M4 Max voice refresh evidence

Published from retained artifacts on 2026-09-12. Runtime/deployment claims are historical September 8 observations. No live services or model choices changed during publication, and the earlier runtime/harness edits are not included. The capture's dirty source and unrecorded evidence-author model limit historical reproducibility and reviewer-independence claims.

This is the sanitized public evidence bundle for the dated finding. It records a local, same-host Apple Silicon voice-lane review, not the reference Mini-to-Dark topology. Private host names, paths, service identities, UIDs, PIDs, audio, base64 payloads, and raw SDK event streams were removed. Model snapshot paths are represented as <model-cache>.

Campaign boundary

• Campaign ID: 2026-09-08-m4-max-voice-refresh
• Capability: local voice LLM, STT/TTS round trip, and Realtime protocol
• Repository revision: b85b5d27bc8da84216bcc7f5b5704dd340a884a6 plus later harness-fence work; bounded candidate sweep and restoration complete.
• Measured hardware: Apple M4 Max laptop with 48 GB unified memory.
• Evidence labels: functional, bounded diagnostic quality, failed capacity, and incomplete compatibility-only Qwen3.8 evidence.
• Decision: no-promotion for the LLM lane; the existing Qwen3 4B LLM configuration is unchanged, while Kokoro FastAPI 0.8.2 was separately deployed.

Common campaign artifacts

• artifact-manifest.json — ten-role ledger and hashes
• source-registry.json — dated source provenance
• summary.json — bounded machine-readable decision
• friction-log.md — retained failures and durable fixes
• restoration.json — verified final restoration record
• publication-summary.md — derivative claim ledger

Raw run evidence

The native/ directory retains sanitized original JSON schemas, including all four preflights, three spoken suites, diagnostic quality, strict capacity, baseline/audio voice records, SDK session summaries, feasibility, runtime and audio-version identity, artifact verification, thinking controls, candidate startup failure, final deployment/restoration, and the source review. native/redactions.json records every removed private path, host identity, and audio/base64 field.

• native/candidate-preflight.json and native/qwen36-preflight.json retain the failed answers, validators, and timing evidence.
• native/candidate-capacity.json retains all ten strict requests and their canary outcomes.
• native/final-restoration-receipt.json and native/final-deployment-receipt.json retain the verified LLM restoration and accepted Kokoro 0.8.2 deployment without private digests.
• sanitized-preflight.json, sanitized-quality-and-capacity.json, and sanitized-voice-summary.json are explicitly labeled derivative navigation summaries.

Decision and publication

The 9B candidate passed six functional preflight groups but did not pass the strict spoken, patch-format, or controlled-output gates. Qwen3.8 preliminary preflight passed 5/6 groups but shared-prefix tools passed only 1/3 and it was too slow for the voice goal; it did not advance to capacity. Qwen3.6-35B-A3B also passed 5/6 preliminary groups but failed strict JSON; spoken and SDK voice work then passed 36/36 strict and one accepted SDK session, but neither result overrides the JSON failure. None of these results authorizes an LLM route, serve, or client-catalog change. The separately tested Kokoro 0.8.2 runtime was deployed after verified endpoint and fresh Realtime checks; it does not promote an LLM candidate.

See the publication summary for claim-safe copy.