wai.access.services: caption quality anyone can recompute, audio description, sign language and a completeness claim (ACC) #396

Merged

dcharlot merged 8 commits from wai/access-services into main 2026-10-21 21:38:39 -04:00

Conversation 0 Commits 8 Files changed 21 +5602 -2

dcharlot commented 2026-10-21 12:30:56 -04:00

Owner

Copy link

What this adds

wai.access.services (wai/extensions/access-services.md): access services for a room session, and figures about them that any verifier recomputes from signed records.

• Caption quality from the signed revisions (§3, §4).

  • Latency is each cue's first display on the room clock less the capture of its first word. A sources record places the words on a take, mapped through the take's opening commitment (capture-binding §9.3), with the take checked to derive from a record_start to the commitment's author. Words can also be placed directly on the room clock. The class capture, stated or none says which applies. The figure gives the mean (rounded down), min, max, p50 and p95, all at a fixed rank.
  • Stability is the revisions after first display, the words they changed (substitutions and insertions of rᵢ against rᵢ₊₁, so growth of a partial is not a change), and how many cues were corrected.
  • Accuracy is the words of the first revision shown against a signed reference transcript, or against the reviewed final revision. A prefix alignment reads the reference only as far as the partial reaches. With a transcript, a full alignment also covers the final text. The class transcript, reviewed or self names what the reference is.
  • Words fold ASCII case and edge punctuation, and the walk back of the alignment is one fixed rule, so the substitution, deletion and insertion counts are the same for every verifier.
  • A quality record states the figure and names its inputs by digest. It is refused figure_mismatch or figure_inputs_mismatch when it does not match what the verifier recomputes.

• Audio description (§5). A description record holds:

  • the script, with each event's interval, DAPT content descriptor and text;
  • the ducking of the programme audio under each event;
  • the voicing, either a separate track or a sink voice with a pinned model.

  Script and voice each carry an origin label: human, machine or machine-reviewed. A voice must start between 35 ms early and 45 ms late of the end of the duck, and end no later than 45 ms after the duck begins to rise. These are DAPT §5.5's bounds, which DAPT takes from EBU R 37. The record maps, written by hand, onto a DAPT document (W3C CRD 2026-06-26): the ducking becomes Mixing Instructions on the Script Event, and the voice an Audio Recording or Synthesized Audio. Player obligations cite EN 301 549 V4.1.1 clauses 7.2.1–7.2.3 and 7.3.

• Sign language (§6). A signing record binds each object of a wai.avatar.pose stream, or of a video of a signer, by digest to the final caption revisions it translates. A review (scope translation) signs it off. Its figures are each segment's delay behind the speech and, for video, a floor: QVGA and 20 fps, from EN 301 549 V4.1.1 clauses 6.5.2 and 6.5.3, which clause 4.2.4 NOTE 3 names.

• Completeness claim (§7). For each declared captions, description and signing track, the verifier checks:

  • the track is present;
  • it is finished: cues resolved, events voiced, cues signed;
  • it is in the declared language;
  • it is within the latency and offset bounds the claim states;
  • its review is present, names this signing, and is not rejected.

  §7.3 lists the tolerance each text states, and where none does. The claim also reports how much of each description is spoken over captioned dialogue, as a figure, not a refusal.

• §9 maps each figure onto 47 CFR 79.1(j)(2), (j)(3) and (k)(2)(iv)/(vi), 79.3(a)(3), the UK regulator's 2013 and 2024 statements on live subtitling, and EBU Tech 3370 §2.3.2–2.3.3's authoring delay. It also lists what is not computed: NER-style severity grading (which needs a person), capitalisation and edge punctuation, speed and placement.

Implementation:

• the access_services module of wai-rs, feature access_services = ["timed_text", "capture_take"], with no new dependency;
• wai_access_services_vectors gen|verify and tests/access_services.rs;
• the corpus live-conformance/access/;
• the second implementation, live-conformance/access_services_verify.py;
• a CI step in wai-conformance.yml plus its paths: entry;
• the REGISTRY entry and export (registry/capabilities.json, vocab/v2.ttl, the REGISTRY tables), with 7 DocumentKind wire identifiers and the tag:…/wai/access scheme;
• a SPEC §5 paragraph and the record lists, the CHANGELOG, the live-conformance index and corpus README;
• the timed-text §1 scope note now points audio description here.

Design decisions and why

• JSON records with JCS signatures, not binary. They follow synthesis and ai-usage, merged on main, using jwp-receipts §2.4.3 encodings and §3.1 step 1's strict reading. The signature is Ed25519 over wai:access\x01 ‖ JCS(record − sig), and the record id is SHA-256(JCS(record)). Signed integers (latencies, offsets) are decimal strings, because the JCS subset holds no negative number.
• A source is the aligner's statement; the media's time is not. On a take, the capture clock is the device's signed commitment, and the alignment is attributable to the signer and checkable against the master. The latency class makes the difference visible. Who may capture, and for whom, stays a take or room verifier's check. §12 says so.
• First-display accuracy uses a prefix alignment that ends at the greatest best column. A growing partial then makes no errors for words it has not reached, and a wrong last word is a substitution, not an insertion. Both cases are covered by a unit test.
• A voice must lie within its event, since DAPT truncates an audio at its Script Event's end (§4.9.1). The writer can then express every voice, and the sync bounds apply inside it.
• Sink-synthesized description is written as preRecording with tta:speak. DAPT has a SHOULD NOT for Audio objects in a pre-recording script and for Synthesized Audio in an as-recorded one. §5.5 states this departure and why.
• A signing binds final revisions only (sign_revision_not_final). It translates what the captions resolve to; a superseded partial is not that.
• Records are checked kind by kind: sources, references, descriptions, signings, reviews, quality, claims. A record another names is checked before it, whatever the order given.
• Synthesis coordination, read, not edited. The voice and signer origins are stated as synthesis §3.1's class synthesized/generate/speech-synthesis, with oversight fully_autonomous or human_validated. The review reuses synthesis §9.1's vocabulary with scope translation.

Checks (all cargo through rb on the build host)

check	command	result
unit tests	cargo test --locked --no-default-features --features access_services --lib access_services::	6 passed
corpus test	cargo test --locked --no-default-features --features access_services --test access_services	4 passed (regenerates byte for byte, no other file, every outcome, every §10 reason refuses a case, §10 equals Reason::ALL)
timed-text (text edited)	cargo test --locked --no-default-features --features timed_text --test timed_text	4 passed
registry	cargo test --locked --no-default-features --lib codecs::registry	14 passed
registry export	cargo run --no-default-features --bin wai -- registry export then registry export --check	regenerated, then "up to date"
reference verify	wai_access_services_vectors verify ../live-conformance/access	PASS, PASS
second implementation	python3 -I live-conformance/access_services_verify.py live-conformance/access	38 vectors and 4 documents PASS; 125 of 125 mutated copies fail
timed-text verifier	python3 -I live-conformance/timed_text_verify.py live-conformance/captions	PASS
clippy	cargo clippy --locked --no-default-features --features access_services --all-targets	no notes in access_services* or codecs/registry.rs (two fixed: is_multiple_of, a type alias); the pre-existing notes elsewhere are unchanged
lockfile	cargo metadata --locked --offline --format-version 1	ok, no dependency added
repo checks	check-tests-in-ci.py, check-workflow-concurrency.sh, check-spec-refs.py, check-wai-section-refs.py, check-no-vanity-metrics.sh, test-public-text-checks.sh, check-comparative-prose.sh, check-orphan-sources.py	all exit 0. The section-ref checker first flagged a [begin, …) interval before a link, so the intervals were reworded.
names	check-public-names.sh over the 15 touched files with the private list	clean

Falsifications (committed first, each restored after)

defect introduced	detector	result
Rust: drop the 35 ms early bound	--test access_services	exit 101: description_early accepted where refusal is stated
Rust: prefix alignment ends at the least best column	--lib and --test access_services	exit 101: unit test, and the lesson's figures differ
Python: drop sign_revision_not_final	access_services_verify.py	exit 1: sign_partial reaches track_incomplete
Python: drop the early bound	access_services_verify.py	exit 1: description_early accepted
125 mutated corpus copies (stated figures, flipped record signatures, altered events, moved places and reasons, a byte in each DAPT file)	access_services_verify.py	every copy fails

Regulatory and standards facts, each read at its source

• EAA Annex I §IV(b)(ii) and Art. 31(2): read from a reproduction of the adopted text, because the Official Journal could not be fetched; the extension says so.
• EN 301 549 V4.1.1 (2026-09): the ETSI PDF.
• WCAG 2.2: W3C.
• 47 CFR 79.1 and 79.3: eCFR as of 2026-10-01.
• DAPT: the CRD of 2026-06-26.
• EBU Tech 3370 v1.0.
• The UK regulator's statements of 16 Oct 2013 and 15 Apr 2024: archived PDFs, since the regulator's site refuses fetches.
• Two facts from the sweep were corrected at source:
  • EN 301 549 V4.1.1 says "subtitles", not "captions";
  • the 2024 statement's 4.5 s is a mean across live programming, not a per-programme limit.

Not done

• Spoken subtitles (EN 301 549 7.1.5), extended audio description, 79.103(e) previewability, and component-level access attributes with HLS CHARACTERISTICS and DASH Accessibility mappings: R7's INGEST part, left out of this slice.
• The objects a signing or a voicing names are digests only. The corpus has no pose stream, video or audio, and the figures do not decode media.
• NER severity grading needs a person, so it is not computed. The figure gives the error counts a grader starts from.
• Sync of a sign track has no stated standard tolerance; the claim states its own (§7.3).

🤖 Generated with Claude Code

## What this adds `wai.access.services` (`wai/extensions/access-services.md`): access services for a room session, and figures about them that any verifier recomputes from signed records. - **Caption quality from the signed revisions (§3, §4).** - **Latency** is each cue's first display on the room clock less the capture of its first word. A sources record places the words on a take, mapped through the take's opening commitment (capture-binding §9.3), with the take checked to derive from a `record_start` to the commitment's author. Words can also be placed directly on the room clock. The class `capture`, `stated` or `none` says which applies. The figure gives the mean (rounded down), min, max, p50 and p95, all at a fixed rank. - **Stability** is the revisions after first display, the words they changed (substitutions and insertions of rᵢ against rᵢ₊₁, so growth of a partial is not a change), and how many cues were corrected. - **Accuracy** is the words of the first revision shown against a signed reference transcript, or against the reviewed final revision. A prefix alignment reads the reference only as far as the partial reaches. With a transcript, a full alignment also covers the final text. The class `transcript`, `reviewed` or `self` names what the reference is. - Words fold ASCII case and edge punctuation, and the walk back of the alignment is one fixed rule, so the substitution, deletion and insertion counts are the same for every verifier. - A quality record states the figure and names its inputs by digest. It is refused `figure_mismatch` or `figure_inputs_mismatch` when it does not match what the verifier recomputes. - **Audio description (§5).** A description record holds: - the script, with each event's interval, DAPT content descriptor and text; - the ducking of the programme audio under each event; - the voicing, either a separate track or a sink voice with a pinned model. Script and voice each carry an origin label: human, machine or machine-reviewed. A voice must start between 35 ms early and 45 ms late of the end of the duck, and end no later than 45 ms after the duck begins to rise. These are DAPT §5.5's bounds, which DAPT takes from EBU R 37. The record maps, written by hand, onto a DAPT document (W3C CRD 2026-06-26): the ducking becomes Mixing Instructions on the Script Event, and the voice an Audio Recording or Synthesized Audio. Player obligations cite EN 301 549 V4.1.1 clauses 7.2.1–7.2.3 and 7.3. - **Sign language (§6).** A signing record binds each object of a `wai.avatar.pose` stream, or of a video of a signer, by digest to the **final** caption revisions it translates. A review (scope `translation`) signs it off. Its figures are each segment's delay behind the speech and, for video, a floor: QVGA and 20 fps, from EN 301 549 V4.1.1 clauses 6.5.2 and 6.5.3, which clause 4.2.4 NOTE 3 names. - **Completeness claim (§7).** For each declared captions, description and signing track, the verifier checks: - the track is present; - it is finished: cues resolved, events voiced, cues signed; - it is in the declared language; - it is within the latency and offset bounds the claim states; - its review is present, names this signing, and is not rejected. §7.3 lists the tolerance each text states, and where none does. The claim also reports how much of each description is spoken over captioned dialogue, as a figure, not a refusal. - **§9** maps each figure onto 47 CFR 79.1(j)(2), (j)(3) and (k)(2)(iv)/(vi), 79.3(a)(3), the UK regulator's 2013 and 2024 statements on live subtitling, and EBU Tech 3370 §2.3.2–2.3.3's authoring delay. It also lists what is not computed: NER-style severity grading (which needs a person), capitalisation and edge punctuation, speed and placement. Implementation: - the `access_services` module of `wai-rs`, feature `access_services = ["timed_text", "capture_take"]`, with no new dependency; - `wai_access_services_vectors gen|verify` and `tests/access_services.rs`; - the corpus `live-conformance/access/`; - the second implementation, `live-conformance/access_services_verify.py`; - a CI step in `wai-conformance.yml` plus its `paths:` entry; - the REGISTRY entry and export (`registry/capabilities.json`, `vocab/v2.ttl`, the REGISTRY tables), with 7 `DocumentKind` wire identifiers and the `tag:…/wai/access` scheme; - a SPEC §5 paragraph and the record lists, the CHANGELOG, the live-conformance index and corpus README; - the timed-text §1 scope note now points audio description here. ## Design decisions and why - **JSON records with JCS signatures, not binary.** They follow synthesis and ai-usage, merged on main, using jwp-receipts §2.4.3 encodings and §3.1 step 1's strict reading. The signature is Ed25519 over `wai:access\x01 ‖ JCS(record − sig)`, and the record id is SHA-256(JCS(record)). Signed integers (latencies, offsets) are decimal strings, because the JCS subset holds no negative number. - **A source is the aligner's statement; the media's time is not.** On a take, the capture clock is the device's signed commitment, and the alignment is attributable to the signer and checkable against the master. The latency class makes the difference visible. Who may capture, and for whom, stays a take or room verifier's check. §12 says so. - **First-display accuracy uses a prefix alignment that ends at the greatest best column.** A growing partial then makes no errors for words it has not reached, and a wrong last word is a substitution, not an insertion. Both cases are covered by a unit test. - **A voice must lie within its event,** since DAPT truncates an audio at its Script Event's end (§4.9.1). The writer can then express every voice, and the sync bounds apply inside it. - **Sink-synthesized description is written as `preRecording` with `tta:speak`.** DAPT has a SHOULD NOT for Audio objects in a pre-recording script and for Synthesized Audio in an as-recorded one. §5.5 states this departure and why. - **A signing binds final revisions only** (`sign_revision_not_final`). It translates what the captions resolve to; a superseded partial is not that. - **Records are checked kind by kind:** sources, references, descriptions, signings, reviews, quality, claims. A record another names is checked before it, whatever the order given. - **Synthesis coordination, read, not edited.** The voice and signer origins are stated as synthesis §3.1's class `synthesized`/`generate`/`speech-synthesis`, with oversight `fully_autonomous` or `human_validated`. The review reuses synthesis §9.1's vocabulary with scope `translation`. ## Checks (all cargo through `rb` on the build host) | check | command | result | |---|---|---| | unit tests | `cargo test --locked --no-default-features --features access_services --lib access_services::` | 6 passed | | corpus test | `cargo test --locked --no-default-features --features access_services --test access_services` | 4 passed (regenerates byte for byte, no other file, every outcome, every §10 reason refuses a case, §10 equals `Reason::ALL`) | | timed-text (text edited) | `cargo test --locked --no-default-features --features timed_text --test timed_text` | 4 passed | | registry | `cargo test --locked --no-default-features --lib codecs::registry` | 14 passed | | registry export | `cargo run --no-default-features --bin wai -- registry export` then `registry export --check` | regenerated, then "up to date" | | reference verify | `wai_access_services_vectors verify ../live-conformance/access` | PASS, PASS | | second implementation | `python3 -I live-conformance/access_services_verify.py live-conformance/access` | 38 vectors and 4 documents PASS; 125 of 125 mutated copies fail | | timed-text verifier | `python3 -I live-conformance/timed_text_verify.py live-conformance/captions` | PASS | | clippy | `cargo clippy --locked --no-default-features --features access_services --all-targets` | no notes in `access_services*` or `codecs/registry.rs` (two fixed: `is_multiple_of`, a type alias); the pre-existing notes elsewhere are unchanged | | lockfile | `cargo metadata --locked --offline --format-version 1` | ok, no dependency added | | repo checks | `check-tests-in-ci.py`, `check-workflow-concurrency.sh`, `check-spec-refs.py`, `check-wai-section-refs.py`, `check-no-vanity-metrics.sh`, `test-public-text-checks.sh`, `check-comparative-prose.sh`, `check-orphan-sources.py` | all exit 0. The section-ref checker first flagged a `[begin, …)` interval before a link, so the intervals were reworded. | | names | `check-public-names.sh` over the 15 touched files with the private list | clean | ## Falsifications (committed first, each restored after) | defect introduced | detector | result | |---|---|---| | Rust: drop the 35 ms early bound | `--test access_services` | exit 101: `description_early` accepted where refusal is stated | | Rust: prefix alignment ends at the least best column | `--lib` and `--test access_services` | exit 101: unit test, and the lesson's figures differ | | Python: drop `sign_revision_not_final` | `access_services_verify.py` | exit 1: `sign_partial` reaches `track_incomplete` | | Python: drop the early bound | `access_services_verify.py` | exit 1: `description_early` accepted | | 125 mutated corpus copies (stated figures, flipped record signatures, altered events, moved places and reasons, a byte in each DAPT file) | `access_services_verify.py` | every copy fails | ## Regulatory and standards facts, each read at its source - EAA Annex I §IV(b)(ii) and Art. 31(2): read from a reproduction of the adopted text, because the Official Journal could not be fetched; the extension says so. - EN 301 549 V4.1.1 (2026-09): the ETSI PDF. - WCAG 2.2: W3C. - 47 CFR 79.1 and 79.3: eCFR as of 2026-10-01. - DAPT: the CRD of 2026-06-26. - EBU Tech 3370 v1.0. - The UK regulator's statements of 16 Oct 2013 and 15 Apr 2024: archived PDFs, since the regulator's site refuses fetches. - Two facts from the sweep were corrected at source: - EN 301 549 V4.1.1 says "subtitles", not "captions"; - the 2024 statement's 4.5 s is a mean across live programming, not a per-programme limit. ## Not done - Spoken subtitles (EN 301 549 7.1.5), extended audio description, 79.103(e) previewability, and component-level `access` attributes with HLS `CHARACTERISTICS` and DASH Accessibility mappings: R7's INGEST part, left out of this slice. - The objects a signing or a voicing names are digests only. The corpus has no pose stream, video or audio, and the figures do not decode media. - NER severity grading needs a person, so it is not computed. The figure gives the error counts a grader starts from. - Sync of a sign track has no stated standard tolerance; the claim states its own (§7.3). 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 4 commits 2026-10-21 12:30:57 -04:00

wai.access.services: caption quality from the signed revisions, audio description, sign language and a completeness claim ... 00bbc69536

The extension (extensions/access-services.md), the access_services module of wai-rs
behind its own feature, and the corpus live-conformance/access/ written from fixed seeds
by wai_access_services_vectors: three accepted sessions and a negative control for every
reason of §10, and the DAPT document of each description.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Merge remote-tracking branch 'origin/main' into wai/access-services a8b60cff57

access_services_verify.py: the access-services corpus's second implementation ... 6e747e412a

Written from the extension's text: records read strictly and their JCS signatures, words
and their alignment, the caption figures, description rules and DAPT documents (parsed with
xml.etree), signing, reviews and claims; every mutated copy of the corpus must fail. The
extension now links synthesis, merged on main, for the origin of a synthesized voice or
signer and for its review vocabulary.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

wai.access.services: registry entry and export, SPEC §5, CHANGELOG, CI step, corpus README; timed-text's scope points to it ... 4dacdb2038

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot referenced this pull request 2026-10-21 12:56:12 -04:00

dash_read_bounds: time the short and long runs in turn, the least of seven each #400

dcharlot added 1 commit 2026-10-21 14:41:05 -04:00

Merge origin/main into wai/access-services (IETF draft, timing-test fix) ... 6ceff94397

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot referenced this pull request 2026-10-21 20:37:41 -04:00

room e2e: the window's play wait states where the seek back stood, and has the heavy bound #423

dcharlot added 3 commits 2026-10-21 21:38:22 -04:00

Merge origin/main into wai/access-services (settlement, contribution, shared lists) ... 9782a2b6dd

Co-Authored-By: Claude Opus 5.5 <[email protected]>

registry export: regenerated after merging main ... e3d105b66b

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Merge origin/main into wai/access-services (room e2e window wait) ... 5b25cdda59

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot merged commit 3e9848b878 into main 2026-10-21 21:38:39 -04:00

dcharlot referenced this pull request from a commit 2026-10-21 21:38:41 -04:00

Merge pull request 'wai.access.services: caption quality anyone can recompute, audio description, sign language and a completeness claim (ACC)' (#396) from wai/access-services into main

dcharlot referenced this pull request 2026-10-22 03:14:15 -04:00

wai.media.upscale: super-resolution as an exact derivation, residuals against it, per-title priors, and an account of sink-made pictures (C3) #408

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!396

No description provided.