Parity — what is witnessed, and where

Two columns, because they are different claims. Witnessed locally means a check in this repo runs and passes; the check is named, so a green row can always be re-run. Witnessed running means the same claim has been watched holding with a person talking to the line.

Every row in the second column reads not yet. Nothing in this repository has been run as a voice call: the graph names four extensions that do not exist, and no audio has been through it. The rows below describe a design and the checks that hold the design’s own files to each other — which is a real thing to check, and is not the same thing as a working line.

That distinction is the whole point of the ledger. data-agent-service spent a day discovering that a runbook naming three parameters which did not exist reads as deployable, and the fix was not a better runbook but a check that compares the runbook to the definition. These rows are that check’s equivalent here, and they stop at exactly the line they can prove.

The design, as configuration

Capability	Witnessed locally	Check	Witnessed running
Every addon the graph names is declared, and nothing installs at run time that was not pinned	🟢	make test	not yet
The host is the only node that reaches the model — there is no second path to an answer	🟢	make test	not yet
mcp_client_python is absent, so run_query can never reach the conversational model	🟢	make test	not yet
Barge-in reaches the model, the voice and the wire, not one of the three	🟢 structurally	make test	not yet — an interrupt is a timing claim and only a call can make it
A backend is a descriptor: every configured backend has one, and each declares its dispatch and its fast tools	🟢	make test	not yet
No general layer carries the first backend’s vocabulary	🟢	make test	not yet
Every setting the graph reads, and every switch the plan lists, is in the template	🟢	make test	not yet
The TEN pin is one version across the template, the compose file and the Dockerfile	🟢	make test	not yet
The upstream stack is reachable and its ask service is healthy before anything starts	🟢	make doctor	not yet

The line itself

Capability	Witnessed locally	Check	Witnessed running
The image builds on amd64	🟢 in CI, every push	the image (amd64) job — 904 MB, 326 s	not yet
The image builds on arm64 from the release assets, and faster-whisper installs there	🟢 in CI, on a native arm runner — ctranslate2 4.8.1 has an aarch64 wheel	the image (arm64) job — 986 MB, 213 s	not yet
The server starts and registers the graph, on both architectures	🟢 in CI	the smoke step: /health answers and /graphs lists analyst_line	not yet
A backend descriptor loads, interpolates its settings, and every way it can be malformed is refused at start-up rather than at run time	🟢	make test — 15 checks over das_tools/descriptor.py	not yet
A pre-rendered phrase and a synthesized one are one code path to everything downstream	🟢 by construction	local_tts serves both through get(); no test reaches it without the runtime	not yet
Everything the caller hears leaves through one path, so there is never a second speaker with no arbiter	🟢 structurally	make test — one tts_text_input send in das_host	not yet
A refusal, an abstention and an error each have their own fixed phrase, and the refusal does not sound like missing data	🟢	make test — 8 checks over the host’s policy	not yet
Speech reaches the graph, is recognised, and a final transcript reaches the host	🟢 run — 2.7 s of synthesized speech in, is_final transcripts out, on arm64	make up, then feed PCM to the WebSocket	not yet
Every extension in the graph loads and the session stays up	🟢 run — all seven nodes, /list shows the session alive	make up and POST /start	not yet
The host takes the turn and calls the model	🟠 partly — the call is made and fails against the llm-stub, which answers plain JSON where the SDK streams. The stub cannot stand in for a model here	needs ANTHROPIC_API_KEY	not yet
A person speaks and hears an answer	🔴 not run — everything up to the model is proven; TTS has never been reached	make up && make call with a key	not yet
The panel’s arithmetic is right: p95 by nearest rank, a span needs both marks, the first mark wins, and an unmeasured phase reads as nothing rather than zero	🟢	make test — 11 checks run through node	not yet
The panel stores nothing — no transcript, no question, no audio	🟢	make test — the page uses no storage API	not yet
A mis-transcribed entity produces a confirmation, not a dispatch	🟢 the decision; 🔴 the call	make test — 15 checks over the uncertainty rules	not yet
First audio under 900 ms at p95	🔴 not run — the panel can measure it; nothing has been measured	make up, then a call	not yet
A definitional question is answered without dispatching	🔴 not run	the tiers witness	not yet
A refusal is spoken from a fixed phrase and never paraphrased	🔴 not run	the refusal witness	not yet
Semantic turn detection beats fixed silence	🔴 not run — needs a GPU this machine does not have	switch #1, off vs on	not yet

Upstream, for reference

The ask contract this repo consumes is witnessed in data-agent-service: 24/24 direct and 24/24 through the gateway, with the model stubbed. Its four behaviour checks have not run. See that repository’s docs/parity.md.