Release 0.9.1
What changed
- Preserve semisynthetic and rendered prediction-count buckets in per-dataset and
per_source_accuracysummaries. - Add regression coverage for the complete multiclass reporting path.
- Add and refine v0.9 taxonomy, scoring, and safetensors documentation.
The prediction-summary change is bookkeeping-only. It does not alter inference, sampling, metrics, scores, rankings, outcomes, or runtime behavior. Existing lossy summary payloads can be reconstructed from stored parquet records without rerunning inference.
Full Changelog: v0.9.0...v0.9.1