2.1.0

PESTE 2.1.0 corrects Whisper evaluation for recordings longer than 30 seconds and publishes the
corresponding official result refresh.

Highlights:

• Whisper preprocessing is non-truncating and automatically uses native sequential decoding when
  a padded batch exceeds the model's segment limit.
• All 29 Whisper candidates were recalibrated and completed fresh 871-recording official runs,
  including all 13 test recordings longer than 30 seconds.
• The accuracy, speed, and Pareto boards were regenerated for 42 models from 29 refreshed Whisper
  bundles and 13 unchanged non-Whisper 2.0 bundles.
• Result rows expose model, source, runtime-image, and hardware provenance.
• Calibration, legacy Whisper metadata handling, and publication validation were strengthened.

The release uses the fixed PESTE decoding policy. Model-specific tuning, temperature fallback,
prompts, external language models, output repair, quantization, and precision fallback remain
outside the benchmark contract. See the
benchmark contract
and changelog.

Official Whisper runs used source revision
e7833ac24e5e2db15a66ae267066bff576d872c5 and runtime image
ghcr.io/armanjr/peste-benchmark@sha256:7c81292cfd47fe2ed956bf5b81f8cef5d78d1e9558299650fd9d883973bd14ef.

2.0.0

docs: streamline README

1.0.0

docs(readme): soften tagline and align citation version to 1.0.0

Drop the unqualified "reproducible" claim from the tagline and set the
citation version to 1.0.0 to match the package version in pyproject.toml.