Skip to content

Local publication summary

Local documentation is task-authorized. No external posting, publishing or push was performed. Canonical planned URL: https://fakoli.github.io/anvil-serving/findings/2026-10-06-m4-max-voice-quality-qualification/

Short public copy

Apple M4 Max voice tests found no qualified replacement. Critical STT, TTS human/client gaps and strict 8K failures remain. Exact configs: https://fakoli.github.io/anvil-serving/findings/2026-10-06-m4-max-voice-quality-qualification/

Claim ledger

Claim Evidence Limit
Human STT WER 3.064/2.786/3.343% Derived tables, three canonical native projections, graph-data hashes 24 distinct cases × 3; critical failures; paired uncertainty overlaps zero
TTS 60/60 control/Breeze q8; 59/60 BF16; 56/60 Fish Five full-attempt native artifacts and final TTS assessment ASR proxy; no human/client winner; partial cells not ranked
LLM semantic 132/131/138/144/144 of 144 Every semantic attempt and independent audits; derived tables Different configurations/budgets; historical exact and capacity gates remain
All completed strict 8K arms have 0/3 eligible responses Native capacity rows and independent audits Actual prompts, terminal receipt and capture/canary limits; no speed/useful-context claim
Cache zero disables free-buffer reuse Exact source identities/worker traces and official MLX cache documentation No active-memory hard cap; private LM fixture unshipped; no isolated causal proof
Bounded restoration passed Final audit, starting/ending identity and audio/protocol receipts No full control/microphone/playback/interrupt/soak qualification; two unsafe Kokoro snapshots and all completeness unverified
No promotion; controls retained Final assessment, original plan and coverage Retention is not new qualification; full Mac gate 31 failed / 9,808 passed / 687 skipped

All copy is derivative. Original SHA/bytes, internal native hashes, redaction/omission locations and projected SHA/bytes resolve through the ledger. No audio, weights, personal identity or private fixture source is public. The late wrapper and post-measurement lint amendment preserve their actual chronology.