No reproducible measured findings are tracked in this repository.
Earlier drafts contained CER, WER, real-time factor, held-out, routing, device, precision, and ensemble conclusions without the complete inputs and artifacts needed to audit them from a clean clone. Those numbers are not current baselines and are not retained as historical evidence.
A future comparison must track or durably identify:
- The evaluation manifest, corpus split, audio identities and checksums, reference transcripts, and their provenance and licence.
- Every model repository, immutable revision, weight checksum, backend version, dependency lock, normalization option, and language code.
- The complete command, hardware, macOS version, resolved device and dtype, and whether any device retry occurred.
- Per-file transcription JSONL, failures, raw timing and resource telemetry, aggregate output, and enough logs to diagnose exclusions.
- A clean-clone procedure that reproduces the aggregate from the tracked or immutable inputs without relying on an operator's existing caches.
Small smoke runs may validate that a backend executes, but they are not model selection evidence. Published comparisons must use the same committed manifest and scoring contract for every model. Bot-generated summaries and review comments are advisory until a person verifies the underlying artifacts.