Fixture-backed evidence path

Demo spine

This page shows the current Voice Bench SA pipeline using safe fixture data only. It is a working shape for consent, voice-note intake, transcript correction, independent review, and export evidence.

01

Consent

Every task starts with contact provenance, clear permission, and a stop/delete path before any voice data is requested.

02

Voice note

The fake fixture accepts two WhatsApp-style audio submissions and rejects unsafe or incomplete media before export.

03

Transcript

Audio is paired with transcript candidates and participant corrections so local names, speech, and phrasing can be checked.

04

Review

A separate reviewer validates the correction path before any row can enter the export-ready batch.

05

Export

The generated manifest has two rows and points at the planned Google Cloud Storage object, with no live cloud write yet.

Export manifest

Two fixture rows are ready for the first Google Cloud smoke upload

Manifest object
manifests/smoke/offline-demo-export-manifest-v0.jsonl
Rows
2 accepted audio submissions
Audio outputs
audio/smoke/audio_intake_accept_voice_note_ogg.ogg
audio/smoke/audio_intake_accept_voice_media_webm.webm
Task
voice_simple_greeting_v0

Export gates

The demo refuses unsafe or unsupported claims

  • Fixture data only; no live participant audio is shown.
  • Source audio extensions are preserved for OGG and WEBM.
  • Cloud upload is planned, not performed by this page.
  • Gemini, user, and revenue evidence remain separate gates.
  • Local proof command: npm run xprize:local-proof:check.

Gemini QA contract

The model check is specified before live provider calls

Contract
gemini_transcript_qa_v0
Model target
gemini-3.5-flash
Readiness command
npm run gemini:qa:check
Current mode
Self-test only; no live Gemini call yet.

Current status

Honest by default

Email Voice Bench