python -m mrf_honest.ai.eval --records <assessments.jsonl> --output evals/ai/results/<date>-<cohort>-<provider>-<model>.json
runs mrf-honest narrate over committed assessment records and counts, per
record and overall, the claims the model generated, the claims shown (every
cited quote occurs verbatim in the named committed document under corpus/),
and the claims withheld. It does not measure whether a shown claim is a
correct reading of the passage it quotes, and no gold explanations exist; a
verified citation proves the passage exists and says those words.
Each result records provider, model, prompt version, UTC date, the Git commit,
and the records file, and tests/test_ai_narration.py refuses a result
without that provenance. Numbers are never written by hand; an ai extra and
MRF_AI_PROVIDER / MRF_AI_MODEL (Anthropic API or Amazon Bedrock through the
public SDK) are required to run it.