Every Model Call Through the Gate: A Governance-Mediated Voice Pipeline with Chunk-Level Fidelity Verification — Systems Paper — mmv-voice: Local Whisper Transcription, MMV-Governed Formatting, Speaker Attribution with Graceful Degradation, and a Pre-Registered Fidelity Evaluation

mmv-voice is an open-source (AGPL-3.0) pipeline that turns an audio recording into verified structured text — speaker-attributed turns, cleaned prose, meeting minutes, and a digest for a downstream secretary system — under one architectural rule: every language-model call passes through the frozen MMV governance stack; no raw model API is ever invoked. The pipeline is local-first end to end by default: Whisper large-v3-turbo transcribes on local GPU with automatic language identification (~100 languages, with language-conditioned downstream prompting); speaker attribution uses audio diarization when the optional dependency is available and degrades gracefully to governed text-based attribution when it is not; formatting (filler removal, punctuation, paragraphing) runs through the MMV Medium release binding (a 12B-class local model under the frozen route-transformer / post-validator stack); an optional cloud tier (120B-class) exists behind an explicit per-session consent dialog, because selecting it sends transcript text off-device — the default is always local. The system's distinctive component is chunk-level fidelity verification: because any generative "cleanup" of a transcript can silently drop, add, summarize, or shift meaning, every formatted chunk is checked (again through the governed stack) against its source for meaning-content preservation, suspicious chunks are flagged inline rather than silently accepted, and a verification report ships with the output. Fidelity is thus treated not as an assumed property of formatting but as a per-chunk, inspectable claim. We describe the pipeline, the engine-switching and consent design, and a pre-registered evaluation protocol now awaiting its human ground-truth labeling: character-error- rate scoring of raw and formatted text against hand-labeled reference segments — separating transcription fidelity from formatting fidelity, the quantity the verifier is supposed to protect — plus time-overlap speaker- attribution accuracy with symbol-mapping, with acceptance thresholds (formatted CER within 2 points of raw; attribution ≥ 0.85 practical, < 0.7 remediation) published before any number exists. No fidelity results are reported here, deliberately: the protocol's point is that when the numbers arrive, the bar they must clear was already on record. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.