[daniele21]

Changes

• Introduce suite_public.yaml with separate benchmark lineage (36 smoke / 140 core / 550 extended), leaving the historical 468-case default untouched.
• Pin source Parquet artifacts to immutable HF revisions with DuckDB-based cached source adapters and source checksums.
• Add SQuAD gold aliases/normalization + balanced answerable/unanswerable quotas, GSM8K final-numeric gold parsing, and separate JSONSchemaBench compliance scorer with bounded safe schema filters.
• Add model-bench --suite suite_public.yaml CLI switch and persist the selected suite's checksum/identity in run artifacts.
• Extend deterministic tests and documentation (PUBLIC_BENCHMARKS.md).

Validation / methodology

• Unit and config tests added; live provider smoke + E2E remains to be run on local hardware.
• Results are explicitly MCB subsets, not official full-benchmark scores. No model was contacted and no results have been fabricated.
• Default canonical suite, controlled datasets and historical run evidence remain unchanged.

Suggested follow-up gates

cd experiments/model-capability-benchmark && uv sync --frozen --extra dev && uv run --frozen pytest -q tests/test_public_benchmarks.py tests/test_dataset_adapters.py && uv run --frozen model-bench --suite suite_public.yaml plan --profile core --capabilities all