feat(benchmark): Phase 2 (task 4) — actionable per-rule signals, ledger, FP knob

Complete Phase 2: make benchmark output drive per-rule improvement, wire the
loop-owned ledger, and turn a confirmed false positive into a score knob. Zero
new dependencies; detection stays byte-identical (recall 44/47 = 93.6%, FP 0%)
and the 90/10 gate holds.

- Per-rule attribution (4.1): BenchmarkMiss/BenchmarkFalsePositive carry rule_id;
  BenchmarkMetrics.per_rule aggregates matched/missed/false_positives per rule.
- Recommendation delta (4.2): evaluate_benchmark emits RuleRecommendation
  (keep/shadow/loosen) written to rule_policy_delta.json; the inert
  next_improvement_candidate stub is now rule-aware ("rule:<id>" or
  "language_hunter").
- Loop ledger (4.3): cli loads the ruleset with the rule_policy.json overlay
  (empty -> byte-identical; an entry overlays a rule's status/floors at load);
  read_rule_ledger/write_rule_ledger added for the evolve loop. A disabled rule
  in the ledger no longer fires (covered by an integration test).
- FP -> reachability knob (4.4): a confirmed false-positive learning lowers a
  finding's effective reachability multiplier in the score
  (fp_reachability_overrides + scoring reachability_override) instead of deleting
  the rule, so recall stays high while the score stops over-penalizing.

93 tests pass; ruff + mypy clean. Spec tasks 4.1-4.4 checked off — Phase 2
(Core: rules-as-data + benchmark signals) is complete. The loop-driven decision
to apply ledger flips remains the gated evolve loop (Phase 3/4).

Co-Authored-By: Claude Opus 4.8 <[email protected]>

src/openultrasast/benchmark.py

@@ -60,6 +60,7 @@ class BenchmarkMiss:
     path: str
     evidence: str
     reason: str
+    rule_id: str | None = None
 
 
 @dataclass(frozen=True)
@@ -79,6 +80,7 @@ class BenchmarkCalibrationRecord:
     failed_stage: str
     next_improvement_candidate: str
     reason: str
+    rule_id: str | None = None
 
 
 @dataclass(frozen=True)
@@ -87,6 +89,17 @@ class BenchmarkFalsePositive:
     path: str
     line: int | None
     reason: str
+    rule_id: str = ""
+
+
+@dataclass(frozen=True)
+class RuleRecommendation:
+    rule_id: str
+    action: str  # keep | shadow | loosen
+    matched: int
+    missed: int
+    false_positives: int
+    rationale: str
 
 
 @dataclass(frozen=True)
@@ -101,6 +114,7 @@ class BenchmarkMetrics:
     runtime_seconds: float | None
     model_usage: dict[str, object]
     artifact_links: dict[str, str]
+    per_rule: dict[str, dict[str, int]]
 
 
 @dataclass(frozen=True)
@@ -123,6 +137,7 @@ class BenchmarkResult:
     false_positives: list[BenchmarkFalsePositive]
     baseline_deltas: list[BenchmarkBaselineDelta]
     calibration_records: list[BenchmarkCalibrationRecord]
+    rule_recommendations: list[RuleRecommendation]
 
 
 def load_benchmark_manifest(path: Path) -> BenchmarkManifest:
@@ -224,6 +239,8 @@ def evaluate_benchmark(
     false_positives = [_false_positive(finding) for index, finding in enumerate(findings) if index not in matched_finding_indexes]
     baseline_deltas = _baseline_deltas(run.manifest.expected, baseline_findings or [])
     calibration_records = [_calibration_record(miss) for miss in misses]
+    per_rule = _per_rule_metrics(findings, matched_finding_indexes, misses)
+    rule_recommendations = _rule_recommendations(per_rule)
     expected_total = len(run.manifest.expected)
     actual_total = len(findings)
     missed_total = len(misses)
@@ -249,11 +266,13 @@ def evaluate_benchmark(
             runtime_seconds=runtime_seconds,
             model_usage={},
             artifact_links=_artifact_links(scan_run_dir),
+            per_rule=per_rule,
         ),
         misses=misses,
         false_positives=false_positives,
         baseline_deltas=baseline_deltas,
         calibration_records=calibration_records,
+        rule_recommendations=rule_recommendations,
     )
 
 
@@ -266,6 +285,9 @@ def write_benchmark_artifacts(run: BenchmarkRun, manifest_path: Path, result: Be
     (run.root / "external_baseline_deltas.json").write_text(
         json.dumps(_jsonable([asdict(delta) for delta in result.baseline_deltas]), indent=2, sort_keys=True) + "\n"
     )
+    (run.root / "rule_policy_delta.json").write_text(
+        json.dumps(_jsonable([asdict(rec) for rec in result.rule_recommendations]), indent=2, sort_keys=True) + "\n"
+    )
 
 
 def _matching_finding_index(expected: ExpectedFinding, findings: list[StaticFinding], used_indexes: set[int]) -> int | None:
@@ -301,6 +323,7 @@ def _miss(expected: ExpectedFinding) -> BenchmarkMiss:
         path=expected.path,
         evidence=expected.evidence,
         reason="no OpenUltraSAST finding matched the expected benchmark vulnerability",
+        rule_id=expected.rule_id,
     )
 
 
@@ -310,9 +333,57 @@ def _false_positive(finding: StaticFinding) -> BenchmarkFalsePositive:
         path=finding.path,
         line=finding.line,
         reason="finding did not match any expected benchmark vulnerability",
+        rule_id=_rule_id_of(finding.finding_id),
     )
 
 
+def _rule_id_of(finding_id: str) -> str:
+    return finding_id.split(":", 1)[0]
+
+
+def _per_rule_metrics(
+    findings: list[StaticFinding],
+    matched_finding_indexes: set[int],
+    misses: list[BenchmarkMiss],
+) -> dict[str, dict[str, int]]:
+    per_rule: dict[str, dict[str, int]] = {}
+
+    def bucket(rule_id: str) -> dict[str, int]:
+        return per_rule.setdefault(rule_id, {"matched": 0, "missed": 0, "false_positives": 0})
+
+    for index, finding in enumerate(findings):
+        rule_id = _rule_id_of(finding.finding_id)
+        key = "matched" if index in matched_finding_indexes else "false_positives"
+        bucket(rule_id)[key] += 1
+    for miss in misses:
+        if miss.rule_id:
+            bucket(miss.rule_id)["missed"] += 1
+    return dict(sorted(per_rule.items()))
+
+
+def _rule_recommendations(per_rule: dict[str, dict[str, int]]) -> list[RuleRecommendation]:
+    recommendations: list[RuleRecommendation] = []
+    for rule_id, counts in per_rule.items():
+        matched, missed, fps = counts["matched"], counts["missed"], counts["false_positives"]
+        if fps > 0:
+            action, rationale = "shadow", "rule produced false positives; stage it before it drags precision"
+        elif missed > 0:
+            action, rationale = "loosen", "rule missed expected findings; widen or lower its evidence floor"
+        else:
+            action, rationale = "keep", "rule matched expected findings with no false positives"
+        recommendations.append(
+            RuleRecommendation(
+                rule_id=rule_id,
+                action=action,
+                matched=matched,
+                missed=missed,
+                false_positives=fps,
+                rationale=rationale,
+            )
+        )
+    return recommendations
+
+
 def _baseline_deltas(expected_findings: list[ExpectedFinding], baseline_findings: list[BaselineFinding]) -> list[BenchmarkBaselineDelta]:
     tools = sorted({finding.tool for finding in baseline_findings})
     deltas: list[BenchmarkBaselineDelta] = []
@@ -359,14 +430,18 @@ def _baseline_matches_expected(expected: ExpectedFinding, finding: BaselineFindi
 
 
 def _calibration_record(miss: BenchmarkMiss) -> BenchmarkCalibrationRecord:
+    # A miss tied to a known rule points the loop at that rule; an unattributed
+    # miss (no rule fired for the class) is a coverage gap for a language hunter.
+    candidate = f"rule:{miss.rule_id}" if miss.rule_id else "language_hunter"
     return BenchmarkCalibrationRecord(
         cwe=miss.cwe,
         vulnerability_class=miss.vulnerability_class,
         path=miss.path,
         evidence=miss.evidence,
         failed_stage="benchmark_ground_truth_matching",
-        next_improvement_candidate="rules_or_language_hunter",
+        next_improvement_candidate=candidate,
         reason=miss.reason,
+        rule_id=miss.rule_id,
     )
 
 

src/openultrasast/calibration.py

@@ -249,6 +249,30 @@ def learnings_from_verifications(
     return learnings
 
 
+# A confirmed false positive lowers a finding's effective reachability multiplier
+# by this factor in the project score (it does not delete the rule).
+FP_REACHABILITY_FACTOR = 0.4
+_BASE_REACH_MULT = {"reachable": 1.0, "inferred-file-surface": 0.6, "unknown": 0.4}
+
+
+def fp_reachability_overrides(
+    findings: list[StaticFinding],
+    learnings: list[FalsePositiveLearning],
+) -> dict[str, float]:
+    """Lowered score multipliers for findings whose scope has a confirmed FP.
+
+    Keyed by ``finding_id``; consumed by the scoring model so a confirmed false
+    positive reduces the score penalty instead of the rule being deleted.
+    """
+    overrides: dict[str, float] = {}
+    for finding in findings:
+        vulnerability_class = _vulnerability_class(finding)
+        if any(_learning_applies(finding.path, learning) and learning.vulnerability_class == vulnerability_class for learning in learnings):
+            base = _BASE_REACH_MULT.get(finding.reachability_status, _BASE_REACH_MULT["unknown"])
+            overrides[finding.finding_id] = round(base * FP_REACHABILITY_FACTOR, 3)
+    return overrides
+
+
 def _reason_from_verification(finding: StaticFinding, status: VerificationStatus) -> FalsePositiveReason:
     if finding.reachability_status == "unknown":
         return FalsePositiveReason.UNREACHABLE_PATH

src/openultrasast/cli.py

@@ -18,6 +18,7 @@ from .benchmark import (
 from .calibration import (
     FalsePositiveLearning,
     calibrate_rankings,
+    fp_reachability_overrides,
     learnings_from_verifications,
     load_false_positive_learnings,
     merge_false_positive_learnings,
@@ -105,7 +106,9 @@ def _run_scan(path: Path, config_path: Path, mode: str, fail_on: str) -> ScanOut
     write_harness_config(config=config, processors=[], contract_mode="strict", path=run.root / "harness.json")
     runtime.start(mode=mode, target=run.target)
     policy = runtime.run_stage("policy_load", load_policy)
-    ruleset = runtime.run_stage("ruleset_load", lambda: load_ruleset(DEFAULT_RULESET_DIR))
+    ledger_path = run.target / CALIBRATION_DIR / "false_positive_learnings.json"
+    rule_ledger_path = run.target / CALIBRATION_DIR / "rule_policy.json"
+    ruleset = runtime.run_stage("ruleset_load", lambda: load_ruleset(DEFAULT_RULESET_DIR, rule_ledger_path))
     runtime.run_stage("policy_check", lambda: assert_rules_resolve(ruleset, policy))
     static_hints = runtime.run_stage("static_mapping", lambda: _load_static_hints(config.static_analysis.sarif_paths))
     write_static_hints(static_hints, run.root / "mapping" / "static_hints.json")
@@ -118,7 +121,6 @@ def _run_scan(path: Path, config_path: Path, mode: str, fail_on: str) -> ScanOut
     targets = attach_reachability_hints(targets, entry_points)
     write_preprocess_artifact(snapshot, targets, run.root / "preprocess" / "file_targets.json")
     rankings = runtime.run_stage("rank", lambda: rank_targets(targets))
-    ledger_path = run.target / CALIBRATION_DIR / "false_positive_learnings.json"
     prior_learnings = load_false_positive_learnings(ledger_path)
     rankings, calibrations = runtime.run_stage("calibrate", lambda: calibrate_rankings(rankings, prior_learnings))
     applied_calibrations = [calibration for calibration in calibrations if calibration.applied_learning_ids]
@@ -168,6 +170,9 @@ def _run_scan(path: Path, config_path: Path, mode: str, fail_on: str) -> ScanOut
     )
     cwe_by_rule = {rule.rule_id: rule.cwe for rule in ruleset}
     rule_cwe_by_id = {finding.finding_id: cwe_by_rule.get(finding.finding_id.split(":", 1)[0], "") for finding in findings}
+    # A confirmed false positive (prior-scan learning) lowers a finding's effective
+    # reachability multiplier in the score instead of deleting the rule.
+    reachability_override = fp_reachability_overrides(findings, prior_learnings)
     score_artifact = runtime.run_stage(
         "score",
         lambda: build_score_artifact(
@@ -178,6 +183,7 @@ def _run_scan(path: Path, config_path: Path, mode: str, fail_on: str) -> ScanOut
             min_score=config.score.min_score,
             block_severity_reachable=config.score.block_severity_reachable,
             blocking=config.score.blocking,
+            reachability_override=reachability_override,
         ),
     )
     score_path = run.root / "score.json"

src/openultrasast/ruleset/__init__.py

@@ -5,6 +5,8 @@ from .store import (
     PatternRule,
     RulesetError,
     load_ruleset,
+    read_rule_ledger,
+    write_rule_ledger,
     write_ruleset,
 )
 
@@ -13,5 +15,7 @@ __all__ = [
     "PatternRule",
     "RulesetError",
     "load_ruleset",
+    "read_rule_ledger",
+    "write_rule_ledger",
     "write_ruleset",
 ]

src/openultrasast/ruleset/store.py

@@ -81,6 +81,17 @@ def write_ruleset(path: Path, rules: Iterable[PatternRule]) -> None:
     path.write_text("\n\n".join(blocks) + "\n")
 
 
+def read_rule_ledger(path: Path) -> dict[str, dict[str, object]]:
+    """Read the loop-owned ``rule_policy.json`` overlay (status/floors per rule_id)."""
+    return _load_ledger(path)
+
+
+def write_rule_ledger(path: Path, entries: dict[str, dict[str, object]]) -> None:
+    """Write the loop-owned ``rule_policy.json`` overlay (used by the evolve loop)."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(entries, indent=2, sort_keys=True) + "\n")
+
+
 def _rule_from_dict(item: dict[str, object], path: Path) -> PatternRule:
     try:
         rule_id = str(item["rule_id"])

src/openultrasast/scoring/project_score.py

@@ -29,15 +29,26 @@ class ScoreArtifact:
         return asdict(self)
 
 
-def finding_penalty(finding: StaticFinding, rule_cwe: str | None, policy: Mapping[str, CwePolicy]) -> float:
+def finding_penalty(
+    finding: StaticFinding,
+    rule_cwe: str | None,
+    policy: Mapping[str, CwePolicy],
+    reachability_override: Mapping[str, float] | None = None,
+) -> float:
     """Penalty contributed by one finding: severity weight × reachability multiplier.
 
-    Dynamic-only or unmapped CWEs contribute zero (report-only, never scored).
+    Dynamic-only or unmapped CWEs contribute zero (report-only, never scored). A
+    confirmed false positive can supply a lowered effective reachability multiplier
+    via ``reachability_override`` (keyed by ``finding_id``) instead of deleting the rule.
     """
     pol = policy.get(rule_cwe) if rule_cwe else None
     if pol is None or not pol.static:
         return 0.0
-    return SEV_WEIGHT[pol.severity] * REACH_MULT.get(finding.reachability_status, REACH_MULT["unknown"])
+    if reachability_override and finding.finding_id in reachability_override:
+        multiplier = reachability_override[finding.finding_id]
+    else:
+        multiplier = REACH_MULT.get(finding.reachability_status, REACH_MULT["unknown"])
+    return SEV_WEIGHT[pol.severity] * multiplier
 
 
 def project_score(
@@ -45,13 +56,14 @@ def project_score(
     rule_cwe_by_id: Mapping[str, str],
     policy: Mapping[str, CwePolicy],
     k: float = 60.0,
+    reachability_override: Mapping[str, float] | None = None,
 ) -> int:
     """Compute the 0-100 project score (shadow findings excluded by the caller).
 
     Exponential decay of total penalty: no penalty → 100; one severity-5 reachable
     finding → ~43.
     """
-    total = sum(finding_penalty(finding, rule_cwe_by_id.get(finding.finding_id), policy) for finding in findings)
+    total = sum(finding_penalty(finding, rule_cwe_by_id.get(finding.finding_id), policy, reachability_override) for finding in findings)
     return round(100 * math.exp(-total / k))
 
 
@@ -92,8 +104,9 @@ def build_score_artifact(
     min_score: int = 80,
     block_severity_reachable: int = 5,
     blocking: bool = False,
+    reachability_override: Mapping[str, float] | None = None,
 ) -> ScoreArtifact:
-    score = project_score(findings, rule_cwe_by_id, policy, k=k)
+    score = project_score(findings, rule_cwe_by_id, policy, k=k, reachability_override=reachability_override)
     penalty_total = 0.0
     max_severity = 0
     by_category: dict[str, int] = {}
@@ -109,7 +122,7 @@ def build_score_artifact(
             if pol.dynamic:
                 out_of_scope_dynamic_only += 1
             continue
-        penalty_total += finding_penalty(finding, cwe, policy)
+        penalty_total += finding_penalty(finding, cwe, policy, reachability_override)
         max_severity = max(max_severity, pol.severity)
         by_category[pol.flaw_category] = by_category.get(pol.flaw_category, 0) + 1
     return ScoreArtifact(