[HugoXK]

Partial: #951 — the evidence-card schema and eight new cards. The algorithm-port rows on docs/results.md are still uncarded, so this does not close the issue.

What changed

• The 12-field evidence-card schema is published as a machine-checkable artefact: tools/evidence_card.py --check validates every card in reports/evidence/ and fails on a missing or unknown field instead of skipping it; a missing value is written not reviewed, never omitted.
• Eight new cards in reports/evidence/: one-call path, bring-your-own-agent, efficiency, equal-budget HotpotQA, the merge-cost lever, judge bias across five workloads, ground-truth repair, and the rule-search residual. Each states its held-out set, matched budget, cost, reliability, code pointer and the rest of the twelve fields.
• reports/evidence/README.md indexes all 11 cards; docs/results.md ends by pointing a reader at the directory instead of repeating numbers.
• Two cards are recomputed in CI from committed runs (equal-budget medians plus separates(); the matrix card's 1425 s / 1.39× / 1.92×). The six live-measurement cards say "not reproducible offline" on the card itself rather than implying a reproduction that does not exist.

Size: 12 files, +291/−7 (T1 by path: reports/**, tests/**, docs/**).

Verification at the tip

Tip SHA: ac68dcf2c485be1ba4a7c5d4fe0d90cecb684eb3

Ran: every command below from the repo root of a clean extraction of this tip, PYTHONPATH=., pytest with -o addopts="".

• python3 -m tools.evidence_card --check → 11 cards, exit 0.
• python3 -m pytest -q -o addopts="" tests/test_evidence_card_reports.py → 46 passed.
• python3 -m pytest -q -o addopts="" tests/test_evidence_card.py tests/test_evidence_card_reports.py tests/test_docs_examples.py tests/test_efficiency.py tests/test_matrix_report.py tests/test_baselines.py tests/test_audit_attach.py tests/test_audit_diagnose_script.py → 167 passed.
• python3 -m tools.changelog_fragments check → 34 changelog fragment(s) OK.
• Red at base 0c9e936d9fd48c7712ef2ab2360413fdc927e311: this PR's test file with the eight new card files absent → 33 failed / 13 passed, exit 1.

Not run:

• Full pytest -q; the two pytest commands above are the suites this change touches.
• mkdocs build --strict → skipped, not run at this tip; mkdocs is absent from this box's python, and no earlier run is claimed for this tree.
• The six live-measurement cards are not reproducible offline — each needs its model endpoint and the run behind it, and each card says not reproducible offline on the card itself.

[Birfy]

Verdict: PASS

Tip SHA: ac68dcf2c485be1ba4a7c5d4fe0d90cecb684eb3 · Risk tier: T1 (risk_gate.py from main c94e9d67)

Ran:

• python -m tools.evidence_card --check on the tip → 11 cards, all valid.
• pytest tests/test_evidence_card_reports.py → pass.
• Spot-checked results-efficiency.json against docs/results.md lines 99-126: 5.8× / 1.1× / 1.8× / 1.7× / 2.65× / 3.6 s → 1.2 s all match the page; the retired 7.1× / 5.9× / 2.4× / 3.0× / 193.6 s figures are named as not re-runnable, as the card says.
• Required CI green on this tip.

Not run: the live-API reruns the non-bench cards point at (they say so on the card). Partial for #951, so it correctly does not close the issue.