| Filename | Latest commit message | Latest commit date |
|---|---|---|
|
Some checks failed
wai cross-architecture byte-equality / windows-x86_64 (pull_request) Successful in 30s
wai cross-architecture byte-equality / linux-aarch64 (pull_request) Successful in 42s
joule-code conformance gate / Build + round-trip C1-C4 vectors (pull_request) Failing after 1m3s
WAI conformance / cargo-class corpora · linux-arm64 (pull_request) Successful in 1m1s
WAI conformance / cargo-class corpora · windows (pull_request) Successful in 1m21s
jouleclaw public surface must stay documented / Public-surface audit (jouleclaw-rs) (pull_request) Successful in 58s
WAI conformance / cross-arch byte-equality · linux-arm64 (pull_request) Successful in 20s
WAI conformance / cross-arch byte-equality · windows (pull_request) Successful in 22s
wai cross-architecture byte-equality / linux-x86_64 (pull_request) Successful in 14s
WAI conformance / cargo-class corpora · linux (pull_request) Successful in 29s
WAI conformance / cross-arch byte-equality · linux (pull_request) Successful in 15s
wai cross-architecture byte-equality / macos-arm64 (reference) (pull_request) Successful in 55s
WAI neural decoders end-to-end / synthetic fixtures → Rust decode → reference (±4) (pull_request) Failing after 1m5s
WAI conformance / cargo-class corpora · macos (pull_request) Successful in 13m25s
WAI conformance / cross-arch byte-equality · macos (pull_request) Successful in 3m42s
Toolchain 1.97.1 -> 1.98.0 (released 2026-08-18) in both `rust-toolchain.toml`
files, 33 pins across 5 CI workflows, and 26 workspace `rust-version`
declarations. Nothing pinned to 1.97 remains.
Dependencies refreshed in all 25 workspaces: 1,751 changes, every one
semver-compatible. The vendored `jouledesk-rs/vendor/libwebrtc` tree is
deliberately excluded — it is a patched vendor drop, not a managed dependency.
Note the `1.97` in `jouleclaw-loader-gguf/BENCHMARKS.md` is a benchmark timing
("1.97 ms"), not a version; a blind find-and-replace would have corrupted a
results table, so pins were matched by field rather than by string.
Also completes a cross-workspace regression this repo has been carrying. When
`budget_funds` and `caveats` were added to `jcp_schema::Grant`, the workspaces
that path-depend on `jcp-schema` were fixed one wave at a time, and three were
missed: `joule-ui`, `jouleclaw` and `mesh` still built `Grant` without the new
fields and had been failing to compile since. `cargo test --workspace` inside
`jcp-rs` never builds them, which is exactly why it went unseen; checking every
workspace independently is what surfaced it. Two of the six sites used field-init
shorthand (`budget_uj,` with no colon), so they needed a separate pass — a
reminder that a regex over initializers is not a substitute for compiling.
Verified: 21 of 25 workspaces `cargo check --workspace --all-targets` clean.
Not fixed here, and NOT caused by this change — each verified against the
pristine tree rather than assumed:
- `eoc/eoc-rs`: `RequestBuilder::query` is unavailable because reqwest is taken
with `default-features = false`, which drops the feature providing it.
reqwest is 0.13.4 before and after, and `serde_urlencoded` is absent from the
lock in both, so this predates the refresh. Fixing it is a choice between
enabling the feature and changing the call sites.
- `openpay`: `Hmac::new_from_slice` unresolved. hmac 0.12.1, digest 0.10.7 and
sha2 0.10.9 are identical before and after — same versions in, same result.
- `smart-byte/smart-byte-rs`: never resolved on main. Its manifest asked for two
features that do not exist (`rand_core`'s `alloc`, which no version has ever
had, and `cbc`'s `std`, where 0.2 offers `alloc`); both are corrected here,
which advances it from "will not resolve" to "will not compile" and exposes
the real problem — the manifest wants `rand = "0.10"` while the lock pinned
`rand 0.8.8` and 38 files still use the 0.8-era API (`thread_rng`,
`rand::rngs::OsRng`). That is a two-major-version source migration and wants
its own change, not a side effect of a toolchain bump.
`wai/wai-rs` needs libjxl present to build its default feature set; CI installs
`libjxl-dev`, so this is a local environment gap rather than a repo failure.
|
||
| .. | ||
| aac-rs | ||
| conformance/v1 | ||
| spec | ||
| vocab | ||
| LICENSE | ||
| LICENSE-SPEC | ||
| README.md | ||
| SPEC.md | ||
AAC
A carrier for voice. A communication board is a signed, portable file — every symbol, every language, every access method — that speaks on any device, offline, owned by no one.
AAC is the communication cargo class of the Transaction Science substrate. Where Smart Byte carries value, EOC carries compute, Map carries place, and Joule Code carries coding work, AAC carries utterance — what a non-speaking person means to say — as a content-addressed, portable board with its access model and its energy cost intrinsic.
Augmentative and alternative communication is the assistive technology that gives a voice to people who cannot rely on speech: children and adults with autism, cerebral palsy, ALS, stroke, apraxia, Rett syndrome, and many other conditions. A board is a grid (or a visual scene, or a keyboard) of cells; selecting cells composes an utterance; the device speaks it. The selecting can be done by touch, by a single switch scanned across the grid, by where the eyes rest, by a head pointer, by a joystick — whatever motor channel a person has.
A board in this standard is not an app's private save file and not a vendor API response. It carries exactly what a voice needs to survive moving between devices: identity (a content address over its canonical serialization), a layout (cells placed once and locked, so muscle memory is never invalidated), symbols (referenced by source and stable id, not baked-in pixels), an access model (the method, timing, and calibration the user relies on), language (a graph edge, so a bilingual user code-switches by traversal), and a two-part joule cost on every operation — for transparency, never billing. Communication resolves through an explicit energy-tiered cascade: the overwhelming majority of utterances are cache hits and on-device lookups that never touch a network, and richer help (context-aware prediction, sentence construction, voice cloning) is reached only when something cheaper cannot close the request.
The substrate is intentionally ownerless: no protocol-level token, permissive licensing on the spec text and reference code, and conformance vectors so any implementer can verify canonicity without a trust relationship with the authors. A conforming client opens a board, renders it, drives it with the user's access method, predicts on-device, and speaks — against local storage, with no contact with any company's servers as a condition of operation. Transaction Science is one steward: it publishes the spec, ships the reference implementation, and operates optional commercial services (hosted symbol libraries, voice banking, team coordination) around the substrate — not its proprietor.
The doctrine
A voice is a file, not a subscription. The durable layer of a communication system is not the app and not a recurring charge — a person's ability to speak must not lapse when a payment does. It is the portable, self-describing board that renders and speaks forever from dumb storage, on any device the user already owns.
Seven principles, normative in SPEC.md:
- Free forever. No billing, no tiers, no balance, no in-app purchase. Every cascade level is free at the protocol layer. A conforming implementation never gates communication behind payment.
- Offline first. Levels L0–L2 — cache, symbol lookup, local prediction — run entirely on-device and carry 95%+ of communication. A network drop never silences a user.
- Motor stability. Cells are placed once and locked. Symbols do not move; layouts do not reflow. The user's motor plan is sacred — a board that rearranges itself has destroyed the thing that made it usable.
- Every access method. Touch, switch scanning (1- and 2-switch, row-column, group, linear), dwell/eye-gaze, head tracking, joystick, partial voice. No method is an afterthought; each is a first-class input in the same model.
- Every language. Multilingual is a graph edge, not a separate app. Code-switching is traversal. RTL, CJK, Indic, and Arabic scripts are first-class.
- Open interchange. A board exports to and ingests from the open interchange format. A user brings their existing boards in and takes everything with them when they leave. No lock-in — ever.
- The receipt. Every operation reports a two-part energy cost in microjoules with an honest provenance tag. The cost is shown for transparency and is never converted into a charge.
The cascade
AAC's organizing primitive is the energy-tiered cascade — ten levels, each an order of magnitude costlier than the last, each reached only when nothing cheaper closes the request:
| Level | Name | Typical op | Energy | Offline |
|---|---|---|---|---|
| L0 | Cache hit | recently-used phrase, pinned message, motor-planned sequence | ~1 µJ | ✅ |
| L1 | Symbol lookup | core vocabulary retrieval, symbol-set search | ~10 µJ | ✅ |
| L2 | Local prediction | on-device word/phrase n-gram completion | ~100 µJ | ✅ |
| L3 | Context prediction | time / place / partner / activity-aware prediction | ~1 mJ | — |
| L4 | Semantic expansion | novel sentence construction from partial input | ~10 mJ | — |
| L5 | Conversation AI | conversation repair, social scripts, pragmatics | ~100 mJ | — |
| L6 | Generative communication | letters, posts, longer narratives | ~1 J | — |
| L7 | Voice clone | personal-voice synthesis from recordings | ~5 J | — |
| L8 | Multimodal generation | custom symbol / visual-scene synthesis | ~10 J | — |
| L9 | Adaptive model | retrain the user's prediction model | ~15 J | — |
The boundary at L2/L3 is normative: exactly L0–L2 are offline-capable, and a conforming client serves them with no network. Everything above is enrichment, never a dependency.
Open interchange
A user's years of customization — added words, custom symbols, personal phrases, the exact layout their muscle memory has learned — are the most valuable and least portable thing in assistive technology. AAC makes interchange a property of the standard: a board serializes to a canonical, self-describing form, and the reference engine ingests the open board interchange format and other community exports so a user never starts over when moving in, and never loses anything when moving on. Ingestion is one-directional and additive — the standard reads existing work into the substrate; it does not compete with the tools that produced it.
Contents
SPEC.md— the normative specification (RFC 2119): the board / cell / symbol data model, the access-method abstraction, canonical serialization, the cascade and its offline invariant, the interchange format, and the receipt.spec/— sub-specifications: the board data model, the motor-stability invariant, the access model, the communication cascade, interchange, and the receipt.vocab/— the controlled vocabulary (LEXICON.md) and stable IRIs: the OWL ontology (v1.ttl), the JSON-LD 1.1 context (v1.jsonld), and SHACL shapes (v1.shapes.ttl). Namespacehttps://aac.transaction.science/vocab/v1#.conformance/— published vectors. An implementation is conformant if it reproduces them.aac-rs/— the Rust reference:aac-core(the engine — boards, symbols, vocabulary, motor planning, scanning, dwell, prediction, i18n, interchange),aac-web(a reference communication app built on Leptos), andaac-conformance(the harness).
Reference implementation and demonstrator
The Rust reference lives in aac-rs/. aac-core is the engine a third party implements against — platform-agnostic, dependency-light, no UI. The demonstrator is aac.transaction.science: a free communication app (Rust → WASM, offline-first PWA) that consumes aac-core and renders grid boards, switch scanning, dwell selection, gamepad and eye-tracking input, on-device prediction, and system / banked-voice speech. The demonstrator is a demonstration of the standard — free forever, every access method — not a product to be sold.
Status
v0.1.0 draft. Reference engine ported and building: aac-core (22 modules — boards, symbols, vocabulary, motor planning, switch scanning, dwell, prediction, context, history, multilingual, accessibility, partner tools, analytics, the cascade router, the store, interchange import/export), aac-web (the Leptos reference app), and the aac-conformance harness (cascade routing, energy monotonicity, the offline invariant, and wire round-trips). cargo test and cargo clippy --all-targets are green on the engine and the harness. The reference app builds end-to-end through cargo leptos build --release — both the WASM client (aac-web.wasm, ~3 MB) and the SSR server binary — and boots: it serves the app shell and the offline-first PWA assets. Specification core published in SPEC.md and spec/, with the controlled vocabulary and stable IRIs in vocab/. Not a medical device; not certified as a speech-generating device (SGD) for insurance funding (see SPEC §11).
Licence
- Reference code (
aac-rs/): Apache-2.0 — seeLICENSE. - Specification text (
SPEC.md,spec/,vocab/): CC-BY-4.0 — seeLICENSE-SPEC.