Průzkumník tvrzení

Holding play sounds the tone at the set relative gain, and releasing it leaves no audio node running

Every V1 signal pattern plans exactly the automation the frozen V1 file planned

A sweep pattern plays a frequency that rises through the requested range, and stopping it leaves no audio node running

Two oscillators at 440 Hz and 442 Hz mixed to mono produce a level envelope beating at 2 Hz, and the stereo split puts one on each channel

The spectrum reads a bin-centred full-scale sine at 0 dB relative to full scale and interpolates the peak frequency between bins

The spectrogram maps relative level to colour through a monotone lookup table on a log or linear frequency axis

Starting the microphone requests audio with echo cancellation, noise suppression and automatic gain control off, feeds it to an analyser and never to the output, and stopping it ends every track

The generator-versus-microphone comparison reports the frequency difference in hertz and cents and flags a match, a harmonic, a mismatch or no signal

The sequencer compiles a block timeline into sorted Web Audio automation whose gain never reaches zero and whose frequencies stay inside 20 Hz and the safe maximum

The pattern sequencer plays through the Filter Lab's stage on an engine-owned stage chain that the AudioEngine counts while it sounds and releases after the last pass, live and in the sequence's WAV export, while the Envelope and Additive labs say they shape the instrument only

An enabled 500 Hz low-pass attenuates a 5 kHz tone below a fifth of its level, and bypassing it restores the level

The ADSR envelope schedules attack, decay, sustain and release with Web Audio set, linear and exponential semantics, and a release mid-segment continues from the current value

Additive synthesis plays a PeriodicWave whose harmonic coefficients are exactly the ones the harmonic bars draw

The phase view and the Lissajous figure are computed from the two frequencies and phases, one period for near-unison pairs

The stereo correlation meter reads +1 in phase, -1 inverted and about 0 in quadrature, and resets on silence

Every hearing range and call band shown comes from a cited publication, carries its criterion, and is shown with a disclaimer

A copied link restores the instrument configuration through the validated path, with the gain capped

A saved preset is stored locally and loading it restores the same configuration; blocked storage never throws

WAV export writes a valid RIFF/WAVE file (16-bit PCM or 32-bit float) of the current signal that parses back to the same samples

An exported configuration file carries every specified key and no runtime objects, and importing it restores the configuration through the validated path

The build packs the modular source into one dist/index.html deterministically, inlining styles, vendor scripts and the application as classic scripts

The committed dist/index.html loads nothing at runtime; module scripts, import maps, external scripts and runtime fetches are refused

The built page boots without console errors and plays in Chromium, Firefox and WebKit, both from file:// and from a /oscilla/ sub-path as GitHub Pages serves it

About is the last primary navigation item; it opens a full-width view that explains OSCILLA, with exact source, Majordomus and contact links, laid out without overflow from 320 to 1536 px

The development times the About view states are the commit times of the commits it names

Every deployed page names its release state: only the release commit shows the bare version, and any later commit shows <version>+<distance>.g<short>, so a page never presents itself as a release it is not

The About view's evolution timeline names every published release line and marks the release line of the running build (package.json MAJOR.MINOR) as current, so a minor or major release cannot ship with a stale About page

No text the page or its modules can show claims a medical, therapeutic, wellness, cognitive, pseudoscientific or animal-specific effect, inaudibility, the safety of a frequency, ultrasound below 20 kHz or certification, except in a sentence that negates it

README's opening line, the About lede, the page's meta description, the og:description and twitter:description generated from scripts/social-assets.mjs and package.json's description contain the positioning sentence Q002 chose verbatim within 200 characters, the title surfaces and the brand label carry the name within 70, and a self-description the test neither checks nor exempts with a reason fails it

The first screen is the Playground Q004 chose (option A): its h1 is the positioning sentence Q002 chose, above every panel; at 1536x1024 Source & Pattern, Signal analysis, Microphone Analyzer and Device & Limits stand in one row, MEASURE's own Guided measurement panel (seven steps, Check setup, Frequency and Level UNCALIBRATED) spans the next row inside the first viewport, and the Pattern Sequencer with the compact Studio beside it follows, the compact Studio starting in that viewport with its whole signal path shown; at 390x844 the sentence, Source & Pattern, Signal analysis, the measurement row and the compact Studio run top to bottom (the Microphone Analyzer, Device & Limits and the Pattern Sequencer stay in their workspaces on a phone, as before) with the six frequency steps and Mode folded behind Source & Pattern's own disclosures and the measurement steps folded to the current one; no guided path is shown; Check setup, and Repeat after a result, pressed there open MEASURE and run in it; and every entry path of the committed inventory lands as before

About's Origin tells the start Q005 decided against the record, an anti-bark ad promising 'ultrasound through any phone's speaker' and the reply 'this can be knocked together in one HTML5', keeping the browser-laboratory question only as the framing the project grew into; 'Built in one night' keeps its 5 h 51 min qualified as wall-clock time with several AI workers in parallel; and About names no company behind the ad, no vendor frequency, no personal motive and no figure the knowledge records do not measure

Each deep link of the README landing opens what its text names: the Playground, the Measure workspace with a recipe of two 3-second log sweeps that the decoder accepts whole, and the Studio on the Subtractive Synth template's graph; each claim id the landing cites is a guaranteed claim of this file, and the landing links the extending guide

The derived regions of README.md (every workspace with its deep link and the map of how the workspaces connect, the Studio templates, the Synth engines, the gallery of the gated visual references and the embedded deploy charts) are rendered from the registries by scripts/facts.mjs and scripts/readme.mjs, and npm run readme:check in verify fails on a region edited by hand or a registry changed without regenerating

The landing at /oscilla/lab/ is generated by scripts/landing.mjs from the deployed commit, byte-identical on a rerun, in Czech (lang cs) with Czech copy for every workspace and engine of the registries (a registry id without it fails the generator), a card linking every workspace into the app, every chart of CHARTS with a title and a description, no iframe until the visitor presses a start button, and no script, stylesheet, image or frame from another origin

Every file, module export, link target and npm script that docs/EXTENDING.md names exists, every import of its code examples resolves to an export, and its example Studio node type validates in createNodeRegistry and compiles through compileStudio to a ready node whose adapter flips the polarity, and the guide's unit test of that node passes as printed in its own process

The Measure workspace, separate from the Playground, takes a measurement through setup check, noise check, sweep, analysis and review to a saved experiment, and the instrument cannot play while a measurement owns the output

Aborting a measurement in preflight, noise check, armed, measuring or analysing, by Escape, STOP, page hide or leaving the workspace, ends it ABORTED with no audio node, source, capture, worklet port or input track left

The Measure and Experiments workspaces of the built page run in Chromium, Firefox and WebKit from file:// and from a /oscilla/ sub-path, with the capture worklet loaded from a data URL and no console errors

A frequency-response calibration profile is imported from CSV, TXT or JSON with line-numbered errors, identified by the SHA-256 of its normalised points, interpolated linearly in dB over log frequency and never extrapolated by the application; uncovered frequencies stay uncorrected and marked

No level is labelled dB SPL unless a valid absolute level calibration with an explicit external reference applies; otherwise every level is dB relative (dBFS-like) and marked UNCALIBRATED

The transfer function of the playback and capture chain is recovered from a captured log sweep, a known flat gain within 0.1 dB and known filters within 0.5 dB of their analytic magnitude, with phase only under a robust alignment and a valid range that ends where the signal is less than 10 dB above the noise

The impulse response is derived by regularised spectral deconvolution of the captured sweep, an identity system giving one peak at the pre-roll within one sample and a 0.5 echo reading -6.02 dB within 0.5 dB, and windowing and normalisation return new objects that never alter the original

Octave and one-third-octave band levels integrate power between the base-10 band edges below 0.95 of Nyquist, and the band averager in fast and slow mode reaches 63.2 percent of a step after one time constant whatever the frame rate

A level calibration is stored only with the input it was taken with and applies only to a measurement captured from that input, as the measurement engine checks against the input it captures from; a reading typed before any input is known is refused with the reason, a calibration without a binding (an earlier record) never applies to a known input, a record whose calibration is bound to another input than its own is a finding, and records say when a calibration is not bound to an input

A record whose stimulus this build's engine cannot measure (anything but a log sweep) is kept on import with a finding that says so, and Repeat and "Run this definition" refuse it rather than run a log sweep in its place; records the application saves are log sweeps only

Every saved measurement is an experiment that records its recipe, output level, input device and constraints, calibration, sample rate, runs, quality, algorithm IDs, product version and build, with a configuration hash and a result hash

A saved experiment records the calibration the measurement engine applied and the notes as they were when the measurement started, never a frequency profile or level calibration loaded or created after the run; a stored or imported experiment whose named calibration its own results contradict stays readable, is marked with that finding and is presented and compared as uncalibrated, never as dB SPL

An experiment survives create, encode, JSON and validate identically, its typed arrays written as explicit little-endian base64 that decodes bit for bit

An imported experiment is untrusted input; oversized files, length mismatches, wrong types, non-finite numbers, unknown algorithm IDs, prototype keys, a future schema version or a result hash mismatch are refused with a reason and nothing is stored

Experiments persist in the browser IndexedDB database oscilla-experiments, survive reopening, are deleted only explicitly, and fall back to an in-memory store that says so when IndexedDB is unavailable

A stored completed experiment run cannot be changed; only its name and annotation notes change, through a separate metadata operation, and a duplicate is the same run under a new ID that records its source

Every stored result names the algorithm that produced it by a versioned ID, an unknown ID is refused on import, and every known ID reproduces its golden output

Comparing experiments flags differences in calibration, sample rate, stimulus and algorithm, and shows A minus B only over the overlapping valid range of equivalent experiments, never normalised

Comparing runs lists what changed between them by domain (recipe, algorithms, calibration, conditions, Studio graph, build, result, metadata) with units, classifies Studio layout and view changes as presentation and Studio changes outside both runs' measured paths as not used by the measurement, never as execution, and compares against the one run marked as the baseline, a mark that is metadata and leaves the run immutable

A run records the version of the experiment definition it was executed from, only when it ran exactly that version's recipe; a definition's hash covers its recipe, declared conditions and acceptance criterion but not its name or notes, an edit of those fields appends a version that earlier runs never move to, and a run without an authored definition carries one derived from its own recipe and marked as derived

A saved run's detail traces one stored result point through the analysis, capture, calibration as applied, run, definition version, build and Studio provenance the record stores, and lists a reproducibility checklist whose items are recorded, partial or not recorded, each with its reason; nothing is inferred, a missing field reads "not recorded", no item is recorded without its field, and there is no score

A finding is a user's interpretation stored apart from the experiments it cites, with a categorical status (observation, hypothesis, supported, contradicted, inconclusive; no confidence number), typed references to experiments, comparisons and stored values, and the id and result hash of each cited experiment; a supported or contradicted finding cites at least one reference, no finding write changes an experiment, a reference to a deleted or absent experiment is kept and reported as missing, never removed, and every claim about a cited experiment is derived from one verification that reads the stored record fresh at the moment of the claim (a reference's state on each refresh, Open, linking, backlinks and the words "a stored grid point"), so a reference reads present only when the stored experiment is readable and carries the cited result hash (a different hash is a different record; a hash missing on either side is an identity that cannot be verified; an unreadable record is unreadable), and a finding recorded from a view cites the record on screen and is refused when another tab replaced it

From a stored experiment, definition, finding or Studio project, OSCILLA lists what it is connected to and what depends on it, each entry from one stored field that it names, never inferred from names, recipes or times; an entry reads "stored here" only for a stored record that is readable (its result hash recomputed) and is the record the field names, and otherwise says missing, does not match, not verifiable or unreadable, never hiding or repairing the reference; a build and a frequency profile, which are never stored, read "running here" and "loaded here" instead; the entries under an experiment's detail are those of the record the detail shows

A measurement of two or more runs stores the aggregate (centre, dispersion and repeatability) as its primary response; the stored transfer is the aggregate centre marked as derived, never one run, and comparison uses the aggregate

Every measurement is rated GOOD, USABLE, POOR or INVALID by the versioned rule set oscilla.confidence.v4 from named metrics (signal-to-noise, clipping, dropouts, discontinuities, coverage, repeatability, calibration), always with its reasons

OSCILLA runs the offline measurement analysis in a Worker loaded from a data URL when the platform has Worker, so that no analysis step blocks the main thread, with results bit-identical to the inline fallback

A raw capture lives only for its analysis: once a measurement completes, is INVALID, fails or is aborted the engine references no capture; a saved experiment and its export carry derived results only, and the store and an import refuse a record that carries raw samples; the page says where a run is saved, in About and in the README that the stored impulse response can carry sound the microphone picked up after the sweep

The address names the workspace: one dispatcher turns a hash into a workspace (the Studio keys, then m, then a recipe link, then the Playground), every workspace switch is one history entry, Back and Forward move between workspaces without applying a link's keys again, a refused link is replaced in the address so that Back never reopens it, and a reload reopens the workspace the address names

While anything would be lost on leaving the page (unsaved Studio changes, a completed measurement that is not saved, a level calibration held in page memory, a typed name or rename that is not stored, records held only in the memory store), the page holds exactly one beforeunload listener and the browser asks before a reload; with nothing to lose it holds none and asks nothing

Space has one owner in each workspace, the same at every width: Hold to Play in the seven workspaces that play, load or analyse the instrument's signal (Playground, Analyzer, Filter Lab, Compare, Synthesis, Sequencer, Presets), the Studio transport in Studio, and nothing in Measure, Experiments, Learn and About; one shortcut dialog, opened by Help, the overflow menu and Studio's keyboard button, lists the Space meaning and the keys of the workspace in view and the global keys, each key once

window.OSCILLA ships in the one tested page with a pinned key list, and each member observes, drives only an action a user has through the user's validation (setValues refuses unknown keys, out-of-range values, a toggle value that is not true or false and a call while a measurement runs, whole), or injects only as TEST CONTEXT (showResult refuses a result without testContext, setInputNow refuses outside TEST CONTEXT loopback); an input or a level calibration named in TEST CONTEXT is not one outside it

A StudioModel is plain serializable data: AudioNode, MediaStream, DOM, canvas, typed array, function and class-instance values are rejected before it is stored or serialized.

A normalized StudioModel serializes deterministically, and its studioHash covers execution state only: positions, names, markers, metadata, view and selection never change it.

Only compatible typed ports connect; every rejected connection carries a readable reason and leaves the model, history and revision unchanged.

Instantaneous audio feedback, control cycles and analysis cycles are rejected with the cycle path and a reason before anything reaches Web Audio.

Undo and redo restore exact semantic state: undoing every action returns the initial model, a continuous drag is one history entry, and a new edit after undo clears redo.

Copying and pasting a subgraph creates fresh ids, keeps only the edges inside the selection and offsets positions, as one undoable step.

Studio files are imported as untrusted data: size, nesting depth, counts and the strict schema are checked before use, nothing is evaluated, and a newer schema is refused.

The Studio schema version is its own integer, independent of the product and experiment versions, and changes only through registered stepwise migrations.

Timeline clips are placed in absolute seconds and carry sequencer block payloads checked by the sequencer block rules; timeline edits validate their track and target references and undo exactly.

Automation lanes accept only automatable parameters, one lane per parameter, keep points sorted by time, and refuse an exponential ramp that touches zero.

A Studio measurement topology is typed analysis routing: the sweep reference and the calibrated microphone capture reach the Transfer Analyzer on separate ports that cannot be swapped, and a microphone can never reach Master Output.

Every Studio node type names the existing engine export it compiles to, takes its defaults from the engine, and holds no audio object.

The full Studio graph editor, the Inspector, the timeline and the compact Studio widget on the Playground all render one canonical StudioModel through one store; an edit in one is the model the others show, and no view keeps a synchronized copy.

The running Web Audio topology is compiled from the StudioModel through the existing AudioEngine, with Master Output feeding only the engine safety chain, so the graph that sounds is the graph on screen.

Editing the graph while it plays patches the running graph in transactions with crossfaded routes, keeping the click ratio below a fixed threshold, and leaves no node, source or connection behind after stop.

The Studio runtime states what runs as data: every validation, compiler, runtime and transport diagnostic carries a machine code; a compiled plan has a deterministic planHash; the applied record (revision, studioHash, planHash) is set only when a runtime transaction commits, and a refused edit leaves it as it was; one pure verdict says whether the runtime is not applied, in sync, refused or behind the desired model; and the transport does not adopt a model the runtime refused

Each Studio dispatch, undo, redo, cancelled gesture, document replace, Play, Stop and sync is traced under one correlation id from the intent through the compiled plan (with its planHash) and the runtime apply to each AudioParam value and route gain the node adapters report they wrote, a key no adapter wrote is traced as stored or owned with no audio time, and a refused or unapplied edit is traced as refused with its diagnostic code or as not applied; the trace is bounded, in memory only and covered by no hash.

Studio transport, clips, loop and automation are scheduled on AudioContext time, pattern clips through the existing sequencer compiler, and the playhead only observes the scheduler.

A saved project loads back to the exact model, and a saved patch (nodes, parameters, internal cables and their automation lanes, no tracks or clips) inserts or replaces only on explicit user intent; nothing saved is ever overwritten silently.

A measurement run from Studio goes through the V3 measurement engine with a recipe derived from the graph and its clips, and the saved experiment records the Studio schema version, the studioHash and the execution state of the graph it was run from, and the measured path the recipe was derived from with its own hash, all outside configHash.

Connecting ports, moving nodes, moving, resizing, duplicating and deleting clips, and editing automation points each have a keyboard path that dispatches the same store actions as the pointer path, without pointer dragging; on touch, ports connect by tapping.

The Studio workspace of the built page runs in Chromium, Firefox and WebKit from file:// and from a /oscilla/ sub-path with no console errors, inside the single dist/index.html that loads nothing at runtime.

OSCILLA will state whether Web MIDI is unsupported, needs permission, is denied, has no devices or is connected (naming the devices), and nothing else will change when MIDI is absent.

Note On and Note Off from a MIDI input sound through the AudioEngine, with velocity under the safety level cap, bounded polyphony and click-free starts and stops.

A note held on the computer keyboard, the on-screen keyboard or a MIDI key sounds past the playback limit only while it is held, under the output ceiling, and is released on blur, page hide and Escape.

The first gesture starts audio on a phone as on a desktop: the unlock listens in the capture phase at pointerdown, keydown, pointerup and touchend, never at a bare click, and the page asks for a playback audio session where navigator.audioSession exists, keeping the browser default without it or on a refusal.

A MIDI response mode will produce the same output events for the same input events, scheduled on the audio clock, with no network call or model inference in the path.

Incoming and generated MIDI notes will be shown with their names and velocities beside the spectrum and waveform.

A MIDI jam will be recorded as an editable event track that saves, reloads and replays identically.