Reproducible lab · Synthetic subtitle evidence

Subtitle timing and accessibility boundary benchmark

Exercise the real local SRT/WebVTT engine across safe, blocking, uncertain, malformed and adversarial inputs. The benchmark measures deterministic parser decisions and byte integrity—not language quality, media synchronization or platform acceptance.

Worked synthetic evidenceEvery input is generated and checked in. Results are not customer outcomes, production usage, rankings, independent human certification or a claim that any fixture is an accessible finished caption track.

Engine, sources and results reviewed 2026-08-09. This lab contains no advertising.

26 / 26unit gates passed

70 / 70adversarial expectations matched

445 / 445browser assertions passed

20 / 20independent attacks passed

0false PASS observations

0release blockers

Method

Hash exact core, renderer and synthetic fixture source bytes.
Compute the engine manifest from the core and renderer hashes; abort artifact generation if that identity is absent.
Run the three named fixtures and 70 adversarial cases through the same bounded engine used by the browser.
Generate same-format output, conditional conversion, CSV, JSON, SVG, Markdown and Receipt v1 at a fixed clock.
Repeat artifact generation three times and compare filename, body bytes and SHA-256.
Exercise Chromium and WebKit at desktop, 390 px and 320 px in EN/TR and light/dark: 24 tool cases, 445 assertions.
Mutate one artifact byte and one receipt-core field, then prove the original declared hash no longer matches.

Expected versus observed

Scenario	Expected	Observed	Evidence boundary
Supported 3-cue SRT, declared 8 s	PASS	PASS · 7 artifacts	Receipt core `1204f9ff64aac95dc5359d7c7c64259b7b6390ce20cb67105a4f64a81a236b05`
Overlap plus cue beyond declared media	BLOCK	BLOCK · 7 artifacts	Every output is review-only; receipt core `9ac24f4089fcf8c24b46b708c6db8b187fdc7168fd87542b8a18e2a55a5267a0`
WebVTT STYLE, REGION, settings and dense text	INDETERMINATE	INDETERMINATE · 6 artifacts	Unsupported rendering surfaces remain explicit; receipt core `360825c4bd5085bc7033f99d0f111e0ba424d8024d15eb68f610e3a5f71038d5`
Missing media duration	INDETERMINATE	INDETERMINATE	The engine cannot prove the out-of-media boundary.
WebVTT non-breaking-space timing separator	Reject	Rejected	`invalid_timing_line`; no artifact published.
10,001 cues or input over 5 MiB	Reject	Rejected	Bounded input fails closed before delivery output.
WebVTT cue IDs converted to SRT	No lossy conversion	Conversion withheld	Same-format VTT evidence remains available.
Formula and script-shaped cue text	Inert evidence	PASS	CSV is formula-neutralized, HTML is inserted as text, and SVG omits raw cue text.

Tamper experiment: integrity fails visibly

Original fixture receipt

Declared receipt-core SHA-256

1204f9ff64aac95dc5359d7c7c64259b7b6390ce20cb67105a4f64a81a236b05

Declared normalized SRT SHA-256

b6a82ed5491773946c10d62df8de6c99016b4b9e6f4cb81ec369f61bb8dc847b

One-field / one-byte mutation

Changing result_summary.cue_count from 3 to 4 recomputes the receipt core as:

322afec69b36dbbed29e648a55dc239070eb0666e41ddeb20d04e977afca87dd

Flipping one byte in the SRT recomputes it as:

d05db1fcecd691c2ce7cad4c240b0add006048fc8b505b8e51d3601938c67ae9

Observed: both mismatches are BLOCK conditions.

This detects integrity drift only. A matching receipt does not establish authorship, source authenticity, factual correctness or accessibility.

Coverage and deliberate uncertainty

Syntax and timing

SRT/WebVTT structure, timestamp grammar, start/end order, source order, overlap, media boundary, gaps and bounded inline timestamps are checked.

Review heuristics

Duration, estimated characters per second, line count and line length are deterministic profile signals. Crossing one requires review; it is not automatically an accessibility failure.

Unsupported surfaces

WebVTT STYLE, REGION, cue settings, markup and identifiers are inventoried. Where faithful rendering or conversion is not proven, the engine returns uncertainty or withholds conversion.

Not measured

Audio accuracy, shot changes, frame-level sync, translation, meaningful non-speech audio, speaker correctness, player rendering, safe area and real-user comprehension.

Exact source and report manifest

Tool pageb3d2a7367895c5cf1e1c6e0854c649174edb1b39a0978bde6e9e6cfdf74ed3af
Engine core478b13db5ea9551a3d4ea8c16853b9b443477166831ba56f1745b8e815a6e87f
Renderera48e2b82e59e6c53cdcc13ebc202738d534354eba6ee58a847fa2a4822ffb8fb
Synthetic fixture module8e80ae62e039a57fcd3a5058bd375e73636b46d1c268e2cec582eb6c250ef967
Engine manifeste7424750d6c2376baad7472d2074c5a74294458a9717883e44940ede77246604
Unit report172f53011dda6a77e335b8eeab6ab77c05424d5367752ef044361417a4c7ce60
Browser report2a31604bddffa620a4dc3947a22c35381970130e3dd3588f3b28cb20c07793ac
Independent red-team reportb22a153d57dfefb7ded5941c300310358c067f9ecf29b7a32d14aef403014467

Reproduce

shasum -a 256 app/tools/subtitle-qa-studio/index.html \
  app/js/renderers/subtitle-qa-core-20260809.js \
  app/js/renderers/subtitle-qa-20260809.js \
  app/fixtures/subtitle-qa-v1/fixtures.mjs
node scripts/subtitle_qa_unit_20260809.mjs
node scripts/subtitle_qa_browser_20260809.mjs
node scripts/subtitle_qa_independent_redteam_20260809.mjs
node scripts/subtitle_cluster_gate_20260809.mjs

Run from the repository root with Node.js, Chrome/Chromium and Playwright WebKit available. A source change must update its evidence; copying old hashes into new content is a failing provenance condition.

Sources, privacy, limits and correction

W3C WebVTT, 20 May 2026 Candidate Recommendation DraftFormat and parsing basis; reviewed 2026-08-09.
W3C WAI Captions/SubtitlesAccessibility context, not a claim that this parser audits WCAG.
Netflix timing guidancePlatform-specific editorial reference, not a FastTool certification profile.
BBC Subtitle GuidelinesBroadcaster-specific editorial reference, not a compliance claim.

Fixtures are synthetic; no user subtitle data is collected by this lab. The linked workbench processes supported bytes locally. Normal static assets and consented analytics may use the network, but user subtitle text, filenames and hashes are excluded from analytics.

Supported input and claim boundaries are detailed in the paired workflow guide. Verify packets with the Receipt Verifier. Report a reproducible defect to corrections@fasttool.app with the scenario, source hash and observed output.