note 08).

When to seek professional help

These are general indications drawn from the sources above and from the site’s clinical boundary. They are not a test.

Grief that remains intense and consuming many months after the loss, and that makes it very hard to carry out work, care for oneself or others, or maintain relationships, is a reason to talk with a general practitioner, a psychologist or a bereavement service.
Persistent low mood, loss of interest, inability to sleep or eat, or heavy use of alcohol or other substances to cope are reasons to seek help at any point, not only after a set period.
Thoughts of ending your life, or of not wanting to go on living without the person who died, warrant contacting a professional without delay; if you are in immediate danger, contact local emergency services.

Seeking help is not evidence that grief has “gone wrong”, and not seeking it is not evidence that it has gone right.

Limits of the evidence

The experimental evidence on rituals rests on one set of three experiments with short-term self-report and largely non-bereavement losses. The field study found no association between rituals and grief reactions. The dual process model and the continuing bonds framework are theoretical. The trajectory evidence comes from spousal loss in later life, and the prevalence estimate pools heterogeneous studies of non-violent deaths. Cultural variation in mourning is large and mostly not captured by these studies. The evidence ledger records how each claim in this note is graded and when it will be reviewed.

Instrumented emotion: why the telemetry here is conceptual

2026-09-14T00:00:00+00:00

The appeal of emotion as data

Dashboards promise that anything important can be observed, quantified and graphed. The artifact borrows that language on purpose: input state, synchronization, root cause. It then declines to deliver the measurement. This note explains why.

Affective computing

In 1997 Rosalind Picard described affective computing, the study and development of systems that can recognise, interpret and respond to human emotion. The field has produced valuable research on emotion, human–computer interaction and assistive technology. It has also given rise to commercial products that claim to infer emotional states from faces, voices, text and physiological signals, often at scale and in public settings.

The evidence problem

Barrett, Adolphs, Marsella, Martinez and Pollak (2019) reviewed the scientific evidence behind the common assumption that particular emotions are reliably expressed by particular facial configurations, for example that a scowl signals anger. They examined studies of how people move their faces when they experience emotions and studies of how observers infer emotions from faces.

Their conclusion was that people do, on average, scowl when angry or smile when happy more often than would be expected by chance, but the relationship is weak and inconsistent. The same facial configuration can accompany different emotional states, the same emotional state can be expressed in many ways, and the relationship varies substantially across contexts, individuals and cultures. A facial configuration alone therefore cannot be treated as a reliable readout of a specific internal state. The authors cautioned that technologies built on that assumption are not justified by the evidence.

A later meta-analysis approached the same question from the side of emotional experience. Durán and Fernández-Dols (2021) tested whether happiness, sadness, anger, disgust, fear and surprise each co-occur with the facial expression predicted for them. In studies that coded full expressions with the Facial Action Coding System (FACS), the average co-occurrence effect size was .13. Including partial expressions and other coding systems raised it to .23, and the average correlation between the reported intensity of an emotion and the intensity of its predicted expression was .30. Co-occurrence was greatest for disgust and least for surprise, and results varied between samples. The authors concluded that the six classic basic emotions do not reliably co-occur with their predicted facial signal.

Asking people: self-report scales

Asking a person is the most direct route to how an experience feels, and several instruments are designed to make that brief and checkable. The Positive and Negative Affect Schedule (PANAS) consists of two 10-item mood scales. In the study that introduced it, the scales were highly internally consistent, largely uncorrelated with each other and stable at appropriate levels over a two-month period (Watson, Clark & Tellegen, 1988). The Self-Assessment Manikin (SAM) is a non-verbal pictorial technique that asks for only three judgements: pleasure, arousal and dominance. Compared with a semantic differential scale that requires 18 ratings, reports of experienced pleasure and felt arousal correlated highly across the two methods (Bradley & Lang, 1994).

Reliability of this kind shows that a scale measures something consistently. It does not show that a report matches what happened in the body, and it does not remove the dependence of self-report on timing and on beliefs about emotion that the methods page describes.

Are bodily signals any clearer?

If faces are unreliable, physiology might seem a better source: heart rate, skin conductance, breathing and blood pressure are measured by instruments rather than judged by observers. The question is whether particular emotion categories have consistent autonomic patterns, sometimes called fingerprints, that would allow the category to be inferred from the signals.

Kreibig (2010) reviewed 134 publications reporting experiments on emotional effects on peripheral physiological responding in healthy people. The review suggested considerable autonomic response specificity in emotion when subtypes of distinct emotions are considered, and emphasised that sound terminology for the affective states studied and the choice of physiological measures are both important for assessing it. In other words, the review found patterns, but they depended on how finely emotions were defined and what was measured.

Siegel and colleagues (2018) tested the fingerprint hypothesis directly with a meta-analysis of 202 studies measuring autonomic reactivity during laboratory emotion inductions in non-clinical adult samples. Mean effect sizes increased for 59.4% of autonomic variables across emotion categories, but the pattern of effect sizes did not clearly distinguish one emotion category from another. Variation within categories was substantial: heterogeneity accounted for a moderate to substantial share of variability (I² of at least 30%) in 54% of the effect sizes. Experimental features such as whether emotion was induced with films or imagery did not explain much of that variation, and correcting for publication bias reduced the estimated effects further. The authors interpret the results as more consistent with the view that an emotion category is a population of variable, context-specific instances than with fixed fingerprints.

The two reviews are not simply contradictory. They differ in method (a narrative review versus a quantitative meta-analysis with pattern classification) and in the level of description (subtypes versus broad categories). Together they support a practical conclusion: emotion is accompanied by bodily change, but inferring a specific emotion from autonomic signals alone, outside controlled conditions and without context, is not supported by consistent evidence. A rising heart rate is a signal; calling it fear, excitement or anger is an inference that needs other information. The methods page describes how this library separates signals, inferences and claims.

Do the channels agree?

Emotion theories commonly assume that an emotion organises experience, behaviour and physiology into one coherent response. Mauss, Levenson, McCarter, Wilhelm and Gross (2005) tested this within individuals, recording experience, facial behaviour and peripheral physiology second by second while participants watched a film that induced amusement and sadness. Experience and behaviour were highly associated, but physiological responses were only modestly associated with either. Greater intensity of amusement was associated with greater coherence between behaviour and physiology; greater intensity of sadness was not.

A review of experiential, physiological and behavioural measures reached a related conclusion. Each response system carries unique sources of variance, which limits how far measures converge, so there is no gold-standard measure and the three kinds of measure cannot be assumed to be interchangeable (Mauss & Robinson, 2009). The same review found that the bulk of the available evidence favours measures reflecting dimensions rather than discrete states. For any monitor this has two consequences: a physiological channel cannot stand in for self-report, and agreement between channels cannot be assumed where it has not been checked.

What neuroimaging adds

Brain imaging is sometimes presented as a way around these limits. Lindquist, Wager, Kober, Bliss-Moreau and Barrett (2012) pooled the neuroimaging literature on human emotion to compare two hypotheses: that each discrete emotion category consistently and specifically corresponds to distinct brain regions, and that emotion categories are constructed from more general networks that are not specific to any category. They found little evidence that discrete emotion categories can be consistently and specifically localised to distinct brain regions. Instead, a set of interacting regions that are also involved in non-emotional psychological operations was active during the experience and perception of a range of emotion categories. As with bodily signals, reading a state back from the activity of a region is a reverse inference, whose limits the methods page explains.

Outside the laboratory

Measuring emotion in everyday life replaces some laboratory problems with others. Ambulatory assessment gathers self-report, observational and physiological data from people in their natural environment, in real time or near real time. That minimises retrospective bias and yields ecologically valid data, and it raises questions of acceptability, compliance, privacy and ethics (Trull & Ebner-Priemer, 2013). Wilhelm and Grossman (2010) describe how emotional effects in physiological recordings can be masked by social engagement, mental and physical workload, food intake, and circadian and quasi-random variation in metabolic activity, so meaningful interpretation requires a high degree of context specification.

Hoemann and colleagues (2020) used physiologically triggered experience sampling: when cardiac activity changed substantially in the absence of movement, participants’ self-reports and peripheral physiological activity were recorded. Clustering analyses revealed variability in the number and nature of physiological patterns that recurred within individuals, and in the affect ratings and emotion labels associated with each pattern; some broad patterns also recurred across individuals. The authors read the findings as support for the view, also taken by Siegel and colleagues, that emotion categories are populations of variable instances tied to situation-specific needs.

Why the artifact does not measure

Validity. Inferring internal states from observable signals is least reliable in exactly the conditions a crowd provides: noise, movement, varied lighting, intoxication, diverse individuals and rapidly changing contexts. Even in everyday recordings of a single person, physiological signs of emotion can be masked by social engagement, workload, food intake and circadian variation (Wilhelm & Grossman, 2010). Stark and Hoey (2021) add that emotion-AI systems rest on particular conceptual models of emotion and on proxy data, that these choices shape the systems’ ethical and social implications, and that the paradigms computer scientists have developed or adapted should not be taken at their word as producing ground truth about human emotions.

Consent. A person looking at an artwork has not agreed to have their emotional state analysed, stored or displayed. A piece about instrumenting people should not quietly instrument its own audience.

Coherence. The artifact is about the difference between a successful response and a solved problem. Producing confident numbers it cannot justify would repeat the very confusion it describes.

The detection sandbox turns this argument into something you can operate. It takes five synthetic signals, shows the features derived from them and every reading they support at once, and lets you change only the setting to watch the leading reading change while the signals stay where they are. Nothing in it measures anyone.

How to read the numbers on the page

Every metric in the observability table is an artistic value, chosen to express a relationship rather than to report an observation. Relief is shown high but marked temporary. The root cause is shown at full scale and marked unchanged. The page states this next to the table, and the site contains no analytics, trackers or cookies, which can be verified in its source code.

Method: building an evidence-gated artifact with Majordomus

2026-09-14T00:00:00+00:00

What was built

This note documents that process with the same standard the research notes apply to psychology: report what was measured, separate it from interpretation, and state what cannot be concluded.

The recorded timeline of the first release

All times are UTC on 14 September 2026 and come from the Git history, the GitHub API, the Majordomus ledger and the pipeline records.

Time	Event	Source
09:45	Working session begins with the poster rebuild	local file timestamps
09:54:06	GitHub repository created; poster committed	GitHub API, Git
10:26:23	Majordomus layer initialised; first supervised task started	Majordomus ledger
10:29:16	Scope drift reported after files were moved; task restarted	Majordomus ledger
10:39:39	A malformed scope reported by the pre-push check; task restarted before any push	Majordomus ledger
10:42:45	Site committed; the pre-commit hook ran the Majordomus health check with zero failures	Git, hook output
10:43:02–10:45:45	Pages pipeline: 7 jobs, 2 minutes 43 seconds, all passed	GitHub Actions
10:45:42	Release v0.1.0 published after live verification	GitHub API
10:47:25	Task closed as completed; verification command was the live smoke test (exit 0, 7 seconds)	Majordomus ledger

Derived durations:

repository creation to first verified release: 51 min 36 s
Majordomus initialisation to first verified release: 19 min 19 s
push to verified release: 2 min 57 s

At that release the site was checked by 7 validation stages and 30 browser tests. Later phases (link previews and structured data, then this research library and the advice library) followed the same loop and are visible in the repository’s release history.

The kinds of instructions

The work was driven by about fifteen messages from one person. They fell into five types:

A one-line request to create a project on GitHub.
Art direction for the poster, followed by a choice between three offered approaches.
Structured execution briefs. Three long specifications of increasing scope: a static artwork page, an interactive microsite, and a multilingual publishing system with dozens of numbered requirements and an explicit completion gate.
Short directives during work: deploy, validate in a browser, release automatically, use Majordomus and enforce it.
Editorial direction: apply HTML5 Boilerplate practices, improve link previews and search visibility, and replace promotional copy with serious, sourced psychology.

None of these messages specified implementation details such as template syntax or test cases. Those were decisions of the agent, and the supervision layer’s job was to keep them checkable.

What Majordomus contributed

Majordomus did not write code. It constrained and recorded the work:

Scoped tasks. Every change ran as a task with declared paths. When files were moved outside that scope, majordomus watch reported drift, and the task was restarted with a correct scope. When a later scope was malformed, majordomus check reported every file as out of scope before the push, which would otherwise have been refused by the pre-push hook.
A health check on every commit. majordomus doctor runs in the pre-commit hook. Before the first commit it reported that the README did not link the agent bootstrap file; the defect was fixed before it reached the repository.
A finish contract. majordomus finish refuses to close a task as completed unless its verification command passes. For this site that command is the live smoke test, so “completed” means “deployed and checked”, not “written”.
Continuity. Checkpoints and handovers recorded the current state and the next action whenever a task was restarted, so each restart took seconds and depended on recorded facts rather than on the agent’s memory.
Rules in the repository. Project rules such as “no deployment claim without live verification” live in .ai/repo/rules/, next to the code they govern.

Did it make development faster?

The honest answer is that no speed-up factor can be stated. There was one session and no comparable session without supervision, so there is no control condition.

What the records do show is where time was not lost. No push was rejected by CI, no deployment had to be rolled back, and the first pipeline run passed all seven jobs. Three defects (two scope errors and a documentation bootstrap failure) became local findings instead of failed pipeline runs. Each would have cost at least one additional push-and-wait cycle of several minutes, and a failed deployment would have cost more.

The larger effect is on confidence rather than raw speed. Every “done” in this project is backed by a recorded command and its exit code, which a reader can check with majordomus history in the repository.

Limits of this account

The timeline starts at the first recorded file; planning that happened before it is not included.
Durations depend on the tools, the hardware and network latency of this session.
The session also relied on conventional safeguards unrelated to Majordomus, including local validation, browser tests and continuous integration. Their effects are not separated here.

Music and emotion regulation: what listening can and cannot change

2026-09-14T00:00:00+00:00

Why people reach for music

Few regulation tools are as easy to use as music. It is available on a phone, it needs no instruction, it can be used alone or with others, and it acts on attention, bodily arousal and meaning at the same time. Researchers who study everyday music use start from this observation. Cook, Roy and Welker (2019) open their study by noting that people frequently use music to regulate their emotions, and Saarikallio (2011) describes emotional self-regulation as one of the most important reasons people engage with music at all ages.

That frequency matters for this library. Music is the most obvious medium of the artifact’s process, repetition, synchronization and discharge, and it is the medium through which many people seek something they would call catharsis. The questions here are therefore practical: what changes when people use music to regulate emotion, how reliable those changes are, and what kind of change the evidence does not cover.

Two terms need to be kept apart throughout. Music interventions are structured uses of music, usually listening, delivered in a study or a care setting and compared with a control condition. Music-based emotion regulation (glossary) is what people do on their own initiative, choosing music to create, change or maintain a state. The first is studied mostly with randomised trials; the second mostly with interviews and questionnaires. They answer different questions.

De Witte, Spruit, van Hooren, Moonen and Stams (2020) pooled randomised controlled trials of music interventions aimed at stress reduction. They ran two multilevel meta-analyses of 104 RCTs, containing 327 effect sizes and 9,617 participants, one for physiological outcomes and one for psychological outcomes. Physiological outcomes included heart rate, blood pressure and hormone levels; psychological outcomes included experiences such as restlessness, anxiety and nervousness.

Music interventions had an overall significant effect on stress reduction in both domains. The standardised mean difference was d = 0.380 for physiological outcomes and d = 0.545 for psychological outcomes. In the conventions commonly used to describe effect sizes, these are small-to-moderate and moderate effects. The type of outcome assessment moderated the results: effects were larger on heart rate (d = 0.456) than on blood pressure (d = 0.343) or hormone levels (d = 0.349).

Three features of this evidence shape how it should be read.

Physiological and psychological outcomes are not the same thing. The larger effect on self-reported stress than on physiological measures is consistent with the general finding that how people feel and what their autonomic and endocrine systems do are related but only partly overlapping (Research note 05). A person can report feeling calmer before heart rate or cortisol changes much, and the reverse can also occur. Neither measure is the “real” stress; they are different signals that support different inferences.

A pooled effect is an average across varied studies. The trials differed in who took part, what music was used, how long sessions lasted and which outcomes were measured. The moderator analysis on outcome type is itself a sign that effects are not uniform. A pooled d does not predict what a particular piece of music will do for a particular person.

The outcomes are stress states, not stressors. The trials measure change in stress-related responses, typically during or shortly after the intervention. They do not measure whether the circumstances that produced the stress changed. That distinction is the subject of Research note 04, and it applies here without modification.

Who reaches for music, and what that does and does not show

Miranda and Blais-Rochette (2020) asked a narrower question: are people who are more emotionally unstable more likely to use music to regulate their emotions? They meta-analysed 13 correlational studies with a total of 2,641 participants and found a significant positive correlation between neuroticism and emotion regulation through music listening, r = .22 (95% CI 0.17 to 0.27), with no evidence of significant heterogeneity across studies. The authors describe the relationship as relatively moderate and suggest that people higher in neuroticism may be more prone to use music listening as an accessible resource for managing negative emotions in everyday life.

This finding is easy to over-read in either direction. It does not show that music use increases neuroticism, and it does not show that music use reduces it. It shows that the two tend to co-occur. A plausible reading is that people who experience negative emotion more often reach more often for a tool that is always at hand. Whether that tool then helps them, leaves them where they were, or in some cases keeps a negative state active, is a separate question that correlational designs cannot answer.

Strategies people describe

Saarikallio and Erkkilä (2007) built a theoretical model of music in adolescents’ mood regulation from group interviews and follow-up forms with eight adolescents, analysed with grounded-theory methods. The model describes mood regulation by music as a process of satisfying personal mood-related goals through various musical activities. The strategies it distinguishes, which later became the basis of questionnaire measures, are:

Entertainment: creating a pleasant atmosphere and maintaining or enhancing a good mood.
Revival: gaining new energy or relaxing when tired or stressed.
Strong sensation: seeking intense emotional experiences.
Diversion: distracting oneself from unwanted thoughts or feelings.
Discharge: releasing negative emotion such as anger or sadness through music that expresses it.
Mental work: using music as a framework for contemplating and reappraising emotional concerns.
Solace: seeking comfort and the feeling of being understood when sad or troubled.

The list is useful because it shows that “listening to music to feel better” covers several different functions. Diversion moves attention away from a concern. Discharge expresses the emotion. Mental work stays with the concern and changes how it is understood, which resembles reappraisal. Solace resembles seeking support. These are not interchangeable, and the research on emotion regulation more broadly suggests they should not be expected to have the same consequences.

Saarikallio (2011) extended the question to adulthood with group interviews of 21 people aged 21 to 70. The basic nature of regulation, including the goals and strategies, remained highly similar across adulthood. Changes were also evident and fell into three types: change with age, fluctuations related to life events, and changes around the transition to retirement. Both studies are qualitative and small. They describe what people do and how they explain it; they do not measure effects.

Cook, Roy and Welker (2019) approached the same topic with a survey of 794 university students, relating music preferences to reported use of music for regulation. Preferences for pop, rap/hip-hop, soul/funk and electronica/dance were positively associated with using music to increase emotional arousal. Soul/funk preference was also associated with using music to increase positive and decrease negative emotion. More broadly, a preference for energetic and rhythmic music was positively associated with every form of musical emotion regulation examined. The authors note that future research would need to test how effective different kinds of music actually are; the study measured preferences and reported use, not outcomes.

A field with loose concepts

Baltazar and Saarikallio (2016) conducted a critical, integrative review of research on affect self-regulation through music, systematically searching publications from January 1994 to June 2014 and selecting 34 for analysis. They examined what the studies said about four levels: goals, strategies, tactics and mechanisms. The levels had not been studied with equal weight or precision, and the review found a considerable degree of inconsistency in terminology and conceptual imprecision across publications, made worse by the lack of a shared model.

This is not a technical footnote. When one study’s “discharge” is another study’s “venting” and a third study’s “emotional release”, pooled conclusions about any one strategy become hard to draw. The practical consequence for readers is caution about claims of the form “music strategy X works”: the strategy may not be defined the same way across the studies being summarised.

Pleasure, reward and careful wording

Salimpoor and colleagues (2011) used [11C]raclopride positron emission tomography together with psychophysiological measures while listeners heard music that produced intense pleasure for them. The PET measure was consistent with endogenous dopamine release in the striatum at peak emotional arousal. With functional MRI in the same listeners and stimuli, activity in the caudate was more associated with anticipation of peak responses, and activity in the nucleus accumbens with the experience of them. The authors conclude that intense pleasure in response to music can lead to dopamine release in the striatal system.

The finding explains part of why music is so rewarding, and why people return to the same pieces. It should not be stretched further. It concerns pleasure at emotional peaks in selected listeners with self-chosen music. It does not show that music regulates distress through dopamine, and reward during listening is not evidence that anything outside the listening episode has changed. Reward is precisely the kind of consequence that reinforces a behaviour whether or not the behaviour addresses a problem (Research note 04).

When music may maintain a state instead of changing it

None of the verified sources in this note measured whether using music for discharge or strong sensation maintains rumination or delays resolution. That gap should be stated plainly rather than filled with a plausible story.

What can be said is narrower. First, discharge through music is the musical counterpart of expressing emotion, and laboratory research on venting anger found that expression focused on a grievance did not reduce anger and aggression (Research note 02). Listening to music that matches one’s sadness is not the same as hitting a punching bag while thinking about a provocateur, and those findings cannot be transferred directly. They do show that the felt release of expression and its measured effect can diverge. Second, the correlation between neuroticism and music-based regulation means that people who most often feel bad are also those most often using music to manage it, so the question of which uses help them is not academic. Third, the strategies Saarikallio and Erkkilä describe differ in whether they keep attention on the concern (discharge, mental work) or move it away (diversion), and differ in whether they change interpretation (mental work) or only the state. Whether those differences lead to different outcomes over time is an open empirical question.

A reasonable working position is that music reliably changes states, that some uses of music engage with the source of a difficulty and others do not, and that the evidence does not yet tell us which use is better for whom.

Music in groups

Much of the research above concerns solitary listening. Music is also a collective activity. Moving and making sound in time with others adds processes that solitary listening lacks: interpersonal synchrony, which in experiments increased liking and cooperation, and shared exertion, which in group dance raised pain thresholds and closeness. These findings are reviewed in Research note 03. They suggest that the social benefits of collective music are real and distinct from its effects on individual mood, and the same limit applies: the outcomes measured are connection and state during or shortly after the activity.

Practical implications, with limits

Use music deliberately for state change. The trial evidence supports using music to reduce stress-related responses in the short term. Choosing music with that aim is a reasonable, low-risk step.
Notice which function you are using. Asking whether a given listening session is for diversion, discharge, solace or working something through makes it easier to see what it can and cannot accomplish.
Check what changed afterwards. If only the state changed, that is relief and it has value. If the same concern returns unchanged each time, pair music with a strategy that addresses it, such as reappraisal.
Do not treat music as a substitute for care. None of the evidence here supports using music in place of treatment for persistent anxiety, depression or trauma-related difficulties.

These implications follow from the evidence but have not themselves been tested as a package. They are reasoned suggestions, not validated protocols.

Limits of the evidence

The trial evidence is broad but heterogeneous in populations, music and outcomes, and it measures short-term stress-related responses. The correlational meta-analysis on neuroticism cannot establish direction or benefit. The strategy research is largely qualitative or survey-based, depends on self-report, and uses concepts that a systematic review found inconsistent. The neuroimaging evidence concerns pleasure at musical peaks, not regulation of distress. Across all of it, the outcome that matters most for the question of catharsis, whether the conditions producing distress change, was not measured. See the evidence ledger for how these grades are assigned and the methods page for the distinction between measurement and interpretation used throughout this library.

Relief is not resolution: reinforcement, avoidance and memory

2026-09-14T00:00:00+00:00

Two outcomes that feel alike

Relief is a change in how a state feels. Resolution is a change in the conditions that produce the state. Both can follow the same evening, and from the inside they are easy to confuse: in both cases the person feels better afterwards.

Relief as a reinforcer

In learning theory, a behaviour that removes an aversive state is negatively reinforced: it becomes more likely to occur again in similar circumstances. Negative reinforcement is adaptive when the behaviour also addresses the source of the aversive state. When it does not, the behaviour can become self-perpetuating, because each repetition is rewarded by relief while the source remains.

Hayes and colleagues (1996) described experiential avoidance as the attempt to escape, suppress or change unwanted internal experiences, such as thoughts, emotions and bodily sensations, even when doing so causes harm over time. They argued that it is a functional process shared by many psychological difficulties, and it later became a central target of acceptance and commitment therapy. Not every cathartic experience is avoidance. But when the main function of an activity is to make an internal state go away, it is worth asking whether the state is being processed or postponed.

How the lift works, and how long it lasts

A meta-analysis of experiments on acute aerobic exercise found a reliable increase in positive activated affect, such as feeling energetic or enthusiastic, after single sessions (Reed & Ones, 2006). Collective rhythm adds reward and social closeness (Research note 03). These are real effects, but they are acute: they describe how people feel during and after an episode, not changes in the conditions they return to.

The body can change state without the world changing with it

Physiological recovery and resolution are easy to conflate because recovery is felt in the body. McEwen (1998) described stress responses as protective in the short term and introduced allostatic load for the cost of their repeated or prolonged activation, including responses that fail to shut off when a challenge has passed. In that framework, returning to a resting state after activation is valuable in itself. It is also, by definition, a change in the body’s state rather than in the challenge.

That is the point where relief and resolution separate. An evening of exertion and music, a good cry or a calming breathing exercise can bring heart rate, muscle tension and felt distress back toward baseline. If the conditions that triggered the response, a conflict, a debt, a loss, are still in place, the same response is likely to be engaged again on return to them. Recovery lowers the cost of each activation; only a change in conditions, or in how they are appraised, reduces how often activation happens.

The distinction also clarifies what different kinds of discharge do. In a meta-analysis of anger management activities, activities that decrease arousal reduced anger and aggression, while activities that increase arousal, such as hitting a bag or running, were ineffective overall for those outcomes (Kjærvik & Bushman, 2024; see Research note 02). The finding concerns anger and aggression, not whether any activity changed the situation that provoked them. The temporal model on the methods page separates these time scales, and the evidence ledger records how strongly each claim is supported.

Repetitive thought

Rumination, a repetitive and passive focus on distress, its causes and its consequences, predicts longer and more intense episodes of negative mood and is associated with the onset of depression (Nolen-Hoeksema, Wisco & Lyubomirsky, 2008). Watkins (2008) showed that repetitive thought is not uniformly harmful. Abstract, evaluative forms (“why does this always happen to me”) tend to be unconstructive, while concrete, process-focused forms (“what exactly happened, and what could I do next”) can support problem solving and recovery.

An intense shared experience can interrupt rumination for hours. When nothing about the way a situation is thought through has changed, the same pattern tends to return.

How relief is remembered

Retrospective evaluations of experiences are dominated by their most intense moment and by how they end, while their duration is largely neglected (Fredrickson & Kahneman, 1993). In a well-known demonstration, participants preferred to repeat a longer episode of cold-water discomfort that ended slightly less unpleasantly over a shorter one that did not (Kahneman et al., 1993).

The same rule applies to pleasant episodes. A strong peak and a good ending make an evening remembered as more restorative than its lasting effect justifies, which makes repeating it attractive.

Reading the figures

Figure 1 shows the loop that follows from these mechanisms: an unresolved state leads into repetition, synchronization and discharge, discharge produces temporary relief, and the unchanged state leads back to the start. Figure 2 contrasts two trajectories: discharge with a sharp drop and a return to the original baseline, and processing with a slower decline to a lower level. The second figure illustrates the argument; it is not a plot of data from any single study.

The practical question

After any cathartic experience, the useful question is not only “did it help?” but “what exactly changed?” If the answer is how you feel, that is relief, and it is valuable. If the answer includes what you understand, decide or do differently, something has begun to resolve. The advice library turns this distinction into concrete strategies.

Rumination and repetitive thought: why going over it again can feel like processing

2026-09-14T00:00:00+00:00

Why this belongs in a library about catharsis

The catharsis idea has a cognitive cousin. Where the hydraulic model says that an emotion must be let out, its cognitive version says that a painful experience must be gone over until it is processed. Going over it again, replaying the conversation, asking why it happened, imagining what should have been said, feels like doing exactly that. Much of the time it does not lead anywhere, and the evidence on rumination helps explain why.

This note distinguishes kinds of repetitive thought, reviews experimental, experience-sampling and clinical evidence on their consequences, and addresses the question in its title: why unconstructive repetitive thought is so easily mistaken for processing.

Not all repetitive thought is the same

Watkins (2008) reviewed the literature on repetitive thought, a broad category that includes rumination, worry, reflection, problem solving and mental rehearsal. Repetitive thought about oneself and one’s concerns is common and can have either constructive or unconstructive consequences. The review proposed that the difference depends on three factors: the valence of the thought content, the context in which the thinking occurs, and the level of construal, that is, whether thinking is abstract or concrete.

The last factor is the most useful in practice. Abstract repetitive thought is evaluative and general: why does this always happen to me, what does it say about me, what if it never changes. Concrete repetitive thought is specific and process-focused: how exactly did the situation unfold, what happened just before things went wrong, what could be done differently next time. The review found that abstract processing tends to be associated with unconstructive outcomes, while concrete processing can support problem solving and recovery.

A related distinction appears in questionnaire research between brooding, a passive comparison of one’s situation with some unachieved standard, and reflection or reflective pondering, a more purposeful turning inward to understand or solve a problem. The term is defined in the glossary.

What rumination is associated with

Nolen-Hoeksema, Wisco and Lyubomirsky (2008) reviewed response styles theory and the research it generated. Rumination, a repetitive, passive focus on one’s distress and its causes and consequences, predicts the onset of depressive episodes and is associated with several other forms of psychopathology; in experiments, positive distraction relieved depressed mood relative to rumination.

At the level of pooled evidence, Aldao, Nolen-Hoeksema and Schweizer (2010) meta-analysed 114 studies relating emotion regulation strategies to symptoms of anxiety, depression, eating and substance-related disorders. Rumination showed the largest association with symptoms of all strategies examined. Because most of the pooled studies are correlational, this establishes a robust association, not that rumination causes symptoms (Research note 08).

Watkins and Roberts (2020) reviewed later research and summarised the consequences attributed to rumination: magnifying and prolonging negative mood, interfering with problem solving and instrumental behaviour, reducing sensitivity to changing contingencies, acting as a vulnerability across anxiety, depression, psychosis, insomnia and impulsive behaviour, interfering with therapy, and maintaining physiological stress responses. These are summaries of a large literature of varying designs, and the causal claims among them rest mainly on the experiments described next.

Experiments: what induced rumination does

Correlations cannot show direction, so experiments that induce rumination are central. Lyubomirsky and Nolen-Hoeksema (1995) ran three studies with dysphoric and nondysphoric students, who were induced either to focus ruminatively on their feelings and personal characteristics or to distract themselves. In the first study, dysphoric participants who ruminated endorsed more negative, biased interpretations of hypothetical situations than dysphoric participants who distracted, or than nondysphoric participants. In the second, dysphoric ruminators were more pessimistic about positive events in their future. In the third, they generated less effective solutions to interpersonal problems. In all three studies, dysphoric participants who distracted were as optimistic and effective as nondysphoric participants.

The result matters for the catharsis question because the ruminative induction was, in content, an invitation to think about one’s feelings. In low mood, that focus did not produce better understanding; it produced more negative interpretation and worse solutions.

Watkins and Moulds (2005) then asked whether the harm lay in focusing on symptoms at all or in the style of that focus. In 40 depressed patients and 40 never-depressed controls, they assessed social problem solving before and after versions of symptom-focused rumination known to induce either abstract or concrete self-focus. In depressed patients, concrete self-focus improved problem solving relative to abstract self-focus. The authors conclude that the particular mode of symptom focus, rather than symptom focus itself, determines the effect of rumination on problem solving. The same topic, thought about in a different mode, had a different consequence.

Daily life: a loop between rumination and mood

Laboratory inductions last minutes. Moberly and Watkins (2008) used experience sampling to see how rumination and mood relate in ordinary life. Ninety-three adults recorded momentary ruminative self-focus and negative affect at quasi-random intervals eight times a day for a week. Dispositional rumination scores were associated with more momentary rumination over the week. At the same moment, rumination was positively associated with negative affect. Over time, ruminative self-focus predicted negative affect at the next occasion, and negative affect also predicted ruminative self-focus at the next occasion. Brooding, but not reflective pondering, was associated with higher mean negative affect.

The reciprocal pattern is consistent with a self-sustaining loop: low mood prompts rumination, and rumination deepens low mood. This is an association in repeated observations, not a demonstration of cause, but it fits the experimental findings and helps explain why rumination can persist for hours or days once it starts.

Why going over it again feels like processing

The research above does not directly measure what ruminating feels like from the inside, so this section is an interpretation that is consistent with the evidence rather than a finding.

First, rumination is about the problem. It has the same topic as problem solving, and it often carries the intention to understand. Watkins (2008) emphasises that repetitive thought about a concern can be constructive, which makes it difficult to tell from within whether a particular episode is helping.

Second, abstract questions feel deep. “Why did this happen?” and “what does it mean about me?” seem more serious than “what will I say on Tuesday?” Yet the experiments show that it is the concrete mode that improved problem solving.

Third, Watkins and Roberts (2020) propose a model, which they call H-EX-A-GO-N, in which rumination develops from dwelling on unresolved goals into a learnt habit, involving the tendency to process negative information abstractly, particularly with poor executive control and negative biases in information processing. If rumination is partly a habit triggered by reminders of unresolved goals, it will start automatically and feel purposeful even when it produces nothing new. The model is a proposal that integrates experimental evidence, not an established account.

Fourth, rumination can bring a kind of temporary relief: the sense of having attended to the problem. Relief that follows a behaviour tends to make that behaviour more likely, whether or not it solves anything (Research note 04). That link has not been tested directly for rumination in the sources cited here.

A practical test follows from this analysis. Processing tends to yield something: new information, a changed interpretation, a decision, or an action. A loop that returns to the same question in the same words, with the same feeling, and yields none of these is more likely to be rumination than processing.

What reduces unconstructive repetitive thought

Several lines of evidence point to ways of changing rumination rather than trying to exhaust it.

Changing the mode. The experiments above show that shifting from abstract to concrete thinking about the same problem changed its effect on problem solving. Asking how, specifically, rather than why, is the core of this shift.

Changing the perspective. Kross, Ayduk and Mischel (2005) found that reflecting on an upsetting experience from a self-distanced perspective, as an observer, reduced emotional reactivity and rumination compared with immersed recall.

Changing attention. In the experiments of Lyubomirsky and Nolen-Hoeksema, distraction left dysphoric participants as optimistic and effective as nondysphoric participants, and the review by Nolen-Hoeksema and colleagues reports that positive distraction relieved depressed mood. Distraction does not solve the problem, but it can interrupt the loop long enough for concrete thinking to become possible later.

Containing it in time. McGowan and Behar (2013) randomly assigned 53 people with high trait worry to two weeks of stimulus control training, a 30-minute time- and place-restricted worry period each day, or to a control condition of focused worry, in which participants were instructed not to avoid naturally occurring worry. Stimulus control produced greater reductions in worry, anxiety, negative affect and insomnia, though not in depression or positive affect. The study concerns worry rather than rumination, is small and preliminary, and cannot show long-term effects.

Treatment that targets rumination. Watkins and colleagues (2011) randomised 42 people with medication-refractory residual depression to treatment as usual or treatment as usual plus up to 12 sessions of rumination-focused cognitive-behavioural therapy. Adding the rumination-focused therapy improved residual symptoms and remission rates, and the effects were mediated by change in rumination. The trial lacked an attentional control group, so it cannot separate the specific content of the therapy from non-specific effects of additional therapy.

Sleep. In a meta-analysis of 65 trials, interventions that improved sleep also reduced rumination (g+ = -0.49) (Scott et al., 2021; Research note 13).

Relevance to catharsis

Rumination and venting share a structure. Both keep attention on the source of distress, both feel like doing something about it, and both can be followed by a sense of having discharged or processed something. In the venting experiments, hitting a bag while thinking about the provocateur, a combination of exertion and rumination, produced the most anger (Research note 02). In the rumination experiments, focusing on one’s feelings in low mood produced more negative interpretation and worse problem solving.

The alternative in both cases is not suppression. It is a change in how the problem is held: concretely rather than abstractly, from a distance rather than immersed, in a bounded time rather than all day, and with a view to a next step.

Practical implications, with limits

Notice the loop. The same question, in the same words, with no new information, is a signal to change mode.
Switch from why to how. Describe concretely what happened, when and in what sequence, and what could be done next.
Step back. Describe the situation as an observer would.
Set a time for it. If nothing can be done now, postpone the concern to a fixed, limited period.
Use absorbing activity to interrupt. Distraction is not a solution, but it can break the loop.

These are reasoned from the evidence; the combination has not been tested as a package, and the advice on interrupting rumination with concrete steps sets out its limits. Rumination that persists for weeks with low mood or anxiety warrants professional assessment; structured treatments that target it exist.

Limits of the evidence

Most evidence linking rumination with symptoms is correlational. Experimental inductions are brief and mostly conducted with students or small clinical samples. The experience-sampling study shows reciprocal association, not cause. The treatment trial was small and lacked an attentional control, and the worry study was preliminary and concerned worry. The explanation of why rumination feels like processing is an interpretation consistent with the evidence, not a measured finding. The evidence ledger records the grade and review date of each claim.

Sleep and emotion regulation: what experiments on sleep loss show

2026-09-14T00:00:00+00:00

Why sleep belongs in a library about catharsis

Emotional strain and sleep are entangled in ordinary experience. After an argument, a loss or a stressful day, people lie awake replaying events, and the next day everything feels harder. A library about release and regulation might seem to have little to say about sleep, because sleep is not something a person does to an emotion. But that is precisely why it is useful: sleep changes the conditions under which every other regulation strategy operates, and, unlike most of those strategies, it can be manipulated in controlled experiments with healthy volunteers.

This note separates three kinds of evidence. Experiments that deprive or restrict sleep show what sleep loss does to emotional functioning over a day or a few days. Longitudinal studies show how poor sleep relates to later mental health over months and years. Trials of sleep interventions provide an indirect test of whether improving sleep improves mental health. Each answers a different question.

Terms: deprivation, restriction and fragmentation

Sleep deprivation in research usually means total sleep deprivation: staying awake through a night under supervision. Partial sleep restriction shortens sleep, for example to a few hours a night for several nights. Sleep fragmentation interrupts sleep with repeated awakenings. These manipulations resemble different real-life situations: a night shift, a week of short nights before a deadline, or broken sleep while caring for an infant. They are not the same as insomnia, which is a persistent difficulty falling or staying asleep, and they do not model what happens over years.

What the experiments show

Palmer and colleagues (2024) conducted a preregistered systematic review and meta-analysis of more than fifty years of experimental research. They included studies that experimentally reduced sleep via total sleep deprivation, partial restriction or fragmentation in healthy populations and measured positive affect, negative affect, general mood disturbance, emotional reactivity, anxiety symptoms or depressive symptoms. The analysis covered 1,338 effect sizes across 154 studies with 5,717 participants aged 7 to 79.

All forms of sleep loss reduced positive affect, with standardised mean differences between -0.27 and -1.14 depending on the form of sleep loss. All forms increased anxiety symptoms (SMD 0.57 to 0.63). And all forms blunted arousal in response to emotional stimuli (SMD -0.20 to -0.53). Findings for negative affect, for ratings of how pleasant or unpleasant emotional stimuli were, and for depressive symptoms were mixed and depended on the type of sleep loss. The authors also detected nonlinear effects of the amount of sleep lost and differences depending on which stage of sleep was restricted, rapid eye movement sleep or slow-wave sleep. They conclude that extended wakefulness, shortened sleep and night-time awakenings adversely influence emotional functioning.

Tomaso, Johnson and Nelson (2021) ran three related meta-analyses on 241 effect sizes from 64 studies, from early childhood to late adulthood. Sleep loss had a moderate effect on negative mood (g = 0.45), stronger in younger samples and larger for total deprivation than for restriction. It had a large effect on positive mood (g = -0.94). After correction for publication bias, there was a modest negative effect on emotion (g = -0.11), which did not depend on whether the emotional stimuli were positive or negative. Sleep restriction had a small negative effect on adaptive emotion regulation (g = -0.32) and no significant effect on maladaptive emotion regulation (g = 0.14). All studies of adaptive emotion regulation had been conducted with youth samples.

Reading the pattern carefully

Three features of these results deserve attention because they depart from the simplest story.

The most consistent change is less positive affect, not only more negative affect. Both syntheses find consistent reductions in positive mood: large in one (g = -0.94) and small to large, depending on the type of sleep loss, in the other. Effects on negative affect are present in one synthesis and mixed in the other. A tired person may not feel dramatically worse so much as flatter, with less capacity for enjoyment and interest. That loss matters for regulation, because positive experiences are among the resources that help people recover from stress.

“Reactivity” is not one thing. Popular accounts often say that sleep loss makes people emotionally overreactive. Palmer and colleagues found that sleep loss blunted arousal responses to emotional stimuli, and Tomaso and colleagues found a small effect on emotional responses that did not depend on valence. At the same time, anxiety symptoms rose. These can coexist: a person can feel more anxious overall and respond with less differentiated arousal to specific stimuli in a laboratory task. The honest reading is that sleep loss disturbs emotional functioning in several ways, and that “more reactive” is too simple a summary.

Regulation has been studied less than emotion. Palmer and Alfano (2017) reviewed the field using the process model of emotion regulation (Gross, 1998) as an organising framework. They observed that most sleep-emotion research had focused on emotions generated as a consequence of inadequate sleep rather than on the regulatory processes that may alter those experiences, with constructs often poorly delineated. They found evidence of maladaptive changes at multiple stages of emotion generation and regulation and called for experiments that identify which points in the regulation process are most vulnerable to sleep loss. The small effect on adaptive regulation in the Tomaso meta-analysis, limited to young people, shows how thin the direct evidence on regulation still is.

A lowered threshold for stress

Minkel and colleagues (2012) asked whether sleep loss simply amplifies stress responses or interacts with stressors in a more specific way. In two laboratory experiments with a total of 53 healthy adults, one group was kept awake for a night and the other slept. Participants then completed cognitive tasks under low-stress or high-stress conditions, manipulated through task difficulty, time pressure and feedback. After the low-stress condition, sleep-deprived participants reported greater subjective stress, anxiety and anger than rested controls. After the high-stress condition, negative mood and stress rose about equally in both groups.

The authors suggest that sleep deprivation lowers the psychological threshold at which cognitive demands are perceived as stressful, rather than selectively increasing the size of negative responses to severe demands. The study is small and concerns one night of total deprivation, but the pattern is plausible and practically important. After a short night, ordinary irritations that would otherwise pass may be experienced as stressful, and the anger they produce may be read as a sign that something must be said or done.

Sleep and emotional memory: a proposed model

Walker and van der Helm (2009) reviewed findings from basic and clinical research and proposed a model in which sleep modulates affective neural systems overnight and reprocesses recent emotional experiences, which would help restore appropriate next-day reactivity. They proposed a specific hypothesis about rapid eye movement sleep and emotional memory processing, and suggested that it may offer insight into the association between sleep abnormalities and mood disturbances.

The model is influential and has stimulated much of the later experimental work. It is a hypothesis, not an established mechanism. The later meta-analytic finding that sleep loss blunts rather than heightens arousal responses to emotional stimuli, and that effects depend on the sleep stage restricted, is a reminder that the neural account remains under active test. Imaging findings in this area describe measured activity associated with sleep conditions; they do not show what a person feels (Research note 05).

Over months and years: insomnia and depression

Experiments last a night or a week. For longer time scales the evidence is observational. Baglioni and colleagues (2011) meta-analysed 21 longitudinal epidemiological studies that measured insomnia complaints at baseline and depression at follow-up. Using a random-effects model, the overall odds ratio for insomnia predicting depression was 2.60 (95% CI 1.98 to 3.42). When outliers were removed, the studies were no longer heterogeneous and the fixed-effects odds ratio was 2.10 (95% CI 1.86 to 2.38). The authors summarise this as non-depressed people with insomnia having about twice the risk of developing depression compared with people without sleep difficulties. They note that the included studies did not always consider other intervening variables.

An association of this kind cannot show that insomnia causes depression. Insomnia can be an early sign of a developing depressive episode, and both can share causes such as chronic stress. What the finding does show is that persistent poor sleep is not a trivial companion of distress.

Trials that improved sleep

A stronger test of the causal question comes from randomised trials. Scott and colleagues (2021) meta-analysed 65 randomised controlled trials, comprising 72 interventions and 8,608 participants, that reported the effects of an intervention that improved sleep on mental health. Improving sleep led to a medium effect on composite mental health (g+ = -0.53), depression (g+ = -0.63), anxiety (g+ = -0.51) and rumination (g+ = -0.49), a small-to-medium effect on stress (g+ = -0.42) and a small effect on positive psychosis symptoms (g+ = -0.26). Greater improvements in sleep quality led to greater improvements in mental health, a dose-response relationship. The authors interpret the findings as suggesting that sleep is causally related to the experience of mental health difficulties.

Two cautions apply. The meta-analysis selected trials in which the intervention improved sleep, which is appropriate for the question asked but means it does not estimate how often attempts to improve sleep succeed. And the abstract does not break the effects down by type of intervention, so the result supports the value of better sleep, not of any particular technique.

The effect on rumination is worth noting for this library. Repetitive negative thinking is one of the processes most consistently associated with prolonged distress (Research note 15), and lying awake is often where it happens.

What about everyday sleep habits?

Advice on sleep usually takes the form of sleep hygiene: regular timing, limiting caffeine and alcohol, a quiet room, and so on. Irish and colleagues (2015) reviewed the empirical evidence for such recommendations with an emphasis on the general population. They note that sleep hygiene education has been ineffective as a treatment in clinical sleep medicine. Epidemiological and experimental research generally supported associations between individual recommendations and night-time sleep, but the direct effects of individual recommendations on sleep remained largely untested in the general population.

This is a useful correction to confident lists of sleep rules. The habits are reasonable, low-risk and associated with better sleep, and the evidence that following them improves sleep for a given person is weaker than their popularity suggests. Persistent insomnia is a different matter, for which structured treatments delivered by clinicians exist.

Relevance to catharsis and regulation

The findings above suggest a practical inversion of the catharsis intuition. After an emotional strain, the impulse is often to process it immediately and intensely: to talk late into the night, to write until the early hours, or to replay the event until it feels settled. If that comes at the cost of sleep, the next day begins with less positive affect, more anxiety and a lower threshold for stress, conditions under which the strain is likely to feel larger, not smaller.

Protecting sleep does not resolve anything by itself. It preserves the capacity to regulate, to reappraise and to act the following day. The emotional experience that seems to demand release at midnight is not the same experience in the morning, and the evidence does not support the idea that it must be dealt with before sleeping.

Practical implications, with limits

Treat sleep as part of recovery after a strain. Where possible, keep the usual sleep window rather than using the night to process the event.
Expect a lower threshold after a short night. Irritation and anxiety the next day may partly reflect sleep loss; postpone consequential conversations or decisions when possible.
If thoughts keep you awake, park them. Writing down the concern and a concrete next step, and returning to it at a set time the next day, is consistent with evidence on repetitive thought, though it has not been tested as a sleep intervention here (Advice).
Seek help for persistent insomnia. Difficulty sleeping on most nights for weeks, especially with low mood or anxiety, warrants talking with a clinician; effective treatments exist.

These implications are reasoned from the evidence; they have not been tested as a package.

Limits of the evidence

Experimental studies manipulate short periods of sleep loss in healthy volunteers, often in laboratories, and their emotional measures vary widely. Evidence on regulation processes specifically is limited, and one set of findings on adaptive regulation comes only from young people. Longitudinal associations between insomnia and depression cannot establish direction. The trial meta-analysis supports improved sleep in general rather than specific everyday habits, and the evidence for individual sleep hygiene recommendations in the general population is weak. The neural model of overnight emotional processing is a hypothesis. The evidence ledger records the grade and review date of each claim.

Stress physiology, social buffering and the body in the crowd

2026-09-14T00:00:00+00:00

Stress as a physiological process

The body responds to challenge with a coordinated set of changes: activation of the sympathetic nervous system and release of catecholamines within seconds, and activation of the hypothalamic–pituitary–adrenal axis and release of cortisol over minutes. These responses mobilise energy, sharpen attention and support action. They are adaptive.

McEwen (1998) described this process as allostasis, achieving stability through change, and introduced the concept of allostatic load for the cost of repeated or prolonged activation. Stress mediators protect in the short term but contribute to damage when they are activated too often, when the body fails to habituate to repeated challenges, when the response does not switch off after the challenge ends, or when the response is inadequate. Over time, allostatic load has been linked to cardiovascular, metabolic, immune and neural changes. The central point for this library is that recovery after activation matters as much as the activation itself.

Regulating the response

Ochsner and Gross (2005) reviewed neuroimaging studies of the cognitive control of emotion. Strategies such as reappraisal recruit prefrontal and cingulate regions associated with cognitive control, and these regions modulate activity in systems that generate emotional responses, including the amygdala. Controlled attention and reinterpretation therefore have measurable neural consequences, which is consistent with behavioural evidence that reappraisal changes emotional experience (Research note 08).

Cohen and Wills (1985) reviewed evidence on whether social relationships protect health directly or only under stress. They found support for both. Integration in a social network had a general beneficial association with well-being, while the perceived availability of support, the belief that others would help if needed, protected people specifically against the adverse effects of stressful events. This buffering effect was most evident when the support available matched the needs created by the stressor.

The distinction is useful for thinking about crowds. Collective experiences can strengthen a sense of belonging and closeness (Research note 03). Buffering, however, depends on support that is perceived as available when a particular stressor arises, which is a property of ongoing relationships rather than of a single event.

Music as a regulation tool

Saarikallio and Erkkilä (2007) interviewed adolescents about how they use music in everyday life and identified several distinct mood-regulation strategies. They included creating a pleasant atmosphere, reviving energy, seeking strong sensations, diverting attention from worries, working through thoughts and feelings, finding solace, and discharge, the release of negative emotion such as anger or sadness through music that matches or expresses it. Participants experienced music as an effective and flexible way to manage mood.

The study is qualitative and involves adolescents, so it describes strategies rather than measuring their long-term effects. It is valuable here because it shows that people already distinguish discharge from other uses of music, such as working through a problem.

The body in the crowd

A night of dancing combines several of the processes described in this library. Sustained exertion produces an acute lift in positive affect (Reed & Ones, 2006). Synchrony and exertion raise pain thresholds and increase closeness (Tarr et al., 2015). Loud, rhythmic music supports discharge and diversion. The shared setting provides, at least for a few hours, a strong sense of belonging.

Each of these is a real physiological and psychological effect. None of them removes a chronic stressor. If the conditions that produce repeated activation, such as financial strain, conflict, isolation or overwork, persist, the stress response will be engaged again when a person returns to them. From the perspective of allostatic load, the useful questions are whether the body gets adequate recovery between activations and whether anything reduces how often the stress response has to be engaged. Relief contributes to the first. Only change contributes to the second.

The venting hypothesis: what experiments show about letting anger out

2026-09-14T00:00:00+00:00

The hypothesis

The hydraulic model of emotion predicts that expressing anger discharges it. Hitting a punching bag, shouting into a pillow or playing an aggressive game should therefore lower anger and reduce later aggression. The advice has been repeated in popular psychology for decades, and it feels plausible because venting is often followed by a sense of release.

The experiments

Bushman, Baumeister and Stack (1999) tested the idea directly. Participants read a message claiming that hitting a punching bag relieves anger, a message claiming the opposite, or a neutral message, and were then provoked. Those who read the pro-catharsis message were more inclined to choose the punching bag. In a follow-up, participants who had vented were subsequently more aggressive toward the person who had provoked them, not less. Belief in catharsis did not protect against aggression; it steered people toward a strategy that increased it.

Bushman (2002) then compared three responses to provocation. One group hit a punching bag while thinking about the person who had angered them (rumination). A second group hit the bag while thinking about getting physically fit (distraction). A control group sat quietly for the same time. The rumination group reported the most anger and behaved most aggressively. The control group, which did nothing, behaved least aggressively. Distraction fell in between.

What the evidence does and does not show

The studies are laboratory experiments with brief provocations, mostly with university students, and aggression is measured with standardised tasks such as choosing the intensity of an unpleasant noise delivered to another participant. They show that venting anger does not reduce short-term aggression and can increase it. They do not show that every form of emotional expression is harmful, and they do not describe what happens over months.

The proposed mechanism is rehearsal. Venting while attention stays on the provocation keeps anger-related thoughts active and makes aggressive responses more accessible. Physical arousal alone is not release: exertion raises arousal, and if attention remains on the grievance, that arousal is experienced as anger.

What reduces anger instead

Strategies that lower arousal, such as waiting, slow breathing or genuine distraction, let anger decline without rehearsing it. Cognitive reappraisal changes the interpretation of the provocation before or while the emotion unfolds; people who habitually use reappraisal report more positive and less negative emotion than people who habitually suppress (Gross, 1998; Gross & John, 2003). Reflecting on an upsetting experience from a self-distanced perspective, as an observer rather than reliving it, reduced emotional reactivity and rumination compared with immersed recall (Kross, Ayduk & Mischel, 2005).

These strategies share one feature. They change the relationship to the emotion rather than trying to empty it out.

A meta-analytic test: turning arousal up or down

The experiments above are single studies with particular procedures. Kjærvik and Bushman (2024) asked the broader question across the literature: do anger management activities that increase physiological arousal work differently from activities that decrease it? Arousal-decreasing activities included deep breathing, mindfulness and meditation. Arousal-increasing activities included hitting a bag, jogging and cycling. Their meta-analytic review was based on 154 studies with 184 independent samples and 10,189 participants.

Arousal-decreasing activities reduced anger and aggression, with a pooled effect of g = -0.63 (95% interval -0.82 to -0.43), and the result was robust. The effect held across genders, races, ages and cultures; in students and non-students, criminal offenders and non-offenders, and people with and without intellectual disabilities; whether delivered digitally, by researchers or by therapists; in group and individual sessions; and in field and laboratory settings. Arousal-increasing activities, by contrast, were ineffective overall (g = -0.02, interval -0.13 to 0.09), and their results were heterogeneous and complex. The authors conclude that the findings do not support venting anger or going for a run as effective anger management activities.

Two cautions keep the result in proportion. The null average for arousal-increasing activities means no reliable reduction in anger on average, not that exercise is harmful; the heterogeneity means some studies found different results, and the pooled estimate does not identify which conditions made the difference. And the review concerns anger and aggression, not mood, fitness or social connection, where exercise has other documented benefits (Research note 04).

Catharsis is not the same as venting

The meta-analysis sharpens a distinction that everyday language blurs. Catharsis, in the sense this library examines, names an experienced release: the feeling, after expressing an emotion, that something has been let out. Venting names a behaviour: expressing anger, often with high physical arousal, in the expectation that this reduces it.

The two can come apart. A person can feel released after hitting a bag and still be as angry, measured by self-report or behaviour, as someone who did nothing, or angrier. The feeling of release is real as an experience, and it is also what makes venting attractive to repeat. It is not a measurement of whether anger and aggression have gone down. On the measured outcome, the evidence favours lowering arousal over raising it. The distinction between how a change feels and what it changes is developed in Research note 04.

Relevance to the artifact

The process shown in the artifact, repetition, synchronization and discharge, is not venting anger at a target, and the findings above should not be stretched to cover dancing or music. Their relevance is narrower and still important: the feeling of release is not evidence that the cause of an emotion has been dealt with, and some kinds of release keep the emotion rehearsed. That is the distinction the next notes develop.

What catharsis is: from purgation to an empirical question

2026-09-14T00:00:00+00:00

A word with three careers

The Greek katharsis meant cleansing. In medical writing it described purging the body, and in religious practice ritual purification. Aristotle gave the word its most influential aesthetic sense: in the Poetics he describes tragedy as achieving, through pity and fear, the katharsis of such emotions (1449b).

Aristotle did not define the term further, and commentators have argued for centuries about what he meant. Three readings recur. Purgation treats emotions as something expelled. Purification treats them as refined into a better form. Clarification treats catharsis as a kind of understanding reached through the drama. The disagreement is not only philological. It already contains the question this project is built around: does release change something, or does it only relieve?

The clinical version

In 1895 Josef Breuer and Sigmund Freud published Studies on Hysteria. They described a cathartic method in which patients recalled distressing memories, often under hypnosis, and expressed the emotion attached to them. They called the discharge abreaction. Breuer’s patient, known as Anna O., called the treatment the “talking cure”.

Freud later abandoned hypnosis and the cathartic method in favour of free association and the analysis of resistance. One reason he gave was that relief after abreaction was frequently temporary: symptoms returned when the underlying conflict had not been worked through. The method faded in clinical practice, but its hydraulic picture survived in popular culture. Emotions build pressure, and pressure has to be let out.

The empirical version

Modern psychology turned catharsis into a claim that can be tested: does expressing an emotion reduce it?

For anger, controlled experiments found the opposite of what the hydraulic model predicts. Venting anger while focusing on its cause kept the anger active and was followed by more aggression, not less (Research note 02).

For distressing experiences more broadly, the picture is more nuanced. In studies of expressive writing, participants who wrote about both the facts and their feelings about a traumatic event later showed health benefits compared with control groups (Pennebaker & Beall, 1986). Subsequent analyses suggested that benefit was linked to building a coherent account of the experience, that is, to cognitive processing, rather than to emotional discharge alone (Pennebaker, 1997).

Research on emotion regulation reframed the question again. Instead of a binary choice between release and suppression, it asks which strategies change the course of an emotion, at what point in its unfolding, and with what costs (Gross, 1998).

A working definition

In this project, catharsis means a subjectively felt release of emotional tension that follows expression, physical exertion or immersion in an experience. The definition deliberately says nothing about whether the cause of that tension has changed. Whether it has is the empirical question, and the answer differs by emotion, by method and by context.

Why the distinction matters

When relief is mistaken for resolution, people repeat whatever made them feel better and are surprised when the underlying difficulty returns. The artifact’s status line, problem_solved: false, is a compact way of keeping those two outcomes apart. The following notes examine each part of that claim: anger and venting, collective rhythm, relief and reinforcement, and the limits of measuring emotion at all.

Allostasis, allostatic load and stress recovery

2026-09-14T00:00:00+00:00

What “the body changes state” can mean

Accounts of catharsis often have a bodily plot. Before, the chest is tight and the jaw clenched; during, there is shouting, crying or exertion; after, the body is loose and quiet. That description is phenomenologically accurate for many people. The question for this essay is what physiology says about such changes of state, and what it does not.

Stress physiology offers a vocabulary for this: activation, mediators, reactivity, recovery, and two concepts that attempt to connect short-term responses with long-term health, allostasis and allostatic load. Each has a defined meaning, a body of evidence and critics. This essay explains them in turn and then asks what they imply about relief and resolution.

From homeostasis to allostasis

Homeostasis

The classical idea of physiological regulation is homeostasis: internal variables such as temperature, blood glucose and blood pressure are kept within a range around a set point, and deviations are corrected by feedback. When something pushes a variable away from its set point, mechanisms detect the error and push it back.

Allostasis as predictive regulation

Peter Sterling and Joseph Eyer introduced the term allostasis in a 1988 chapter on arousal pathology (Sterling & Eyer, 1988). The word is usually glossed as “stability through change”. Sterling later set out the argument in detail (Sterling, 2012). The premise of the standard homeostatic model, he argued, is flawed: the goal of regulation is not to preserve constancy of the internal milieu but to continually adjust it to promote survival and reproduction. Feedback, which corrects errors after they occur, is inherently inefficient. Efficient regulation instead requires anticipating needs and preparing to satisfy them before they arise.

Sterling lists advantages of such predictive regulation: errors are reduced in magnitude and frequency; the capacities of different components are matched; resources are shared between systems; and errors are remembered and used to reduce future errors. In his account this strategy requires central coordination by the nervous system, which integrates many variables with prior knowledge to predict needs, set priorities and govern anticipatory behaviour, such as moving somewhere warmer before becoming cold.

The practical consequence of the idea is that set points are not fixed. Blood pressure rises in anticipation of standing up or of a demanding meeting, not only in response to a fall. Physiological change during emotional episodes is, on this view, largely anticipatory preparation for action (allostasis).

Allostatic load

The concept

McEwen and Stellar (1993) extended allostasis over time. Their article emphasises the hidden cost of chronic stress to the body over long periods, which acts as a predisposing factor for the effects of acute stressful events. They define allostatic load as the cost of chronic exposure to fluctuating or heightened neural or neuroendocrine response resulting from repeated or chronic environmental challenge that an individual reacts to as particularly stressful. The formulation stresses cascading relationships, beginning early in life, between environmental factors and genetic predispositions, which lead to large individual differences in susceptibility.

Protective and damaging effects

A later review put the central idea in terms of mediators (McEwen, 1998). Stress mediators such as cortisol and catecholamines are protective in the short term: they mobilise energy, support cardiovascular responses and prepare the organism for demands. When they are chronically activated or poorly terminated, they contribute to damage. The same systems that help in an acute challenge can, when repeatedly or persistently engaged, contribute to wear and tear (claim).

The reasoning here is important for catharsis. Allostatic load is not a quantity of stored stress waiting to be released. It is the cumulative physiological cost of how often, how strongly and for how long systems have been activated. Nothing in the concept suggests that a single intense discharge reduces it.

Measuring load

To test the idea, researchers constructed indices that combine biomarkers from several systems, typically neuroendocrine, immune, metabolic and cardiovascular. A review of this work reported that numerous studies using such an index demonstrated greater prediction of morbidity and mortality beyond traditional detection methods used in biomedical practice (Juster, McEwen, & Lupien, 2010). The review also identifies risk and protective factors associated with increased load.

The site grades this with low confidence. The studies are observational; the composition of the index differs between studies, which makes comparison and synthesis difficult; the review is narrative rather than systematic; and prediction does not establish that the index measures a single underlying process.

What happens during an acute stress response

A brief, deliberately general description helps to interpret claims about bodily “release”. During an acute challenge, activity of the sympathetic branch of the autonomic nervous system typically increases, with effects on heart rate, blood pressure, breathing and sweating, while parasympathetic influence on the heart is reduced. The hypothalamic–pituitary–adrenal axis responds more slowly, with cortisol rising later than the autonomic response and remaining elevated after it. When the challenge ends, these systems return towards baseline at different speeds.

Three cautions follow from the evidence base of this site.

These responses are not emotion-specific. A meta-analysis of 202 studies found that autonomic responses did not clearly distinguish one emotion category from another and varied substantially within categories (Siegel et al., 2018). Exertion, anticipation, excitement and anger can all raise heart rate (Theory 04).
A return to baseline is not the removal of a cause. Physiological variables return towards baseline after most challenges whether or not the challenge has been resolved.
Timing differs across systems. A person can feel calm while a slower system is still elevated, or feel unsettled after heart rate has normalised.

Recovery: the neglected half

If allostatic load reflects how long activation lasts, then recovery should matter as much as reactivity. Linden, Earle, Gerin and Christenfeld (1997) examined this. Of 105 studies of physiological reactivity published in four major journals over two years, 63% had collected recovery data, but only 23% reported them. The authors concluded that recovery had been neglected, argued for its conceptual and ecological importance, and noted that studies using recovery protocols revealed findings that were not apparent from reactivity comparisons alone.

The point is methodological but has a direct bearing on catharsis. A cathartic episode is usually described in terms of recovery: the state after. Physiologically, recovery can be fast or slow, complete or incomplete, and it can be interrupted by renewed activation. How recovery unfolds after an emotional episode is a separate question from how intense the episode was.

Perseverative cognition

Why would activation persist after a stressor has ended? Brosschot, Gerin and Thayer (2006) proposed the perseverative cognition hypothesis. Worry and rumination are common responses to stress, yet biopsychological models had largely focused on activation during stressors. The authors argued that perseverative cognition prolongs stress-related affective and physiological activation, both in anticipation of stressors and after them. They reviewed evidence that worry, rumination and anticipatory stress are associated with enhanced cardiovascular, endocrine, immune and neurovisceral activity, and described the findings as preliminary support (perseverative cognition).

This connects stress physiology with emotion regulation. A person who repeatedly rehearses a conflict, whether alone or while venting, may extend the physiological episode as well as the emotional one. The association is not evidence that rumination causes disease, and the authors are explicit about the preliminary status of their hypothesis. It is, however, a plausible route by which the way an emotion is handled could affect the duration of activation. See also rumination.

Physiological stress does not occur in isolation. The perceived availability of support has been found to buffer the effects of stress, while social integration has a main effect on wellbeing, largely in observational studies (Cohen & Wills, 1985). The Research note on stress and social buffering discusses this. For allostatic load, the relevant implication is that the frequency and duration of activation depend on circumstances and relationships as well as on individual responses.

The critique: does allostasis add anything?

The allostasis framework is widely used but not universally accepted. Day (2005) offers a direct critique. Examining the arguments for supplementing homeostasis with allostasis and allostatic load, he concludes that the concept of allostasis was introduced largely through misunderstandings of homeostasis; that, in terms of understanding how organisms operate, it is not clear that allostasis or allostatic load offer anything not already apparent or derivable from an accurate reading of homeostasis; and that attempts to subsume stress under allostasis collapse the study of homeostatic responses and stress responses together, when there is evidence for pathways that merit the title of stress circuitry.

The site records this critique as a contested position. Its force is conceptual. It does not deny that chronic stress has costs, or that anticipatory regulation occurs. It questions whether new terms are needed, and whether they clarify or blur the concept of stress. Readers should treat “allostatic load” as a useful organising idea whose boundaries are disputed, not as a measured quantity.

Allostasis, catharsis and resolution

What does this body of theory imply for the idea that catharsis releases stress? The following is interpretation, drawing on the evidence above.

Release is not the same as recovery. Many cathartic activities, such as shouting, hitting, crying intensely or exhausting exercise, increase physiological activation before it falls. The felt relief afterwards coincides with a return towards baseline. That return would occur in any case once activation stops. The meta-analytic finding that arousal-decreasing activities reduced anger while arousal-increasing activities were not effective overall is consistent with this reading (Research note 02).

Load is cumulative, not stored. On the allostatic load model, what matters for long-term cost is repeated and prolonged activation. An episode of release does not subtract from load; if it adds another bout of activation, or is followed by rumination, it may extend it. There is insufficient evidence to quantify this for any particular activity.

Resolution acts on exposure and duration. If a stressor recurs, or if it is repeatedly rehearsed in thought, activation recurs. Changes that reduce recurrence (changing a situation), shorten activation (effective regulation, support) or reduce perseverative cognition act on the variables the model identifies. Relief that leaves those unchanged leaves the pattern of activation unchanged.

Feeling calm is informative but limited. A subjective sense of calm after an episode is associated with physiological recovery in many situations, but systems recover at different speeds, and self-report is not a measure of load.

Clinical boundary

Persistent physical symptoms, sleep problems, or distress that does not ease are reasons to consult a medical or mental health professional. Concepts such as allostatic load are research constructs; they are not clinical tests, and nothing on this site assesses any reader’s stress physiology.

What this essay concludes

Stress physiology gives a precise sense to the idea that the body changes state. Activation is anticipatory and protective in the short term; its costs accumulate with repetition and duration; recovery is a distinct and often neglected process; and worry and rumination are associated with activation that outlasts its cause. The allostasis framework organises these ideas, and it is contested as adding little to homeostasis. None of it supports a picture of stress as a stored quantity released by an intense episode. On this account, what shapes long-term physiological cost is how often and how long systems are engaged, which is why the question of what changed after relief matters physiologically as well as psychologically.

Appraisal theories of emotion

2026-09-14T00:00:00+00:00

The problem appraisal theories solve

Three people receive the same message from a manager: “Can we talk tomorrow morning?” One feels anxious, one feels curious, and a third, who has been waiting for a promotion decision, feels hopeful. If emotions were direct reactions to stimuli, identical stimuli should produce identical emotions. They do not. And the same person can feel different emotions about the same event at different times, as its meaning changes.

Appraisal theories explain this variability by placing an evaluation between the event and the emotion. What produces the emotion is not the event itself but what the event means for the person: whether it matters, whether it helps or harms their goals, who or what caused it, and whether they can cope with it. The basic premise, shared across the family, is that emotions are adaptive responses reflecting appraisals of features of the environment that are significant for the person’s well-being (Moors, Ellsworth, Scherer, & Frijda, 2013). See appraisal in the glossary.

According to the same overview, the approach was pioneered in the 1960s by Magda Arnold and Richard Lazarus and developed from the 1980s by several theorists who proposed specific variants and tested predictions. The overview also notes that “appraisal” is used in widely different ways, which is why it is more accurate to speak of appraisal theories than of appraisal theory.

Lazarus: stress, coping and relational meaning

Richard Lazarus came to emotion through the study of stress. In the transactional model developed with Susan Folkman, stress is not a property of events or of people but of the relation between them (Lazarus & Folkman, 1984). Two kinds of evaluation are distinguished. Primary appraisal concerns whether something is at stake: is this relevant, and is it a harm, a threat or a challenge? Secondary appraisal concerns what can be done: what resources and options for coping are available? Coping itself is divided into problem-focused efforts, which act on the situation, and emotion-focused efforts, which act on the emotional response. That distinction anticipates the difference between situation modification and response modulation in the process model (Theory 02).

Lazarus later extended this into a cognitive-motivational-relational theory of emotion (Lazarus, 1991). The article sets out what a theory of emotion must address, including whether physiology should define emotion, whether categories or dimensions are the better strategy, and how biological universals are reconciled with sociocultural variation. It then applies an analysis of appraisal patterns, and of the core relational themes they produce, to anger, anxiety, sadness and pride. A core relational theme is a summary of the relational meaning that characterises an emotion: the kind of person–environment relationship that, if appraised, yields that emotion rather than another. The theory also gives coping a place inside the emotion process, not only after it.

The reasoning is functional. If emotions are responses to what matters, then emotions should be organised by kinds of significance, and different emotions should correspond to different ways in which things can matter. Anger, anxiety and sadness are then not arbitrary categories but responses to recognisably different predicaments.

Scherer: the component process model

Klaus Scherer’s component process model develops the appraisal idea into a more detailed architecture (Scherer, 2009). Emotion is conceptualised as an emergent, dynamic process based on the individual’s subjective appraisal of significant events. Two features distinguish it.

A sequence of appraisal checks. Rather than one global evaluation, the model proposes a series of checks that unfold rapidly and can recur. They are commonly grouped by the questions they answer: Is this event relevant to me? What are its implications for my goals? Can I cope with it or adjust to it? How does it fit social norms and my self-image? The outcome of each check is thought to modify the response as it unfolds.

Synchronised components. Appraisal outcomes are proposed to drive changes in several components: motivational (action tendencies such as approach or withdrawal), physiological, and expressive (face, voice, gesture). Emotion, on this view, is an episode in which these components become synchronised in response to appraisal. The subjective feeling is a further component that integrates and represents these changes, and categorising and labelling the feeling with emotion words comes later still (Scherer & Moors, 2019).

The model makes the relation between appraisal and emotion explicit and, in principle, testable. If the sequence exists, manipulating information relevant to an early check should affect responses earlier than information relevant to a later check. If components synchronise, their coupling should be greater during emotional episodes than outside them. The component process model is thus the most mechanistic version of appraisal theory.

What kind of evidence supports appraisal theories

Evidence for appraisal theories comes in several forms, which differ in what they can show.

Self-report of appraisals and emotions. People asked to recall or imagine emotional events and to rate them on appraisal dimensions show systematic patterns: emotions reported as anger are rated differently on dimensions such as agency and goal obstruction than emotions reported as sadness. These studies establish that appraisal profiles and emotion labels covary. They cannot establish that appraisals came first, because both are reported afterwards and people may reconstruct appraisals that fit the emotion they remember.

Manipulating appraisal-relevant information. Experiments that vary, for example, whether an outcome is caused by another person or by circumstance, or whether it can be controlled, and then measure emotional responses provide stronger causal evidence. The review by Scherer and Moors (2019) concentrates on experimental studies of this kind, including studies of how appraisal outcomes relate to physiology and expression. These findings support the claim that manipulated meaning changes emotional responses.

Timing and component coupling. Studies measuring expression and physiology at fine time scales test whether components change in the order the component process model predicts. Scherer’s 2009 article reviews such evidence in support of the model. Much of this work comes from the model’s proponents, and it is a narrative rather than systematic synthesis.

Overall, the evidence is consistent with the view that evaluations of meaning shape emotional responses. Whether it establishes the specific architecture of any one appraisal theory is less clear, and the site grades those claims as theory.

Appraisal as the target of reappraisal

Appraisal theories provide the mechanism that the most studied regulation strategy is meant to act on. Reappraisal means changing the appraisal: reinterpreting an event, or one’s goals in relation to it, so that the emotion that follows is different. In the process model it is the core of cognitive change (Gross, 2015).

The logic is direct. If anger follows from appraising an event as a blameworthy offence, then reappraising it (the person did not intend it, the loss is smaller than it seemed, the goal it blocked was not that important) should reduce anger. The emotion changes because what produces it has changed.

The evidence that instructed cognitive change alters emotional outcomes is strong at the level of effect: a meta-analysis of experiments found a small-to-medium effect (Webb, Miles, & Sheeran, 2012). Whether the effect runs specifically through changed appraisals, rather than, for example, through distraction by the effort of generating reinterpretations, is less directly tested, because most studies do not measure the appraisals themselves before and after. The site therefore records the mechanism as a theoretical claim with moderate confidence, while the effect itself rests on meta-analytic evidence.

Recent work also distinguishes kinds of reappraisal that map onto appraisal dimensions differently. Some tactics reconstrue the situation (what happened), others repurpose the goals against which it is appraised (what matters), and some act on the appraisal of the emotion itself (whether feeling this way is a problem) rather than of the situation (Uusberg et al., 2023). This refinement follows naturally from appraisal theory: if emotions depend on several evaluations, there are several places where meaning can change.

Appraisal, relief and resolution

Appraisal theory gives a precise way to state the distinction at the centre of this site. Stated as interpretation:

If the appraisal that generates an emotion is unchanged, the emotion can be expected to return whenever the event is encountered or recalled. Expression, distraction or physiological calming may bring relief, but the generating evaluation remains in place.
If the appraisal changes, because the situation changes, because new information arrives, or because the person reinterprets it, the emotion that the event produces changes too. That is one sense of resolution.

This also explains why rehearsing a grievance while venting can maintain anger (Theory 01). Repeatedly describing an event as an intentional, unjust offence re-enacts the appraisal that produces anger. It does not only express the emotion; it renews its cause.

Appraisal theory also warns against a naive reading of resolution. Reappraisal changes meaning, not circumstances. When a situation is harmful and could be changed, an appraisal that makes it seem acceptable may reduce distress while leaving the harm in place. Lazarus’s distinction between problem-focused and emotion-focused coping makes the point: both can be adaptive, and which is appropriate depends on what can be controlled. Whether a changed appraisal is accurate is a separate question from whether it reduces distress.

Where appraisal theory is contested

Does appraisal cause emotion? The strongest challenge is causal. Critics argue that appraisal dimensions may describe how people make sense of emotional episodes after the fact, rather than cognitive steps that produce them. In Russell’s framework, much of what happens in an emotional episode can be explained as core affect, a basic pleasant–unpleasant and activated–deactivated state, attributed to a cause, with appraisal-like processes as one ingredient among several rather than the trigger (Russell, 2003). The theory of constructed emotion goes further, proposing that emotions are constructed from predictions and concepts (Theory 04; Barrett, 2016). Appraisal theorists reply that appraisals need not be conscious, deliberate or verbal, and that experiments manipulating meaning show causal effects. The site records the question as open.

Circularity. If appraisal patterns are defined by the emotions they are supposed to explain, the theory risks being true by definition. This is sharpest in self-report studies, where participants may generate appraisals from their knowledge of emotion words. Experimental manipulation of appraisal-relevant information addresses the problem partially.

Speed and automaticity. Emotions can arise very quickly, sometimes before a person can say what they are reacting to. Appraisal theorists respond that appraisal checks can run at several levels, including fast, automatic ones. Critics argue that extending appraisal to automatic processing makes the concept harder to distinguish from perception or learning.

Differences among appraisal theories. Theories disagree on the number of appraisal dimensions, their order and whether they operate in fixed sequence. The editorial overview of the field acknowledges unresolved issues and critical objections (Moors et al., 2013). Evidence for the general premise is not evidence for any one variant.

What this essay concludes

Appraisal theories make emotions intelligible: they explain why an emotion fits a situation as a person understands it, and why the same situation can produce different emotions. Their central claim, that evaluations of meaning shape emotion, is well supported as a general proposition by experimental manipulations of meaning; their more detailed architectures, including the sequence of checks in the component process model, are supported mainly by their proponents and remain debated. For the question of catharsis, appraisal is the natural place to look for the difference between relief and resolution: expression that leaves the appraisal intact can calm without changing anything, and a changed appraisal can change the emotion itself.

Basic emotions and constructed emotions: the current debate

2026-09-14T00:00:00+00:00

What the debate is about

Ask most people what fear is and they will describe a recognisable thing: a racing heart, wide eyes, an urge to flee. The question behind this essay is whether that description names a natural kind, a biologically given pattern that occurs whenever fear occurs, or a category that people use to make sense of highly variable states. The answer shapes what emotion science looks for, how it interprets its data and, for this site, what it could mean to measure an emotion at all.

Two positions define the debate. They are not the only ones; appraisal theories (Theory 03) and dimensional accounts occupy intermediate ground. But the contrast between basic emotion theory and the theory of constructed emotion is the sharpest version of the question, and it has driven much recent research.

Basic emotion theory

The argument

In its classic statement, basic emotion theory argues that emotions evolved because of their adaptive value in dealing with fundamental life tasks, such as facing danger, losing something valued or meeting an obstacle (Ekman, 1992). Each emotion has unique features: a signal, a physiology and typical antecedent events. Emotions also share characteristics that distinguish them from other affective phenomena: rapid onset, short duration, unbidden occurrence, automatic appraisal and coherence among responses. These shared and unique features are proposed to be products of evolution.

The reasoning is functional and comparative. If certain predicaments recurred throughout evolutionary history, organisms that responded to them with coordinated, fast, pre-organised packages of physiology, action and signal would have had an advantage. Signals, especially facial and vocal expressions, would allow others to anticipate a person’s behaviour. On this view, emotion categories are not inventions of language but discoveries of biology.

How the theory has changed

Later statements refined the claim. Basic emotions are described not as single states but as families of related states sharing a set of characteristics that distinguish one family from another; these responses are preprogrammed and involuntary but also shaped by life experience (Ekman & Cordaro, 2011). Culture and learning influence when emotions are elicited and how they are displayed, without, on this view, removing the underlying family structure.

More recent work by proponents has widened the list considerably. A 2019 review reports studies in which upwards of 20 emotions are signalled in multimodal and dynamic patterns of expressive behaviour, moves beyond simple matching of words to faces, and considers contextual influences upon emotion recognition (Keltner, Sauter, Tracy, & Cowen, 2019). A large self-report study found that experiences elicited by 2,185 short videos were better captured by 27 categories than by dimensions such as valence and arousal, but that the boundaries between categories were fuzzy and linked by smooth gradients (Cowen & Keltner, 2017). That result is notable because it supports categories while contradicting the idea that they are discrete.

Readers should therefore be careful with criticism aimed at “six universal facial expressions”. Current basic emotion theory is broader and more context-sensitive than the version often attacked. The site records the position as stated by its proponents (basic emotion theory).

The theory of constructed emotion

Core affect

Constructionist accounts begin from a different observation: emotional states vary enormously within categories. Anger can involve shouting or cold silence, a racing heart or a slow one. James Russell proposed that at the heart of emotion, mood and any other emotionally charged event are states experienced as simply feeling good or bad, energised or enervated, which he called core affect (Russell, 2003). People have no direct access to what causes changes in core affect. Core affect can be experienced as free-floating, as in mood, or attributed to some cause, and thereby begin an emotional episode. Prototypical episodes such as fear and anger are, in this framework, core affect attributed to something, plus various non-emotional processes. See affect.

Construction through prediction and concepts

Lisa Feldman Barrett’s theory of constructed emotion develops this into a broader account (Barrett, 2016). The article argues that emotion science has searched for the biological basis of folk-psychological categories inherited from philosophy, and proposes instead to begin with how the nervous system is organised and to deduce what emotions might be. The resulting account is computational and predictive. Bodily regulation is proposed to rest on predictions about the body’s needs; interoceptive signals are interpreted in light of past experience; and emotion concepts are used to categorise those signals in context. An instance of fear is, on this view, constructed when sensations are categorised as fear in a situation, not triggered as a fixed programme (theory of constructed emotion).

Two predictions follow. First, emotion categories should show variation rather than fixed signatures: different instances of anger can involve different physiology and behaviour. Second, the ingredients of emotion should be domain-general: the same systems for regulating the body, predicting and conceptualising should be involved in emotional and non-emotional states.

What the evidence shows

Neither theory can be settled by a single study, and both sides cite substantial bodies of work. The following findings have been central.

Neuroimaging

A meta-analytic review of neuroimaging studies compared a locationist hypothesis, that discrete emotion categories consistently and specifically correspond to distinct brain regions, with a psychological constructionist hypothesis, that emotion categories are constructed from more general networks not specific to those categories (Lindquist, Wager, Kober, Bliss-Moreau, & Barrett, 2012). The authors found little evidence that discrete emotion categories can be consistently and specifically localised to distinct brain regions. Instead, a set of interacting regions commonly involved in basic psychological operations of both an emotional and non-emotional nature were active across a range of emotion categories.

Two cautions apply. The pooled studies are correlational: activation during a task is associated with the task, not shown to produce the emotion. And a failure to localise categories to regions does not rule out category-specific patterns distributed across networks, which multivariate methods look for. The article was published with peer commentaries, some of them critical. See also the site’s discussion of reverse inference.

Autonomic physiology

A meta-analysis of 202 studies found that autonomic nervous system responses did not clearly distinguish one emotion category from another and varied substantially within categories (Siegel et al., 2018). An earlier review of 134 publications reported considerable autonomic specificity when emotion subtypes are considered (Kreibig, 2010). The two are not simply opposed: the earlier review grouped responses more finely, while the later meta-analysis tested consistency across studies. The site cites Kreibig only as a partial counterpoint.

Faces

A major review concluded that facial movements are not reliable, specific indicators of emotional states across people, situations and cultures: typical configurations co-occur with emotions more often than chance, but weakly and inconsistently (Barrett, Adolphs, Marsella, Martinez, & Pollak, 2019). Proponents of basic emotion theory respond that expression is multimodal and dynamic and that recognition is context-sensitive, which they take to be compatible with emotion families (Keltner et al., 2019). The two sides partly agree on the data, weak and variable coherence, and disagree on what it implies. The consequences for measurement are discussed in the Research note on measuring emotion.

Self-report

Self-report studies, including the 27-category study above, show that people distinguish many categories of experience reliably. Basic emotion theorists take this as evidence of rich category structure. Constructionists take the same evidence as showing how people use emotion concepts, which is what their theory predicts, and note that gradients between categories fit variation better than discrete kinds.

Reading the debate carefully

The debate is often presented as a choice between “emotions are universal and biological” and “emotions are cultural and invented”. Both sides reject that caricature. Basic emotion theory allows for learning, culture and context. The theory of constructed emotion is explicitly biological; it disagrees about how biology is organised, not whether it matters.

The real points of disagreement are narrower:

Are emotion categories natural kinds? Does fear have a consistent biological core that defines it, or is it a population of varied instances grouped by a concept?
Where does specificity lie? In dedicated programmes for each emotion, or in the combination of general-purpose processes in context?
What counts as disconfirming? Proponents of each side argue that the other’s theory can absorb unfavourable data. Critics of the constructionist view argue that a broad predictive framework is hard to falsify; critics of basic emotion theory argue that each failure to find signatures is explained by adding moderators.

There is also a methodological lesson. Both sides now agree that emotional responses vary more than early theories assumed, and that context has to be part of any account. The dispute is about how to explain that variation.

Implications for “measuring” emotion

For this site, the most practical consequence concerns what can be inferred from a signal. Here the two theories converge more than their rhetoric suggests.

Under basic emotion theory, emotion families are real, but their signals are probabilistic, multimodal and moderated by context. A single facial movement, heart rate change or skin conductance response is not a reliable indicator of which emotion is present. Under the theory of constructed emotion, there is no fixed signature to find, and any inference about an emotion from signals is a categorisation that depends on context and concepts. Either way, the defensible statement is the same: a recorded signal is consistent with several states, and emotion categories can be inferred only with uncertainty, from patterns in context.

This is why the site distinguishes signal, feature, inference and context, and why it treats any claim that a device or model reveals what someone feels as overreach. It is also why the artwork’s telemetry is labelled ARTISTIC: its numbers are not measurements of an emotion, under either theory. See the methods page.

The debate also bears on regulation research. If emotion generation and regulation are distinct processes, regulation strategies act on an emotion that already exists. If emotions are constructed, regulation may be part of the construction itself. Gross and Barrett argued that whether the two can be usefully distinguished depends on one’s theoretical perspective (Gross & Feldman Barrett, 2011), and the process model can be read in either way (Theory 02).

Implications for catharsis

The hydraulic idea of catharsis fits most naturally with a view of emotion as a discrete, bounded quantity that exists before it is expressed. Neither current theory supports that picture straightforwardly. Basic emotion theory describes rapid, short-lived responses rather than stored quantities. Constructed emotion theory describes ongoing construction, in which expressing and describing an emotion are part of how the emotion is made.

As interpretation, the constructionist view offers one reason why expression without reframing might keep an emotion going: repeatedly categorising one’s state as anger, in a context that affirms the grievance, reconstructs anger. The basic emotion view offers a different reason for the same observation: re-exposure to the eliciting appraisal re-triggers the response. Both are compatible with the empirical finding that rehearsal while venting tends to maintain anger (Theory 01). Neither requires a reservoir that empties.

What this essay concludes

The debate between basic and constructed emotion is live, and the site does not take a side on what emotions ultimately are. Current evidence supports three narrower statements. Proponents of basic emotion theory now describe a larger, more context-sensitive set of emotion families than early versions. Meta-analyses of neuroimaging and autonomic physiology provide little support for consistent, specific signatures of emotion categories, with a recorded counterpoint for autonomic subtypes. And under either theory, emotions are inferred from patterns in context with error, never read off a single signal. That last statement is the one this site depends on.

Theories of catharsis: from katharsis to an empirical hypothesis

2026-09-14T00:00:00+00:00

One word, three theories

Catharsis is used today as if it named a single, well-understood process: an emotion builds up, it is expressed, and the person is left lighter. The word has in fact carried at least three distinct theories, developed for different purposes and answerable to different kinds of evidence. The first is aesthetic and belongs to the interpretation of Aristotle. The second is clinical and belongs to the early history of psychotherapy. The third is empirical: a hypothesis about what happens to an emotion when it is expressed, which can be tested and has been.

Keeping these apart matters because each makes a different claim. An aesthetic theory of tragedy says something about what an audience gains from watching suffering represented on stage. A clinical discharge model says something about the cause of symptoms and how to remove it. An empirical hypothesis predicts that a measurable quantity, such as anger or distress, will be lower after expression than without it. Evidence against the third does not refute the first, and the authority of the first cannot rescue the third. This essay takes them in historical order and then asks what they imply for the distinction this site is built around: relief is not the same as resolution (see the working definition and Research note on relief).

Katharsis in Aristotle

In Greek usage katharsis belonged to two domains. In medicine it named the purging of harmful matter from the body; in religion it named ritual purification. Aristotle brought the word into the theory of art in one sentence of the Poetics (1449b), where tragedy is said to accomplish, through pity and fear, the katharsis of such emotions. He does not explain the term there, and no fuller discussion survives in the text as we have it.

Three families of interpretation have grown from that sentence.

Purgation. On the medical reading, tragedy arouses pity and fear in order to expel them, much as a physician induces a purge to remove an excess. The audience leaves with less of these emotions than it brought. This reading is the ancestor of the popular idea of catharsis, and it already contains the hydraulic assumption that an emotion is a quantity which can be drawn off.

Purification. On the ethical reading, tragedy does not remove pity and fear but refines them. Spectators learn to feel them towards the right objects, to the right degree. The emotion is not emptied but educated. Here katharsis is closer to training than to discharge.

Clarification. On the intellectual reading, the katharsis belongs to the events of the drama or to the spectator’s understanding of them. By following an action that is coherent and probable, the audience comes to understand why pitiable and fearful things happen. Katharsis is then a kind of insight reached through emotion, not a release of it.

None of these readings has won. The textual evidence is thin, each can cite other passages in Aristotle for support, and the choice among them depends partly on how one reconstructs his wider theory of emotion and of art. The site records this as an unresolved interpretive question (see Research note 01). The dispute is worth knowing for a psychological reason, however. The three readings map onto three possible answers to the empirical question that came much later: does expression remove an emotion, reshape it, or help a person understand something? Only the first is a discharge theory.

The clinical theory: discharge and abreaction

The second theory of catharsis appears in Josef Breuer and Sigmund Freud’s Studies on Hysteria (1895). Their cases describe patients, most famously Breuer’s patient known as Anna O., whose symptoms were said to ease after they recalled a distressing memory and expressed the emotion attached to it. Breuer and Freud called the expression abreaction and the procedure the cathartic method.

The reasoning behind the method is worth stating carefully, because it is the source of the modern hydraulic picture. The model assumed that an experience normally produces an affect that is worked off through action, speech or association with other ideas. When that working-off is blocked, the affect remains attached to the memory, and the undischarged affect maintains the symptom. Recalling the memory vividly and expressing the emotion fully discharges the affect, and with it the symptom disappears. The model therefore treats affect as something like a conserved quantity: it does not dissipate on its own, it has to go somewhere.

Two features of this theory matter for what followed. First, it is a causal theory of symptoms. It claims not only that people feel better after expressing emotion, but that expression removes the cause of their difficulty. Second, the evidence offered for it consisted of uncontrolled case reports. There was no comparison with patients who recalled memories without strong expression, or who received attention and time without recall. Such reports can show that improvement followed treatment; they cannot show that abreaction produced it. The site records the discharge model as a historical theory, not as an established mechanism.

Freud himself moved away from hypnosis and the cathartic method towards free association, the interpretation of resistance and the slower process he later called working through. One reason given in historical accounts is that relief after abreaction was often temporary: symptoms returned when the conflict behind them had not been changed. The shift can be read as an early, informal version of the distinction between relief and resolution. Discharge could make a patient feel better for a while; something other than discharge seemed necessary for lasting change.

The hydraulic picture, however, outlived the method. It is intuitive, it fits common metaphors (letting off steam, bottling things up), and it offers a simple rule: expression is healthy, holding back is harmful. The popular term for this rule is venting, and the underlying picture is described in the glossary as the hydraulic model.

The empirical theory: a testable hypothesis

Twentieth-century psychology turned the hydraulic picture into a prediction: if a person expresses an emotion, the emotion, and the behaviour it motivates, should subsequently be reduced. This is the catharsis hypothesis in its empirical form. Its virtue is that it can be wrong.

Operationalising catharsis

Testing the hypothesis requires deciding what counts as expression and what outcome should fall. Much of the research concerned anger, because anger has a behavioural outcome, aggression, that can be measured in the laboratory. The typical design angers participants, gives some of them an opportunity to express the anger physically or verbally, and compares their later anger and aggression with those of participants who did something else or nothing.

What the anger studies found

In one of the best-known experiments, angered participants who hit a punching bag while thinking about the person who had provoked them later reported more anger and behaved more aggressively than participants who hit the bag while thinking about physical fitness or who sat quietly (Bushman, 2002). The result is the opposite of what the discharge model predicts. It is consistent with the view that expression accompanied by attention to the provocation rehearses the anger rather than using it up. The details are in Research note 02.

A 2024 meta-analysis of 154 studies broadened the question to activities that increase or decrease physiological arousal (Kjærvik & Bushman, 2024). Arousal-decreasing activities, such as slow breathing or relaxation, reduced anger and aggression. Arousal-increasing activities, including some forms of exercise, were not effective overall, although those studies were heterogeneous. For the empirical hypothesis, the conclusion is direct: in the domain where it has been tested most thoroughly, expression through arousal does not reliably reduce the emotion, and in some conditions it is associated with more of it (claim).

A meta-analysis of emotion regulation experiments points in the same direction from a different angle. Concentrating on one’s feelings was not an effective way to change emotional outcomes, whereas cognitive change had a small-to-medium effect and distraction was effective (Webb, Miles, & Sheeran, 2012). If venting is understood as sustained attention to the feeling and its cause, the regulation literature gives no reason to expect it to reduce the emotion. How venting fits into that framework is discussed in Theory 02.

Where expression did help

The empirical story is not simply that expression is useless. Two lines of research found benefits, and both are informative about why.

In studies of expressive writing, people who wrote about both the facts of a distressing experience and their feelings about it later showed health benefits compared with control groups. Analyses of the language they used suggested that benefit was associated with building a more coherent account over the sessions, with more causal and insight-related words, rather than with the amount of emotional expression itself (Pennebaker, 1997).

In research on the social sharing of emotion, people share most of their emotional experiences with others, and they generally find sharing helpful. Yet a review of the evidence concluded that findings were generally not consistent with the view that sharing alleviates the emotional memory (Rimé et al., 1998). A later experiment separated two things a listener can do. When the listener responded empathically, participants felt more connected and had an impression of feeling better; emotional recovery measured two days later occurred only when the listener stimulated cognitive work, such as reframing (Nils & Rimé, 2012). Theory 05 develops this in detail.

Taken together, these findings suggest that when expression is followed by measurable benefit, the benefit is tied to what the person does cognitively with the experience, not to discharge as such (claim). That is a theoretical interpretation of a pattern across studies with different designs and outcomes, and it is graded accordingly in the ledger.

Reasoning about relief and resolution

The discharge model makes a strong prediction that the evidence allows us to split into two parts. The first part, that people often feel relief after expressing an emotion, is compatible with a great deal of everyday experience and with the finding that people rate sharing as helpful. The second part, that the relief reflects the removal of the emotion’s cause, is not supported for anger and is at best conditional for distress more broadly.

Several mechanisms, none of them hydraulic, can produce the first without the second.

Arousal passes. Emotional arousal declines with time. Whatever a person is doing when it declines may be credited for the decline.
Attention and support. Being listened to can reduce distress in the moment and strengthen a relationship without changing how the event is appraised.
Negative reinforcement. A behaviour that ends an unpleasant state is reinforced whether or not it addresses the cause, so a person learns to repeat it (Theory 06).
Memory for peaks and ends. How an episode ends colours how it is remembered, so an episode that ends with release can be remembered as helpful even if its overall course was not improved.

Resolution, by contrast, requires that something changes: the situation, the meaning the person assigns to it, the person’s ability to cope with it, or what they learn from it. Expression can contribute to any of these, particularly when it is organised into an account or met with help to reconsider. Expression that only rehearses the grievance may instead keep the appraisal that generates the emotion active. This is the sense in which the artifact’s status line, problem_solved: false, is a claim about the gap between the two outcomes rather than a judgement that relief is worthless.

Where the theory remains contested

It would overstate the evidence to say that catharsis has simply been refuted. Several questions remain open.

The meaning of Aristotle’s term. Psychology cannot settle a philological dispute. A clarification or purification reading of katharsis is not tested by punching-bag experiments and is, if anything, closer to what the processing evidence describes.

Which emotion, which expression. Most controlled tests concern anger and aggression in laboratory settings with student samples. Grief, fear and shame differ in their action tendencies and social functions, and conclusions about anger do not transfer automatically.

Emotional arousal in therapy. Several psychotherapies deliberately evoke emotion during treatment, for example exposure-based therapies for anxiety, in which emotional activation is combined with corrective information (Research note on clinical perspectives). These approaches do not rely on discharge; they treat activation as a condition for new learning. Whether and when emotional intensity in itself contributes to outcome is an open empirical question.

Aesthetic experience. People seek out sad music, tragedy and ritual lament, and report that these experiences are valuable. Whether that value involves any reduction of emotion, or instead meaning, connection or reflection, is not answered by laboratory studies of anger (Theory 08).

Measurement. Outcomes such as “reduced emotion” depend on how emotion is measured. Self-reported relief, physiological recovery, later behaviour and residual intensity days later are different outcomes, and a study can find benefit on one and none on another (Research note on measuring emotion).

What this essay concludes

The history of catharsis is a history of a metaphor becoming a hypothesis. The discharge model gave the metaphor a clinical form, and the empirical tradition gave it a form that could fail. Where it has been tested most directly, it has not held: expression that rehearses anger is associated with more anger, not less. Where expression has helped, the evidence suggests that the help came from processing, connection and changed understanding, which are closer to the purification and clarification readings of Aristotle than to purgation.

The practical reading of this, stated as interpretation rather than finding, is modest. Feeling relieved after expressing an emotion is common and not in itself a problem. It is simply not evidence that anything has been resolved. For that, it is worth asking what, if anything, changed. The site’s process model essay and appraisal essay describe where such change can occur.

Collective emotion and ritual

2026-09-14T00:00:00+00:00

Emotion that happens between people

Most of the theories in this section treat emotion as something that happens inside an individual and is then expressed, shared or regulated. Many of the moments people describe as cathartic, however, are collective. A stadium sings together after a loss; a crowd at a vigil falls silent; strangers at a concert cry at the same chord; mourners at a funeral move from grief to shared laughter. People often say these experiences change them, and that the change could not have happened alone.

This essay asks what theories of collective emotion and ritual propose about such moments, what research has measured, and how the answers bear on the site’s distinction between relief and resolution. The empirical side of synchrony is covered in the Research note on collective synchrony; this essay concentrates on the theories behind it.

Durkheim: ritual and the renewal of the group

The sociological theory of collective emotion begins with Émile Durkheim’s study of religious life, published in 1912 (Durkheim, 1912). Working from ethnographic reports, Durkheim argued that religious ritual is fundamentally social. When a group gathers and performs shared actions around shared symbols, participants experience an intensified emotion that exceeds what any of them would feel alone. They experience this force as coming from something greater than themselves. Durkheim’s interpretation was that what they encounter is, in effect, the group: ritual makes the collective present and felt. Participants leave with renewed attachment to the group and its moral order.

Two features of this theory matter here. First, it is explicitly about emotion produced between people, by co-presence and shared action, not about individual emotions that are released in company. Second, its function is integrative: ritual is said to renew solidarity, not to resolve individual problems.

Durkheim’s theory has been criticised for its functionalism, its reliance on dated and secondhand ethnography, and its tendency to present social cohesion as the purpose of ritual. It makes no measured claim. The site cites it for its historical influence and for predictions that later research has tested, which Páez and colleagues summarise as the prediction that collective gatherings, through the social sharing of emotion they generate, bring participants to a state of collective effervescence: a sense of union with others and a feeling of empowerment accompanied by positive affect, with renewed confidence in life and social institutions (Páez, Rimé, Basabe, Wlodarczyk, & Zumeta, 2015).

Collins: interaction ritual chains

Randall Collins extended Durkheim’s account from dramatic religious ceremonies to everyday encounters (Collins, 2004). In his theory, which he describes as a radical microsociology, rituals are not only formal events but any interaction with the right ingredients. In the standard summary of the model, these ingredients are bodily co-presence, a boundary between participants and outsiders, a mutual focus of attention, and a shared mood. When they come together and intensify through feedback, the ritual succeeds.

Successful rituals, Collins proposes, create symbols of group membership and charge individuals with emotional energy: a sense of confidence, enthusiasm and initiative. Failed rituals drain emotional energy. People move from situation to situation over the course of their lives, drawn towards interactions where their cultural resources give them the best emotional energy payoff. The theory thus links micro-level emotional experiences to larger patterns of social life; Collins applies it to phenomena as different as conversation, smoking and social stratification (interaction ritual chain).

For catharsis, Collins’s theory offers an alternative to the hydraulic model. The emotional high after a successful gathering is not the release of stored feeling; it is emotional energy produced by the interaction. And because people seek interactions that produce it, the theory predicts that they will return to rituals that work for them, whether or not those rituals change anything else. The book is a theoretical framework built on microsociological observation, not a quantitative test.

Bernard Rimé’s work on social sharing (Theory 05) provides a psychological bridge between individual and collective emotion (Rimé, 2007). Sharing reactivates emotion in speaker and listener and favours empathy and emotional communion. Listeners share further with others. When an emotional event strikes a community, there are as many sources of sharing as members, and each act of sharing reloads the flow. Rimé proposed that such chains of sharing contribute to an emotional climate, and predicted effects on group cohesion and solidarity and on collective memory. These collective predictions are theoretical; they are consistent with Durkheim’s account while locating its mechanism in ordinary communication.

Collective effervescence: testing Durkheim

The gatherings studies

A century after Durkheim, Páez and colleagues set out to test his predictions as a whole (Páez et al., 2015). The article reports four studies, two correlational, one semi-longitudinal and one experimental, of participation in positively valenced gatherings (folkloric marches) and negatively valenced ones (protest demonstrations). Across studies, participation was associated with stronger collective identity, identity fusion and social integration, with higher personal and collective self-esteem and efficacy, with more positive affect, and with more positive social beliefs. Perceived emotional synchrony with others, called emotional communion, mediated these effects. Higher perceived emotional synchrony was associated with stronger emotional reactions, stronger social support and higher endorsement of social beliefs and values. The authors also distinguished perceived emotional synchrony from flow, which was related mainly to efficacy and self-esteem.

The site grades this evidence at level C. Most of the outcomes come from correlational and semi-longitudinal designs, and the abstract does not describe what the experimental study manipulated. People who choose to attend a march may differ from those who do not, and perceived synchrony is itself a self-report measured alongside the outcomes (collective effervescence).

The meta-analysis

A later meta-analytic review synthesised outcomes of collective effervescence during gatherings and demonstrations across 50 samples (Pizarro et al., 2022). Collective effervescence was associated with self-transcendent emotions, group identity, social integration, wellbeing and collective efficacy. Because the pooled studies are mainly correlational, the associations are not evidence that gatherings cause these outcomes (claim).

Synchrony and the body

Laboratory synchrony

Laboratory experiments isolate one ingredient of ritual: moving together. Moving or tapping in synchrony with others increased liking for a partner and later cooperation, including cooperation that required personal sacrifice (Wiltermuth & Heath, 2009). In a group-dance experiment, synchrony and exertion each raised pain thresholds, an indirect proxy for endorphin activity, and each increased in-group bonding (Tarr, Launay, Cohen, & Dunbar, 2015). These are controlled studies with short tasks, and the outcomes were measured during or shortly after the activity. See interpersonal synchrony.

A ritual in the field

A field study at a Spanish fire-walking ritual measured heart-rate dynamics in 38 participants (Konvalinka et al., 2011). Using recurrence analyses, the researchers identified fine-grained commonalities of arousal over the 30-minute ritual between fire-walkers and spectators who were related to them, but not between fire-walkers and unrelated spectators. Because the fire-walkers and their related observers behaved very differently, the authors suggested that the mediating mechanism may be informational rather than motor. The study shows that a collective ritual can be accompanied by synchronised arousal between performers and bystanders. It measured shared physiological arousal in one ritual and one community; it did not measure shared feeling, wellbeing or later outcomes.

Ritual as regulation

Psychologists have more recently offered a framework that connects these lines of work (Hobson, Schroeder, Risen, Xygalatas, & Inzlicht, 2018). The integrative review proposes that rituals serve three regulatory functions: regulation of emotions, of performance goal states, and of social connection. It distinguishes bottom-up processes, arising from the physical features of rituals such as repetition, rigidity and synchrony, from top-down processes, arising from the meaning rituals have for those who perform them. The framework is designed to generate predictions; it does not test them.

Seen through this framework, the rituals that people describe as cathartic are regulatory in several ways at once. They provide structure when a situation is uncertain; they give intense emotion a sanctioned form and a time limit; they direct attention to shared symbols; and they bring people into physical and emotional alignment. The regulation is real. What it regulates, and for how long, are empirical questions.

Ritual and resolution

How do these theories bear on the distinction between relief and resolution? The following is interpretation, informed by the evidence above.

Shared intensity is produced, not released. Durkheim and Collins both describe emotion generated by gathering, attention and synchrony. The high after a concert or a march fits this description better than the hydraulic picture of stored emotion being let out (Theory 01).

Rituals can mark and support transitions. Funerals, weddings and commemorations do not make grief, change or loss disappear. They give it a shared form, publicly acknowledge what has changed, and mobilise the support of others. In that sense ritual can contribute to resolution, not by ending an emotion but by helping people move into a changed situation with others. This is a theoretical reading; the studies reviewed here did not measure long-term adjustment after such rituals.

Relief can be the reward that brings people back. Collins’s theory predicts that people return to rituals that charge them with emotional energy. The learning principle in Theory 06 makes a similar prediction for relief. A ritual that reliably produces emotional energy or release may be repeated for that reason alone, regardless of whether the difficulties participants bring to it change.

Solidarity is not always benign. The same processes that create belonging can sharpen boundaries between insiders and outsiders. Collins’s model includes a boundary to outsiders as an ingredient, and the gatherings studied include protests, whose collective efficacy is directed against something. The studies cited here measured positive outcomes for participants; they did not measure effects on intergroup relations, and the site has insufficient evidence to say how often collective emotion intensifies hostility.

Evidence is short-term. Most measured outcomes were recorded during an event or shortly after it. There is insufficient evidence that collective rhythm or gatherings reduce the difficulties people bring to them over longer periods (claim).

The artwork at the centre of this site is built around a collective experience of release and ends with problem_solved: false. The theories in this essay explain why both parts can be true at once: a shared experience can generate real emotional energy, connection and relief, and still leave unresolved what brought each person into the crowd.

Where the theories are contested

Functionalism. Durkheim’s account, and to some extent its successors, can read as if rituals exist in order to create cohesion. Critics note that rituals also enforce hierarchy, exclude and coerce, and that the function of a ritual is not the same as its cause.

Measurement of collective emotion. “Perceived emotional synchrony” is a participant’s judgement. Physiological synchrony, as in the fire-walking study, is a different construct. Whether people who report feeling in tune with a crowd are physiologically synchronised with it is not established by the studies cited here, and synchronised arousal does not show shared feeling (Theory 04).

Selection and causality. People who attend gatherings choose to do so, often because they already identify with the group. Correlational designs cannot separate the effects of participation from the characteristics of participants.

From laboratory synchrony to real ritual. Tapping experiments and group dance in a laboratory capture a small part of what a funeral or festival involves. Their results are consistent with ritual theories but do not test them directly.

Durability. Whether effects of collective emotion persist, and for whom, remains largely unknown.

What this essay concludes

Theories of collective emotion describe something the hydraulic model cannot: emotion generated between people through co-presence, shared attention, synchrony and meaning. Durkheim framed it as the renewal of the group, Collins as emotional energy produced by successful interaction, Rimé as the propagation of shared emotion into a climate, and recent psychology as a set of regulatory functions. Research on gatherings and synchrony is consistent with these accounts, mostly through correlational and short-term designs. The picture that emerges is of ritual as a powerful regulator of emotion and belonging. It is not, on current evidence, a reason to believe that a shared experience of release resolves what participants carry into it.

Learning, avoidance and relief

2026-09-14T00:00:00+00:00

Why a learning theory belongs in an essay on catharsis

Most of the theories in this section describe what an emotion is or how it can be regulated. Learning theory asks a different question: what does a person learn from the consequences of what they do when they feel something? That question matters for catharsis because catharsis is, by definition, followed by a consequence that people value: relief. If relief teaches, then what it teaches matters.

The central idea of this essay can be stated in a sentence. A behaviour that brings relief becomes more likely, whether or not it deals with what caused the distress. From that simple principle follows a counterintuitive possibility: the things that make people feel better in the moment can keep their difficulties going. The research note on relief and resolution introduces this; this essay sets out the theory beneath it.

Reinforcement: a note on terms

In learning theory, reinforcement is any consequence that makes a behaviour more likely to recur. Positive reinforcement adds something (praise, food, pleasure). Negative reinforcement removes something aversive (pain, noise, anxiety). Both increase behaviour. Negative reinforcement is often confused with punishment, which makes behaviour less likely; the two are different.

Relief is negative reinforcement in this technical sense. When anxiety, anger, shame or tension falls after a behaviour, that fall reinforces the behaviour. Nothing in the principle requires the behaviour to have addressed the source of the state. Leaving a crowded room reduces anxiety whether or not the room was dangerous. Drinking reduces tension whether or not the problem producing tension has changed. Shouting at someone may reduce the felt pressure of anger whether or not the conflict is any closer to being settled. The relief is real, and the learning is real; what is learned is “this works”, not “this solved it”.

Avoidance learning

Two-factor theory

The classic account of how avoidance is learned is O. H. Mowrer’s two-factor theory, summarised in a 1951 article (Mowrer, 1951). In its standard form it proposes two stages. First, through Pavlovian conditioning, a cue that has been paired with an aversive event comes to elicit fear. Second, any behaviour that removes the person or animal from the cue, and so reduces fear, is instrumentally reinforced by that reduction. Avoidance is thus learned in two steps: fear is acquired, and escape from fear is rewarded.

The theory explains why avoidance can be so persistent. Once a person reliably avoids a feared situation, they never encounter the situation without the escape. They therefore never learn that the feared outcome would not have happened. Each successful avoidance brings relief, which strengthens avoidance, which prevents disconfirmation. The loop is self-sustaining (avoidance learning).

Later developments

Two-factor theory has well-known problems. Avoidance can persist long after observable signs of fear have faded, which is hard to explain if fear reduction is the only reinforcer. Reviews of avoidance learning describe how later models added other processes, including expectancies about what will happen if one does or does not avoid, and a distinction between goal-directed and habitual responding (Krypotos, Effting, Kindt, & Beckers, 2015). The review distils principles of how adaptive and maladaptive avoidance are acquired and maintained, discusses clinical implications, and describes interventions that could reduce maladaptive avoidance and prevent its return.

For this essay, the important point survives these revisions. Whether avoidance is maintained by fear reduction, by expectancies that avoiding keeps one safe, or by habit, successful avoidance prevents the experience that could change the expectation. The theories differ about the mechanism; they agree that avoidance and relief can maintain fear.

Most of this research is based on laboratory fear-conditioning experiments with animals and healthy humans. Its application to everyday emotional life and to clinical problems is an extrapolation, supported by clinical observation and treatment research, but not identical to it.

Safety behaviours

A related clinical concept is the safety behaviour: a subtle form of avoidance that a person uses within a feared situation, such as gripping a railing, rehearsing sentences, staying near an exit or seeking reassurance. Safety behaviours allow a person to stay in the situation while attributing the absence of catastrophe to the behaviour (“I did not faint because I held on”). The expectation that the situation is dangerous therefore survives.

Experiential avoidance

The concept

Hayes and colleagues extended the logic of avoidance from external situations to internal experience (Hayes, Wilson, Gifford, Follette, & Strosahl, 1996). Experiential avoidance refers to the unwillingness to remain in contact with particular private experiences, such as painful thoughts, emotions, memories and bodily sensations, together with efforts to change their form or frequency. The authors proposed it as a functional dimension that cuts across conventional categories of psychological disorder: substance use, some forms of self-harm, compulsive checking, chronic worry and emotional numbing might all, in part, serve to escape or suppress unwanted inner experience.

The reasoning combines learning theory with an observation about language. Humans can re-experience distress through thought and memory in the absence of any external cue, so avoiding a situation does not remove the distress. People therefore attempt to avoid the experiences themselves. Because such attempts often bring short-term relief, they are negatively reinforced; because thoughts and feelings cannot simply be removed, the attempts must be repeated, and may narrow a person’s life around the effort to feel less.

The evidence and its limits

A review of empirical studies found that some of the literature suggests experiential avoidance may be implicated in various forms of psychopathology, and identified a fundamental limitation: a lack of theoretical integration and refinement in how the construct is operationalised and assessed (Chawla & Ostafin, 2007). The reviewers called for clearer operationalisation and for tests of whether the construct adds to the prediction offered by related constructs.

That limitation matters for interpretation. Most studies are correlational: people who score higher on questionnaire measures of experiential avoidance also report more distress and more symptoms. That association is consistent with avoidance maintaining difficulty, but also with distress leading to avoidance, or with both reflecting a general tendency to experience negative emotion. The site records experiential avoidance as a theoretically important construct with moderate confidence.

Why confronting can help: emotional processing and inhibitory learning

If avoidance maintains fear by preventing new learning, then approaching feared experiences should allow that learning to occur. This is the logic of exposure-based treatments.

Emotional processing theory

Foa and Kozak (1986) proposed that fear is represented as a memory structure linking stimuli, responses and their meanings. For that structure to change, it must be activated, and information incompatible with it must be available. Activation alone is not enough, and neither is information presented without activation. This theory is directly relevant to catharsis. It says that emotional arousal can play a part in change, but only as a condition for correcting information, not as discharge. See emotional processing and the Research note on clinical perspectives.

Inhibitory learning

A later account builds on laboratory research on extinction (Craske, Treanor, Conway, Zbozinek, & Vervliet, 2014). Exposure therapy is effective for anxiety disorders, yet a substantial number of people do not benefit or experience a return of fear afterwards. The inhibitory learning model proposes that extinction does not erase the original fear association; it creates a new, competing association (“this cue is not followed by harm”) that must be retrieved to inhibit fear. Fear can return when the new learning is not retrieved, for example in a different context or after time has passed.

The model leads to specific clinical strategies, which the authors distinguish from a fear-habituation approach: designing exposure so that expectations are violated, deepening extinction, occasional reinforced extinction, removing safety signals, varying stimuli, using retrieval cues, conducting exposure in multiple contexts, and affect labelling. The abstract is explicit that these are examples for clinicians, illustrated with case studies. They are not self-help instructions, and this site does not recommend unsupervised exposure.

The link to relief is direct. On this account, the goal of exposure is not that distress falls during a session, which would be relief, but that new learning is acquired and retrievable later, which is closer to resolution. A session in which anxiety stays high but the feared outcome does not occur may teach more than a session in which anxiety falls because a safety behaviour was used.

Avoidance is not always maladaptive

It would be a misreading of this literature to conclude that all avoidance is harmful. Hofmann and Hay (2018) argue for a balanced approach. Avoidance is typically considered a maladaptive response that maintains anxiety disorders, and exposure treatment discourages escape and safety behaviours. But avoidance behaviours can also function as effective coping strategies that enhance a person’s perceived control over the environment and potential threat, and whether they are helpful depends on their function.

Everyday life supports the point. Leaving a genuinely dangerous situation is avoidance, and it is wise. Taking a break from an escalating argument can prevent harm. Choosing not to watch distressing news before sleep is situation selection (Theory 02). The question the learning perspective invites is not “is this avoidance?” but “what does this behaviour do, and what does it prevent me from learning?”

Relief, catharsis and resolution

Applying learning theory to catharsis requires care. The following statements are interpretation, informed by the theories above.

Cathartic behaviours can be negatively reinforced. If shouting, smashing, venting online or exhausting exercise is followed by a fall in tension, the behaviour is reinforced by that fall. Its recurrence does not show that it addresses anything; it shows that it brings relief.

Venting can function as escape. When the discomfort of anger or shame is high, discharging it quickly can serve to end the discomfort, much as avoidance ends fear. That may prevent the slower work of reconsidering the situation, speaking to the person involved or tolerating the feeling long enough to learn that it passes.

Relief is not a reliable signal of progress. Because reinforcement is indifferent to cause, the feeling “that helped” is weak evidence that the underlying situation, appraisal or expectation has changed. This is the learning-theory form of the site’s thesis (claim).

Not all relief is avoidance. Calming down after a conflict so that one can think clearly is relief in the service of resolution. The difference lies in what happens next. The advice entry Ask what changed turns this into a practical question.

Where the theory is contested

Mechanisms of avoidance. Whether avoidance is maintained mainly by fear reduction, by expectancies or by habit is actively debated, and different mechanisms may dominate at different stages.

The status of experiential avoidance. Critics question whether experiential avoidance is a distinct process or overlaps heavily with neuroticism and general distress. Better measurement and longitudinal and experimental designs are needed.

Habituation versus inhibitory learning. Clinicians and researchers differ on how much within-session reduction in fear matters. The inhibitory learning account shifts attention away from it; the extent to which this improves outcomes is still under study.

From laboratory to life. Much of the theory rests on conditioning experiments that are far simpler than human emotional lives. Its explanatory reach in everyday emotion regulation is plausible but not established.

Clinical boundary

Avoidance that restricts a person’s life, persistent reliance on substances or self-harm to manage feelings, and fear that does not diminish over time are reasons to seek assessment from a qualified professional. Exposure-based therapies are delivered with preparation, consent and support, and they are not the same as deliberately confronting distressing experiences alone. In an emergency, contact local emergency services.

What this essay concludes

Learning theory contributes one principle that is simple, well established and easy to forget: relief reinforces whatever preceded it. From that principle follow testable theories of how avoidance maintains fear, a clinically influential but imperfectly measured concept of experiential avoidance, and an account of why approaching feared experiences, under the right conditions, can produce lasting change. For catharsis, the lesson is that feeling better is a consequence with its own power to shape behaviour. That power does not depend on whether anything has been resolved.

The process model of emotion regulation

2026-09-14T00:00:00+00:00

Why a process model

Before the late 1990s, research on how people manage their emotions was scattered across coping research, defence mechanisms, psychophysiology and clinical practice. James Gross’s integrative review proposed a common framework: emotion regulation refers to the processes by which people influence which emotions they have, when they have them, and how they experience and express them (Gross, 1998). The key move was to treat emotion as a process that unfolds over time, rather than as a state that is either present or absent. If emotion unfolds, then regulation can intervene at different points, and strategies can be compared by when they act.

That idea is simple, but it changed the question that catharsis research had been asking. The discharge model offered a binary: express or hold in. The process model asks instead which strategy, applied at which point, changes which component of an emotion, and at what cost. Emotion regulation becomes a family of different operations rather than a single choice.

The process model rests on a simplified description of how an emotion arises, often called the modal model. A situation (external, or internal such as a memory) is attended to; it is appraised in relation to the person’s goals; and the appraisal gives rise to a response that involves experience, behaviour and physiology. The response can change the situation, which starts the cycle again.

This sequence is a working assumption rather than a finding. It draws on appraisal theories (Theory 03), and its authors describe it as a deliberately simplified picture. Its function is to provide places where regulation can act.

The five families of strategies

Each family corresponds to a point in the modal sequence (Gross, 1998; Gross, 2015).

Situation selection acts before a situation arises: choosing to approach or avoid people, places or activities according to the emotions they are expected to bring. Declining an invitation to a gathering where a painful encounter is likely is situation selection. So is deliberately seeking out a friend who usually improves one’s mood. See the glossary entry on situation selection.

Situation modification changes a situation once one is in it: asking a colleague to lower their voice, moving a difficult conversation to a private room, or solving the practical problem that is producing frustration. Situation modification is the family most directly connected to resolution in the everyday sense, because it can change the source of the emotion.

Attentional deployment changes which aspects of a situation receive attention. Distraction directs attention away; concentration directs it towards. Rumination, repeatedly focusing on a distressing event and its causes, can be understood as a persistent and often unintended form of attentional deployment (attentional deployment).

Cognitive change alters the meaning a situation has for the person. The best-studied form is reappraisal: reinterpreting what happened (“the reply was short because she was busy, not because she was dismissive”) or changing one’s goals in relation to it. Because it acts on the appraisal that generates the emotion, cognitive change can alter the emotion at its source.

Response modulation acts once an emotion is under way, on its experiential, behavioural or physiological expression. Expressive suppression, keeping the face and voice neutral while feeling strongly, is the most studied example. So are slowing one’s breathing, drinking alcohol, or exercising to change bodily state (response modulation).

The first four families are sometimes grouped as antecedent-focused, because they act before the emotional response is fully generated, and the fifth as response-focused. The model predicts that antecedent-focused strategies, especially cognitive change, should alter the whole emotional response more efficiently, whereas response-focused strategies change the output while the generating appraisal continues.

What the evidence shows

Experimental evidence across families

The most comprehensive test of the families is a meta-analysis of experiments in which participants were instructed to use a particular strategy after an emotion induction (Webb, Miles, & Sheeran, 2012). Its findings distinguish sharply between strategies that the model groups together or apart:

Cognitive change had a small-to-medium effect on emotional outcomes.
Distraction, a form of attentional deployment, was effective; concentrating on one’s feelings, another form, was not.
Response modulation was mixed: suppressing the expression of emotion had some effect on behaviour, but suppressing the experience of emotion was ineffective.

These results come from short laboratory inductions, mostly with film clips, pictures or recalled events, and effects varied with moderators. They support the model’s central premise that strategies differ, and they show that differences within a family can be as important as differences between families.

Individual differences

Studies of habitual strategy use measure how often people tend to reappraise or suppress. People who report habitually reappraising experience and express more positive and less negative emotion and report better interpersonal functioning than those who habitually suppress (Gross & John, 2003). The design is correlational, so it cannot show that reappraisal causes these outcomes; people who find reappraisal easy may differ in other ways. It is consistent with the experimental pattern.

Choosing a strategy

People do not use one strategy for everything. In three laboratory experiments with emotional pictures or unpredictable electric stimulation, participants relatively preferred reappraisal in low-intensity negative situations and distraction in high-intensity ones (Sheppes et al., 2011). The authors’ interpretation is that reappraisal requires engaging with the emotional meaning, which is feasible at low intensity, whereas distraction blocks processing early, before an intense emotion gathers force. The finding concerns choice, not long-term consequences.

The extended process model

The original model describes what strategies do. The extended process model addresses how regulation happens at all (Gross, 2015). It treats emotion regulation as a form of valuation: a system that evaluates the emotion itself as good or bad for the person’s goals and acts to change it. Three stages are distinguished.

Identification. Is the emotion something to regulate? This depends on noticing it and on valuing a change.
Selection. Which strategy family, and which specific tactic, should be used?
Implementation. How is the chosen tactic carried out in this situation?

Each stage is monitored and can fail. A person may not notice rising anger (identification failure), may habitually choose a strategy that does not fit the situation (selection failure), or may try to reappraise but produce thoughts that increase distress (implementation failure). The extended model gives a vocabulary for regulation difficulties that is more specific than “poor coping” (extended process model).

Recent work has also widened the notion of reappraisal. Rather than one strategy, it is described as a family of tactics that differ, for example, in whether they reconstrue the situation or repurpose the goals it bears on, and in whether they act on the object of the emotion or on the emotion itself (Uusberg et al., 2023). This conceptual map is proposed, not yet tested as a whole.

Where venting and reappraisal sit

The process model was not built to classify venting, and its authors do not do so in these terms. Placing venting within it is this site’s interpretation, recorded as such in the ledger. The placement is nonetheless informative.

Venting, as typically studied, involves two operations. It expresses the emotion outwardly, often with high arousal, which is response modulation. And it directs attention to the provocation and to the feeling, which is attentional deployment of the concentrating kind. Neither operation acts on the appraisal that generates the emotion. The meta-analytic finding that concentrating on feelings was not effective, and the experimental finding that venting while thinking about the provoker increased anger and aggression (Bushman, 2002), are both consistent with this placement. So is the meta-analytic finding that arousal-decreasing activities reduced anger, while arousal-increasing activities were not effective overall (Kjærvik & Bushman, 2024): if an activity works at the response level, lowering arousal is more plausible than raising it. Research note 02 covers the studies.

Reappraisal sits earlier. It changes what the situation means, so the emotion that follows is different, not merely expressed differently. That is why it is the model’s candidate for change that lasts beyond the episode. It is not, however, automatically resolution. Reappraisal changes meaning; if the situation itself should change, meaning alone may not be enough. Situation modification remains the family that acts directly on circumstances.

A useful way to state the relation to catharsis, again as interpretation:

Response modulation can change how an emotion feels or looks now. It is the natural home of relief.
Attentional deployment can shorten or lengthen an episode, depending on whether attention moves away or stays.
Cognitive change can change the emotion that a situation produces, now and on later encounters.
Situation selection and modification can change whether the situation occurs at all.

Some forms of expression combine these. Talking a problem through with someone who helps reconsider it involves cognitive change, and structured writing that builds an account may too. The label “venting” does not tell us which operations are taking place, which is one reason studies of expression disagree.

Criticisms and open questions

Is regulation separable from generation? The model assumes that emotion generation and emotion regulation are distinct processes. From a constructionist perspective, in which emotions are constructed from ongoing predictions and conceptualisation, this distinction may not hold; regulation may simply be part of how an emotion is constructed. Gross and Barrett argued that whether the two can be usefully distinguished depends on one’s theoretical view of emotion, and that making those perspectives explicit is more productive than declaring one correct (Gross & Feldman Barrett, 2011). The debate concerns framing; it does not invalidate the experimental findings, but it limits how literally the modal sequence should be read (Theory 04).

Families overlap in practice. Leaving a room is situation selection, modification and attentional deployment at once. The families are heuristic categories, and real regulation is often a blend.

Effectiveness depends on context. A strategy that reduces negative emotion in a laboratory task may not be adaptive in life. Distraction is effective in the short term but, used habitually to avoid a problem that needs attention, it may maintain the problem. Reappraisal may be less helpful when a situation can and should be changed. Research on regulatory flexibility asks whether matching strategies to situations matters more than using any particular strategy; that literature is not reviewed here.

Laboratory to life. Most experiments induce mild, brief emotions in healthy adults. Transfer to intense, prolonged or clinical distress is uncertain. For clinical applications, see Research note on clinical perspectives.

Outcomes differ. “Effective” usually means lower self-reported negative emotion shortly after an induction. Physiological, social and long-term outcomes can diverge from that measure.

What this essay concludes

The process model’s contribution is to replace a binary with a map. Once strategies are sorted by when they act, the question “does expressing an emotion help?” becomes “which operation does this form of expression perform?” The experimental evidence supports the view that strategies differ and that acting on meaning is, on average, more effective than concentrating on feelings or suppressing experience. Within that map, venting acts late and on the surface of an emotion, while reappraisal and situation change act on what produces it. That is the model’s version of the distinction between relief and resolution, and it is a conceptual lens, not a verdict on any individual’s way of coping.

The social sharing of emotion

2026-09-14T00:00:00+00:00

A common behaviour with a puzzling result

When something emotionally significant happens, people talk about it. They tell a partner about a humiliating meeting, call a friend after a frightening near-miss, describe a loss again and again in the weeks that follow. This behaviour is so ordinary that it is rarely noticed, and it is widely assumed to be how emotions are processed: talking it out makes it better.

The research programme on the social sharing of emotion, developed from the early 1990s by Bernard Rimé and colleagues, took this behaviour as its object of study. It produced two findings that sit uneasily together. The first is that sharing is extremely common and valued. The second is that, in many studies, sharing did not reduce the emotional impact of what was shared. This essay explains the theory built to make sense of both, and why the gap between them is a clear example of the difference between relief and resolution.

In this research, social sharing means talking about an emotional event, and one’s reactions to it, with others in a socially shared language. It is distinguished from private rumination, from writing and from nonverbal expression, although it overlaps with all three.

How common it is

A review of the programme reported that most emotional experiences are shared with others shortly after they occur, a finding first reported by Rimé and colleagues in 1992 and then extended across research procedures designed to overcome the limits of the earlier studies (Rimé, Finkenauer, Luminet, Zech, & Philippot, 1998). The phenomenon was examined in children and in older populations, individual and cultural differences were considered, and the relation between the intensity of an emotion and the extent of sharing was investigated. On the basis of this work, the authors proposed that sharing is an integral part of emotional experience rather than an occasional add-on. A later review summarised the theory and evidence, and described sharing as having interpersonal functions such as eliciting attention, empathy and support (Rimé, 2009).

These studies are largely based on self-report, often retrospective, and come mainly from one research programme. They establish that people report sharing most of their emotional experiences; they cannot show exactly how much sharing happens in daily life, and cultural norms about disclosure vary.

Rimé’s account treats sharing as a process that has consequences for both people involved, and for groups (Rimé, 2007).

Reactivation. Talking about an emotional event brings back its images, thoughts and feelings. Sharing therefore reactivates emotional arousal in the person who shares. It also stimulates emotion in the listener, whose responses in turn stimulate the speaker’s further expression.

Emotional communion and bonding. Because the interaction involves reciprocal stimulation, it favours empathy and a sense of emotional communion. The theory predicts that sharing strengthens the social bond between speaker and listener.

Propagation. People who have heard an emotional account often share it further, in what the research calls secondary sharing. When an emotional event strikes a whole community, there are as many sources of sharing as there are members, and each act of sharing reactivates emotion among those involved. Rimé proposed that such chains of sharing contribute to an emotional climate in the community and may affect group cohesion and collective memory. These collective predictions are theoretical and are less directly tested than the individual findings; they connect this essay to Theory 08.

Notice what this account does not claim. It does not say that sharing discharges an emotion. Its first consequence, reactivation, is closer to the opposite: sharing brings the emotion back.

The puzzle: helpful but not relieving the memory

If sharing reactivates emotion, what does it do for recovery? The review of the programme addressed this directly and concluded that findings were generally not consistent with the view that sharing alleviates the memory of the shared emotional experience (Rimé et al., 1998). In correlational studies, people who had shared an event more did not, on average, show a smaller emotional impact when they recalled it later.

Experimental work made the dissociation sharper. Participants who talked about a negative emotional experience rated the conversation as helpful, but the emotional impact of the memory was not reduced compared with control conditions (Zech & Rimé, 2005). People felt that talking had helped them. On the measure of emotional recovery used in the studies, it had not (claim).

There are several ways to read this, and they are not mutually exclusive.

Sharing serves social rather than emotional ends. People may share mainly to obtain support, maintain relationships or make sense of events together, and value those outcomes.
The benefit is real but is not recovery. People may feel less alone, more understood and calmer after sharing, without the memory losing its emotional charge.
The kind of sharing matters. Some forms of sharing may promote recovery while others do not, and studies that mix them average the effects away.

The third reading led to the most informative line of research.

The theory

The two-mode account distinguishes what a listener, or the conversation as a whole, provides. In a socio-affective mode, the listener offers empathy, comfort, validation and closeness. In a cognitive mode, the listener helps the speaker reconsider the event: putting it in perspective, reframing it, finding meaning or revising assumptions it had disturbed. The account predicts that socio-affective sharing temporarily eases distress and strengthens the relationship, and that cognitive sharing is what contributes to emotional recovery (as summarised in Brans, Van Mechelen, Rimé, & Verduyn, 2013).

An experimental test

Nils and Rimé (2012) tested this directly. Participants viewed an aversive film and then talked about it with a close person. The listener was instructed to adopt a cognitive or non-cognitive mode and, independently, a socio-affective or neutral mode, in a 2 × 2 design. Emotional, cognitive and social outcomes were measured immediately and again two days later, after participants were re-exposed to the film.

The pattern matched the prediction:

Emotional recovery, measured at re-exposure, occurred only when the listener stimulated the participant’s cognitive work.
Cognitive variables, the participant’s basic assumptions, were also more positively modified in those conditions.
Socio-affective responses produced social integration, greater closeness to the listener and less loneliness, together with an impression of feeling better.

The authors concluded that sharing can lead to several distinct benefits depending on how listeners respond: emotional recovery, consolidation of shattered assumptions, social integration and temporary reduction of distress. The title of the article, Beyond the myth of venting, states the implication. This is a single laboratory study with a film stimulus, and its abstract does not report the sample size, so the site grades the claim with moderate confidence.

Evidence from everyday life

Two experience-sampling studies examined sharing of everyday anger, fear and sadness (Brans et al., 2013). In the first, sharing was associated with a shorter duration of the emotional episode, especially when the sharing was socio-affective. The second study asked whether this shorter duration reflected relief rather than recovery by also measuring residual emotional intensity three days after the emotion began. It largely replicated the duration finding; duration and residual intensity were empirically distinct; and no relation between sharing and residual intensity was found, even when the mode of sharing was taken into account.

These results fit part of the two-mode account and complicate another. Socio-affective sharing was associated with episodes ending sooner, which is consistent with temporary relief. But cognitive sharing in daily life was not associated with lower residual intensity, which the laboratory experiment would have led one to expect. Possible explanations include that everyday “cognitive” sharing differs from instructed reframing, that the measures differ, or that the effect is smaller outside the laboratory. The study is observational, so sharing was not manipulated.

The feeling that sharing helps is not an illusion to be dismissed. Several outcomes that people plausibly value are supported by the evidence above:

Connection. Empathic responses increased closeness and reduced loneliness.
Shorter episodes. In daily life, sharing was associated with emotions ending sooner.
Feeling better in the moment. Participants reported an impression of feeling better after empathic responses.
Being understood. Sharing elicits attention, empathy and support, which are interpersonal goods in their own right.

Putting feelings into words may also contribute. In a neuroimaging study, labelling the emotion in facial expressions was associated with lower amygdala activity and higher right ventrolateral prefrontal activity than passive viewing (Lieberman et al., 2007). That is a laboratory association with a specific task, not evidence that conversation regulates emotion through the same route, but it suggests that describing a feeling is not neutral (affect labeling).

The theory also gives reasons why a satisfying conversation can leave the emotional charge of a memory unchanged. Stated as interpretation:

Reactivation without revision. Sharing brings the event back. If the account is told in the same terms each time, with the same appraisal, the retelling rehearses the emotion rather than changing what produces it (Theory 03).
Listeners tend to comfort. Many listeners respond to a distressed friend with sympathy and agreement rather than by questioning the account. That is a kind, socially appropriate response. The experimental evidence suggests it brings closeness and relief more than recovery.
Agreement can confirm the appraisal. A listener who shares the speaker’s outrage may validate the interpretation that generates it. This connects sharing to the findings on venting (Theory 01).

Cognitive work, by contrast, changes something: the meaning of the event, the assumptions it disturbed, or the account the person can give of it. Research on expressive writing points the same way, linking benefit to building a coherent account rather than to expression alone (Pennebaker, 1997). The site records this convergence as a theoretical interpretation, graded by its sources (claim).

The social sharing literature offers perhaps the clearest empirical separation of the two outcomes this site distinguishes.

Relief corresponds to feeling better after talking, shorter episodes and the reduction of distress in the moment. These are associated with socio-affective sharing.
Resolution, in the narrow sense measured here, corresponds to emotional recovery: a lower emotional impact when the event is encountered again, together with changed assumptions. In the one experiment that manipulated listener mode, this was associated with cognitive sharing.

Both outcomes are valuable. The practical reading, as interpretation, is not “stop seeking comfort”, but that comfort and reconsideration are different things, and that a conversation can provide one without the other. The advice entries on building lasting support and writing to make sense translate this into practice with explicit limits.

Open questions and limits

Few manipulations of mode. The two-mode account rests heavily on a small number of experiments. Replication with different stimuli, relationships and outcomes is needed.

Laboratory and everyday results diverge. The experience-sampling data did not find a relation between cognitive sharing and residual intensity. It is unknown whether that reflects measurement, a difference between instructed and spontaneous reframing, or a weaker effect in daily life.

Severe and traumatic events. The studies reviewed here used films and everyday emotions in non-clinical samples. They say nothing direct about talking after traumatic events. The evidence on single-session psychological debriefing, discussed in the Research note on clinical perspectives, is a caution against assuming that talking soon after trauma is always protective.

Culture and relationships. Norms about what may be shared, with whom, and how listeners should respond vary between cultures and relationships. The programme considered cultural differences, but most experimental work comes from Western European samples.

Who benefits from sharing? Sharing also affects listeners, who are exposed to reactivated emotion. The costs and benefits for listeners, and for relationships over time, are less studied.

What this essay concludes

The social sharing of emotion is a near-universal behaviour that people value, and the theory built around it explains why: sharing reactivates emotion, invites empathy and strengthens bonds. The same theory explains why sharing need not reduce an event’s emotional impact. The best available experimental evidence suggests that recovery follows when sharing involves cognitive work, and that empathic sharing brings connection and temporary relief. That is a modest, evidence-bounded version of an old intuition: talking helps, but what it helps with depends on what the talk does.