feat: Phase F F10 negative-result citation policy

docs/research/NEXT_100_PR_MAP.md

@@ -118,7 +118,7 @@ Make failure useful.
 | F7 | Add calibration link from negative-result entries to intake reports (complete). — evidence/negative_result_calibration_link.py: links NRR- rejection records to calibration intake reports for closed-loop learning; tests/evidence/test_negative_result_calibration_link.py. | Closes learning loop. | C |
 | F8 | Add benchmark for whether rejected candidates resemble known failure modes (complete). — FMS- schema: 14 fields, 11 validation rules, PATTERN_REPEATED_THRESHOLD=0.80, pattern_repeated_flag enforcement, calibration_action_recommended; 63 tests. | Improves rejection logic. | C |
 | F9 | Add negative-result dashboard schema (complete). | NRD- aggregates NRR- rejection statistics: rejection rate consistency check, all_rejections_have_nrr enforced, top stage/reason controlled vocabulary, 100% rejection warning. | B |
-| F10 | Add policy that public claims must mention relevant negative results. | Prevents cherry-picking. | D |
+| F10 | Add policy that public claims must mention relevant negative results. | Prevents cherry-picking. | D | DONE |
 
 ## Phase G — Calibration and active-learning rigor
 

src/openamp_foundry/evidence/negative_result_citation_policy.py

@@ -0,0 +1,153 @@
+"""NRC- negative-result citation policy schema.
+
+Enforces that any public positive claim references all known relevant negative
+results.  Cherry-picking — highlighting successes while silently omitting
+failures — is the most common way an AI pipeline overstates its case.
+
+Every public claim about pipeline outputs must be paired with an NRC- record
+that lists relevant NRR- records and confirms they are cited.  A claim with
+uncited negatives is machine-flagged as non_compliant.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+VALID_NRC_VERDICTS: frozenset[str] = frozenset({
+    "compliant",
+    "non_compliant",
+    "no_relevant_negatives",
+})
+
+
+@dataclass
+class NegativeResultCitationPolicy:
+    nrc_id: str
+    claim_id: str
+    pipeline_version: str
+    relevant_nrr_ids: list[str]
+    cited_nrr_ids: list[str]
+    uncited_nrr_ids: list[str]
+    uncited_count: int
+    all_relevant_cited: bool
+    policy_verdict: str
+    dry_lab_only: bool
+    limitations: list[str]
+    created_at: str
+
+
+def validate_negative_result_citation_policy(nrc: NegativeResultCitationPolicy) -> None:
+    if not nrc.nrc_id.startswith("NRC-"):
+        raise ValueError(f"nrc_id must start with 'NRC-': {nrc.nrc_id!r}")
+    if not nrc.claim_id:
+        raise ValueError("claim_id must be non-empty")
+    if not nrc.pipeline_version:
+        raise ValueError("pipeline_version must be non-empty")
+    for nrr_id in nrc.relevant_nrr_ids:
+        if not nrr_id.startswith("NRR-"):
+            raise ValueError(f"relevant_nrr_id must start with 'NRR-': {nrr_id!r}")
+    for nrr_id in nrc.cited_nrr_ids:
+        if not nrr_id.startswith("NRR-"):
+            raise ValueError(f"cited_nrr_id must start with 'NRR-': {nrr_id!r}")
+    relevant_set = set(nrc.relevant_nrr_ids)
+    for nrr_id in nrc.cited_nrr_ids:
+        if nrr_id not in relevant_set:
+            raise ValueError(
+                f"cited_nrr_id {nrr_id!r} not in relevant_nrr_ids"
+            )
+    expected_uncited = [r for r in nrc.relevant_nrr_ids if r not in set(nrc.cited_nrr_ids)]
+    if sorted(nrc.uncited_nrr_ids) != sorted(expected_uncited):
+        raise ValueError("uncited_nrr_ids inconsistent with relevant_nrr_ids and cited_nrr_ids")
+    if nrc.uncited_count != len(nrc.uncited_nrr_ids):
+        raise ValueError("uncited_count must equal len(uncited_nrr_ids)")
+    expected_all_cited = nrc.uncited_count == 0
+    if nrc.all_relevant_cited != expected_all_cited:
+        raise ValueError("all_relevant_cited inconsistent with uncited_count")
+    if nrc.policy_verdict not in VALID_NRC_VERDICTS:
+        raise ValueError(
+            f"policy_verdict {nrc.policy_verdict!r} not in VALID_NRC_VERDICTS"
+        )
+    if not nrc.relevant_nrr_ids and nrc.policy_verdict != "no_relevant_negatives":
+        raise ValueError(
+            "policy_verdict must be 'no_relevant_negatives' when relevant_nrr_ids is empty"
+        )
+    if nrc.relevant_nrr_ids and nrc.all_relevant_cited and nrc.policy_verdict != "compliant":
+        raise ValueError(
+            "policy_verdict must be 'compliant' when all relevant negatives are cited"
+        )
+    if nrc.relevant_nrr_ids and not nrc.all_relevant_cited and nrc.policy_verdict != "non_compliant":
+        raise ValueError(
+            "policy_verdict must be 'non_compliant' when relevant negatives are uncited"
+        )
+    if not nrc.dry_lab_only:
+        raise ValueError("dry_lab_only must be True")
+    if not nrc.limitations:
+        raise ValueError("limitations must be non-empty")
+    if not nrc.created_at:
+        raise ValueError("created_at must be non-empty")
+
+
+def _compute_verdict(relevant: list[str], cited_set: set[str]) -> str:
+    if not relevant:
+        return "no_relevant_negatives"
+    if all(r in cited_set for r in relevant):
+        return "compliant"
+    return "non_compliant"
+
+
+def build_negative_result_citation_policy(
+    *,
+    nrc_id: str,
+    claim_id: str,
+    pipeline_version: str,
+    relevant_nrr_ids: list[str],
+    cited_nrr_ids: list[str],
+    limitations: list[str],
+    created_at: str,
+) -> NegativeResultCitationPolicy:
+    """Build a NegativeResultCitationPolicy.
+
+    uncited_nrr_ids, uncited_count, all_relevant_cited, and policy_verdict
+    are all auto-computed from relevant_nrr_ids and cited_nrr_ids.
+    """
+    relevant = list(relevant_nrr_ids)
+    cited = list(cited_nrr_ids)
+    cited_set = set(cited)
+    uncited = [r for r in relevant if r not in cited_set]
+    uncited_count = len(uncited)
+    all_cited = uncited_count == 0
+    verdict = _compute_verdict(relevant, cited_set)
+    nrc = NegativeResultCitationPolicy(
+        nrc_id=nrc_id,
+        claim_id=claim_id,
+        pipeline_version=pipeline_version,
+        relevant_nrr_ids=relevant,
+        cited_nrr_ids=cited,
+        uncited_nrr_ids=uncited,
+        uncited_count=uncited_count,
+        all_relevant_cited=all_cited,
+        policy_verdict=verdict,
+        dry_lab_only=True,
+        limitations=limitations,
+        created_at=created_at,
+    )
+    validate_negative_result_citation_policy(nrc)
+    return nrc
+
+
+def format_negative_result_citation_policy(nrc: NegativeResultCitationPolicy) -> str:
+    lines = [
+        f"Negative-Result Citation Policy — {nrc.nrc_id}",
+        f"Claim: {nrc.claim_id}  |  Pipeline: {nrc.pipeline_version}",
+        f"Verdict: {nrc.policy_verdict}",
+        f"Relevant negatives: {len(nrc.relevant_nrr_ids)}  "
+        f"Cited: {len(nrc.cited_nrr_ids)}  "
+        f"Uncited: {nrc.uncited_count}",
+    ]
+    if nrc.uncited_nrr_ids:
+        lines.append(f"Uncited negatives: {', '.join(nrc.uncited_nrr_ids)}")
+    lines.append(f"All relevant cited: {nrc.all_relevant_cited}")
+    lines.append(f"Limitations: {'; '.join(nrc.limitations)}")
+    lines.append(f"Created: {nrc.created_at}")
+    lines.append(f"dry_lab_only: {nrc.dry_lab_only}")
+    return "\n".join(lines)