Transaction-Science/open-standards
Watch
1
0
Fork
You've already forked open-standards
 0
Code Issues Pull requests Projects Releases 15 Packages Wiki Activity Actions
open-standards/aac/aac-rs
History
Exact
Repository files (latest commit first)
Filename Latest commit message Latest commit date
Transaction Science 48eaa9a587 ci: every Python install pinned by SHA-256; verifiers run with -I 
Six workflow steps installed whatever PyPI served that day: cryptography
(wai-conformance, wai-http twice, wai-web-wasm), the RDF tools' dependencies
(wai-conformance), PyYAML (standards-build, aac-rs), the ML stack of
wai-neural-e2e, and pip itself (wai-neural-e2e, wai-sota-vectors). A
compromised release would have run on the runners. Each now installs a
hashed requirements file with --require-hashes, the way wai-pack.yml
already does:

- cryptography: wai/pack-conformance/requirements-verify.txt (the same file);
- RDF tools: wai/vocab/requirements-shapes.txt, from requirements-shapes.in;
- PyYAML: aac/aac-rs/tools/pagesets/requirements-ci.txt, from requirements-ci.in;
- wai-neural-e2e: wai/tools/requirements-neural-e2e.txt, from
  requirements-neural-e2e.in;
- wai-sota-vectors: wai/tools/requirements-sota-int-hashes.txt, from
  requirements-sota-int.txt.

Each generated file's header holds the uv command that regenerates it. The
two ML locks take torch and torchvision from the CPU wheel index through
wai/tools/torch-cpu-overrides.txt (2.14.0+cpu on Linux, 2.14.0 on macOS),
with no CUDA packages. The new lock files are added to the path filters of
the workflows that install them.

pillow goes from 12.3.0 to 13.0.0 in requirements-sota-int.txt: 12.x has
heap overflows in the TIFF writer and the JPEG decoder and an argument
injection into Ghostscript (GHSA-332p-7hrc-92wf, GHSA-8g5g-8vmr-9782,
GHSA-w889-685w-c8fh); 13.0.0 is the first fixed release.

The 26 verifier and differential calls that ran without -I now run with it,
so neither the working directory nor PYTHON* variables can put a module in
front of the standard library: pin, int_payload, staged, http (4), measure,
sim, sim_diff, claim, differential, jwp, chain and chain-gap differentials,
signer (2), check_shapes, smart-byte verify_v1, pack_verify and the five pack
differentials (7). Each script already puts its own directory on sys.path.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
2026-10-22 13:56:02 -04:00
..
crates Merge aac/sw-webkit-install: the app's service worker is kept installed in WebKit too (bug 215122), and never hangs installing 2026-10-18 14:23:54 -04:00
data fix(pagesets): the phrase sets' language note says "two independent AI models agreed" 2026-10-15 21:31:47 -04:00
third_party Merge origin/main into aac/voice-quality: bar scanning, First letters, the hotfix 2026-10-04 20:15:01 -04:00
tools ci: every Python install pinned by SHA-256; verifiers run with -I 2026-10-22 13:56:02 -04:00
Cargo.lock sec: close the advisory backlog across every workspace, and move the WASM sandbox off a runtime with three critical escapes 2026-10-22 09:44:59 -04:00
Cargo.toml fix(access-core): a clock that works in browsers, and a scan grid that follows the board 2026-09-28 16:25:49 -04:00
prerender-static.sh aac-web sw: an install waits its turn, keeps what it fetched, and downloads a file once 2026-10-18 12:14:00 -04:00
README.md aac: the reference app is the AAC Reference App, no longer "Express" 2026-10-10 22:42:12 -04:00

README.md

aac-rs — the AAC reference implementation

Rust reference for the AAC open standard. Apache-2.0.

Crate What it is
access-core The shared access kernel: access methods, switch scanning with its response calibration (SPEC §5.2), dwell with two-way proposals (§5.3), gaze estimation, voice. Four dependencies (serde, serde_json, thiserror, web-time).
aac-core The engine: boards, symbols, vocabulary, motor planning, prediction, i18n with each language's core status (§4.1), the cascade, interchange.
aac-web The AAC Reference App, the communication app of the standard (Leptos → WASM, offline-first PWA). Live at aac.transaction.science.
aac-conformance The harness: C1–C7 against the published vectors in ../conformance/v1/.

Test and conform

cargo test --workspace
cargo run -p aac-conformance --bin aac-conformance     # C1–C7; any FAIL exits non-zero
cargo run -p aac-conformance --bin gen-vectors -- ../conformance/v1   # regenerate C5/C6 vectors

Build the app as a static site

bash prerender-static.sh        # → target/site/, ready to rsync; deploy with no_cache_bust

The script builds with the wasm-bindgen CLI version the crate pins (installing a private copy if the machine's differs), prerenders the SSR shell, content-hashes pkg/, and writes a service worker that precaches the shell and every file it loads, so the installed app opens with no network.

Four optional downloads ship beside the app, each built outside git and checked against a catalogue committed here. Each is offered to the person, never fetched by itself, and runs on their device:

  • Natural voices (tools/voices/): neural text-to-speech, offered first for the person's language (in setup, and when the language changes). Of the device's own voices the best speaks (premium, enhanced, then basic), the one the person chose while the device has it, and robotic or novelty voices only when nothing else speaks the language; online voices never (access-core voice.rs, aac-web voices.rs; device-voice facts from Readium Speech, third_party/readium-speech/).
  • The sentence helper (tools/ai/): an on-device language model that turns a few chosen words into whole sentences, and suggests replies. It runs on the graphics chip only, and only in the languages it was tested in.
  • Hearing your partner (tools/asr/): on-device speech recognition that writes down what someone says to the person, for the sentence helper's replies. It listens only when asked, keeps and sends nothing, runs on the graphics chip only, and only in the languages with evidence it is accurate.
  • First letters (tools/ae/): abbreviation expansion. The person types only the first letter of each word (iwtgh), and a small on-device model offers whole sentences (I want to go home), fitting what their partner said. It runs on the listener's runtime, on the graphics chip only, never speaks by itself, and is offered only where we measured it (English).

Set AAC_VOICES=none, AAC_AI=none, AAC_ASR=none or AAC_AE=none to build without them.

Host it under your own site

The AAC Reference App reads four build-time variables (set them for both the cargo-leptos build and the prerender):

Variable Default Use
AAC_BASE_PATH "" Mount path, e.g. /app. Also set LEPTOS_SITE_PKG_DIR=app/pkg for the prerender.
AAC_SW_URL /sw.js The service worker to register (a host usually runs one for its whole site).
AAC_APP_TITLE AAC — reference app The document title.
AAC_HIDE_ENERGY unset Set to hide the energy meter.

aac.aurasense.ai builds it this way at /app (build-app.sh).

Two rules the app keeps

  • The first frame is the same on the server and in the browser. Saved state is restored only after hydration, behind a brief starting screen. Reading it earlier made a returning user's first frame differ from the prerendered page, and hydration aborted.
  • Timing belongs to the person (SPEC §5.1). Settings → Timing sets scan speed and dwell, shows proposals from the person's own responses to accept or decline, and adjusts automatically only if they turn that on.