add stage-separated AMFV-Bench v0 eval

[VibeCodingScientist]

Summary

• Adds AMFV-Bench v0: a frozen, stage-separated eval protocol (decomposer / retrieval / verifier) so later PRs can say retrieval A beats B instead of optimizing against vibes.
• Ships 48 NICE-grounded gold cases (8 per stratum: strongly supported, weakly supported, refuted, insufficient, population mismatch, temporal), with Med-V1 Likert verdicts, scope Z, as-of date T, pyramid tier, and short quoted spans — not chunk hashes.
• Includes a deterministic offline scorer (amfv-eval) that never calls a network or LLM, plus a clinician annotation rubric in datasets/amfv_datasets/eval/taxonomy.md.

This is the evaluation-first artifact from the 12 Aug restart. Population-mismatch cases must abstain (verdict 0); supporting them because a disease name matched is a fail. Temporal cases point at the replacement guideline with replaces set to the withdrawn id.

Test plan

• uv sync --dev
• uv run pytest datasets/test
• uv run ruff check datasets && uv run ruff format --check datasets
• uv run amfv-eval --predictions datasets/test/fixtures/eval/predictions_demo.jsonl (demo file supports a paediatric claim on purpose; scope_abstention should be 0)
• Skim datasets/amfv_datasets/eval/taxonomy.md and 2–3 gold cases for clinical sense

Made with Cursor

[CLAassistant]

All committers have signed the CLA.

[VibeCodingScientist]

Closing so I can review the branch locally before opening a PR.

[VibeCodingScientist]

Superseded by #20 so the red-team scorer fixes and clinician-seed note land in one open PR. This closed PR can be ignored.