[QuentinFuxa]

Retain a reproducible M5 evaluation of #425 and #426 against Whisper small MLX/LocalAgreement and Qwen3 Metal: 30 FLEURS clips per language, three paced passes after warmup, plus a ten-minute Chinese stream per backend. Historical figures, the root README and its badge remain unchanged.

m5-20261107/ holds the complete series: all sixteen schema-3.1 reports, measured in one serial session on AC power (253 power samples, all on AC, no run flagged), with the same commits, model revisions and corpus as September. It adds the series runner, fresh-process startup measurements, the power log, environment versions, the summary and a figure. m5-20260907/ keeps the partial September series unchanged, including its AC-to-battery transition and the #446 direct-translation failures.

Neither candidate passes the agreed rule against both existing backends:

Comparison	English	French	Chinese
Native Qwen vs Whisper small	fail	fail	fail
Native Qwen vs Qwen Metal	fail	fail	fail
Nemotron vs Whisper small	fail	pass	pass
Nemotron vs Qwen Metal	pass	pass	fail

• Native Qwen: 27.65% / 44.68% WER and 28.46% CER, worse than Whisper small and than Qwen Metal with the same weights; EOF p95 is slower than both in every pass. The ten-minute stream fails reproducibly after 101.5 s when finish_streaming() receives a 160-sample tail at a pause.
• Nemotron: EOF p95 is 71% to 94% lower than both baselines in every pass and quality matches September exactly, but English WER is 7.6 points worse than Whisper and Chinese CER 10.8 points worse than Metal. Its ten-minute stream completes without a repetition loop.
• Whisper's ten-minute stream repeats the same phrase 31 times again; Metal's completes without a loop.

The measured revisions are those of the candidate PRs and b2baee9, not current main. Ordinary desktop activity was not controlled. The details and limits are in m5-20261107/METHODS.md and REVIEW.md.