Controlled voice regression after freeze release
Proposed procedure — NOT RUN. Freeze released; live validation and acceptance remain open.
Published 2026-10-12 · Dated evidence, not a live status feed
Controlled voice regression after freeze release
Status: proposed validation procedure. NOT RUN. Not authorization to execute.
1. Record the approved launcher and its exact Pi/PC output route. Identify the recovery path and candidate independently.
2. Verify the installed listener source and binary against the accepted manifest. Record hashes and configuration without publishing private infrastructure.
3. Establish the physical microphone and chosen output device. Do not assume numeric device indices are stable.
4. Capture one agreed reference sentence. Retain permitted audio when available, transcript, event identity, writer result and timing.
5. Follow the same turn through response generation, WAV integrity, transfer and actual playback. Record failures before retrying.
6. Repeat with an ordinary question, an ambiguous recognition case, annotation-only non-speech and a legitimate repetition control.
7. Assess factual support separately from encouragement, voice quality and delivery. Do not rely on another model's approval as proof.
8. In a separately authorized bounded test, verify writer-failure propagation and cleanup. Successful fixtures do not establish full-listener behavior.
9. Record observed outcomes and acceptance scope. Keep independent playback, array firmware, endurance, restart and restore qualification separate.
10. Accept or reject the change, retain evidence and lock the proven configuration before proceeding.