NarraTwin AI
“Can project knowledge become a walkthrough without inventing a claim?”
It is shown anyway, because the machinery is the point, and that machinery is proved against a harness, not a model: no real language model is wired in yet, so every guarantee here is about the system around one. The grounding check reads both directions and fails closed — an answer with no checkable claims fails rather than passing empty. Consent is bound by checksum — a fingerprint of the exact evaluation it covers, and spends itself once used. And nothing in the pipeline, the model included, has the authority to mark its own output publishable; that denial is written as code, not policy.
GateIts own readiness review says No-Go, so it is not deployed

SAMPLE — interface design, not a running capture. No screenshot of the built system exists yet, so this stands in and is labelled rather than implied.
- Not claimed
- Deployment · video · avatar Q&A
What it does, and what it refuses
Works now
- Twenty-five languages proved to agree across five surfaces against a pinned fixture
- Consent is checked, and external web is disabled by policy, before any presenter is generated
- The Google TTS path re-verifies the peer address before the TLS wrap, closing a DNS-rebinding window
Known gaps
- Its own release-readiness review reads No-Go, so it is not deployed
- The evaluation report carries no tested commit SHA, so its 41/41 pass cannot be tied to any code state — the committed harness can re-produce it at any pinned commit
- The eval-smoke result has never been committed; only the harness that produces it is in the repo
- Its architecture document still opens “blocked until Stage 4 gate approval”, and undersells the code
Not claimed
- Deployment
- video
- avatar Q&A
In the Google TTS path, budget is reserved before anything leaves the machine, replays are fingerprinted, and a failure landing where bytes may already have gone becomes BILLABLE_UNKNOWN: refused on retry rather than risk billing twice. The ElevenLabs path in the same file does not do this yet; it retries a timeout straight through.
Discussion
Comments need JavaScript and a GitHub account — the same account system the site's own code already lives in. Without either, read or start the thread directly on GitHub.

