@clear-capabilities/agentic-security-scanner 0.166.0 → 0.168.0

package/src/posture/proof-fabric/store.js

@@ -0,0 +1,83 @@
+// Proof artifact store: `.agentic-security/proofs/<findingKey>/<recordDigest>.json`.
+//
+// Reuses the existing state machinery rather than adding another: `safeWriteState` (refuses unsafe dirs and symlinks),
+// `maybeEncryptForWrite` / `maybeDecryptForRead` (the `proofs` artifact is registered `confidential`), retention
+// (evidence class) and legal holds (a hold on `proofs` blocks purge). Nothing persisted contains a plaintext credential:
+// records are passed through a recursive secret filter first, and the runner never returns its per-run nonce or secrets.
+import fs from 'node:fs';
+import path from 'node:path';
+import { statePath, ensureStateDir, safeWriteState } from '../state-dir.js';
+import { maybeEncryptForWrite, maybeDecryptForRead } from '../encryption-provider.js';
+import { redactSecrets } from '../../llm-validator/redact.js';
+import { canonicalize, sha256 } from './reproducer.js';
+
+const PROOF_RECORD_SCHEMA = 'agentic-security/proof-record';
+const PROOFS_DIR = 'proofs';
+const MAX_RECORD_BYTES = 512 * 1024;
+
+export const findingKey = (id) => `${String(id).replace(/[^A-Za-z0-9._-]/g, '_').slice(0, 80)}-${sha256(String(id)).slice(0, 8)}`;
+
+// The shared redactor catches key=value shapes; bare tokens need their own pass, so persisted evidence cannot carry one.
+const BARE_SECRET_RES = [
+  /\b(?:ghp|gho|ghu|ghs|ghr|github_pat)_[A-Za-z0-9_]{20,}\b/g, /\b(?:AKIA|ASIA)[A-Z0-9]{16}\b/g, /\bsk-[A-Za-z0-9_-]{20,}\b/g,
+  /\bxox[abprs]-[A-Za-z0-9-]{10,}\b/g, /\beyJ[A-Za-z0-9_-]{8,}\.[A-Za-z0-9_-]{8,}\.[A-Za-z0-9_-]{8,}\b/g,
+  /-----BEGIN [A-Z ]*PRIVATE KEY-----[\s\S]*?(?:-----END [A-Z ]*PRIVATE KEY-----|$)/g, /\bSECRET-[0-9a-f]{24}-[a-z0-9_-]+\b/g,
+];
+function scrubString(text) {
+  let t = String(text);
+  try { t = redactSecrets(t).text; } catch { return '[unredactable]'; }
+  for (const re of BARE_SECRET_RES) t = t.replace(re, '[REDACTED-SECRET]');
+  return t;
+}
+
+/** Deep-copy `v` with every string passed through the secret redactor. */
+export function scrubDeep(v) {
+  if (typeof v === 'string') return scrubString(v);
+  if (Array.isArray(v)) return v.map(scrubDeep);
+  if (v && typeof v === 'object') return Object.fromEntries(Object.entries(v).map(([k, x]) => [k, scrubDeep(x)]));
+  return v;
+}
+
+function recordDigest(record) {
+  const { recordDigest: _omit, ...rest } = record;
+  return sha256(canonicalize(rest));
+}
+
+export function saveProof(scanRoot, record) {
+  const safe = scrubDeep(record);
+  safe.schema = PROOF_RECORD_SCHEMA;
+  safe.recordDigest = recordDigest(safe);
+  const text = JSON.stringify(safe, null, 2);
+  if (Buffer.byteLength(text) > MAX_RECORD_BYTES) return { ok: false, reason: 'proof record exceeds the size limit' };
+  if (!ensureStateDir(scanRoot)) return { ok: false, reason: 'no safe state directory for this project' };
+  const enc = maybeEncryptForWrite(scanRoot, PROOFS_DIR, text);
+  if (!enc.ok) return { ok: false, reason: enc.reason };
+  const file = statePath(scanRoot, PROOFS_DIR, findingKey(record.findingId), `${safe.recordDigest}.json`);
+  if (!safeWriteState(file, enc.content)) return { ok: false, reason: 'state write refused' };
+  return { ok: true, file, digest: safe.recordDigest, encrypted: enc.encrypted };
+}
+
+function readRecord(file) {
+  const text = maybeDecryptForRead(fs.readFileSync(file, 'utf8'));
+  const rec = JSON.parse(text);
+  // A record whose content no longer matches its own digest was modified after it was written.
+  if (rec.recordDigest !== recordDigest(rec)) return { tampered: true, record: rec };
+  return { tampered: false, record: rec };
+}
+
+/** Newest-first records for a finding. Tampered records are returned flagged, never silently dropped or trusted. */
+export function listProofs(scanRoot, findingId) {
+  const dir = statePath(scanRoot, PROOFS_DIR, findingKey(findingId));
+  let names = [];
+  try { names = fs.readdirSync(dir).filter((n) => n.endsWith('.json')); } catch { return []; }
+  const out = [];
+  for (const n of names) {
+    const f = path.join(dir, n);
+    try { const { tampered, record } = readRecord(f); out.push({ file: f, tampered, record, mtime: fs.statSync(f).mtimeMs }); } catch { out.push({ file: f, tampered: true, record: null, mtime: 0 }); }
+  }
+  return out.sort((a, b) => b.mtime - a.mtime);
+}
+
+export function latestProof(scanRoot, findingId) {
+  return listProofs(scanRoot, findingId).find((p) => p.record) ?? null;
+}

package/src/posture/rule-compiler/cli.js

@@ -0,0 +1,141 @@
+// CLI surface: `agentic-security rules synthesize|evaluate|promote|status|rollback`.
+//
+// Distinct from the pre-existing `labs synthesize-rule` (posture/rule-synthesis.js), which proposes SUPPRESSION
+// rules from repeated false-positive triage verdicts — a different input, a different output and a different
+// command group on purpose, so the two are never confused at the CLI.
+//
+// Non-interactive by construction: no command reads stdin, every corpus evaluation is bounded (no network, no
+// model call — this is the deterministic half; a model may PROPOSE the candidate pattern this CLI is handed, but
+// it never decides promotion).
+import fs from 'node:fs';
+import { statePath } from '../state-dir.js';
+import { draftRuleIR, validateRuleIR, ruleIRDigest } from './ir.js';
+import { checkEligibility } from './eligibility.js';
+import { buildCorpus } from './corpus.js';
+import { gatePromotion, checkLeakage, PROMOTION_DEFAULTS } from './promotion.js';
+import { recordTransition, stateOf, historyFor } from './lifecycle.js';
+import { loadLedger, saveLedger, saveIR, loadIR, saveEvaluation, writeFixtures, publishCustomRule, unpublishCustomRule } from './store.js';
+
+function loadFinding(root, id) {
+  const p = statePath(root, 'last-scan.json');
+  if (!fs.existsSync(p)) return null;
+  const last = JSON.parse(fs.readFileSync(p, 'utf8'));
+  return (last.findings || []).find((f) => f.id === id || f.stableId === id) ?? null;
+}
+
+function readJsonArg(pathOrInline) {
+  if (!pathOrInline) return null;
+  try { return JSON.parse(fs.existsSync(pathOrInline) ? fs.readFileSync(pathOrInline, 'utf8') : pathOrInline); } catch { return null; }
+}
+
+export async function cmdRules(args, root) {
+  const sub = args._[1];
+  if (sub === 'synthesize') return cmdSynthesize(args, root);
+  if (sub === 'evaluate') return cmdEvaluate(args, root);
+  if (sub === 'promote') return cmdPromote(args, root);
+  if (sub === 'status') return cmdStatus(args, root);
+  if (sub === 'rollback') return cmdRollback(args, root);
+  console.error('usage: rules synthesize <finding-id> --spec <file.json> | evaluate <rule-id> | promote <rule-id> | status [rule-id] | rollback <rule-id>');
+  return 4;
+}
+
+// `--spec` carries exactly what a model or a human proposed: id/title/severity/cwe/family/languages/match/
+// remediation plus the corpus inputs (originalCode/patchedCode/nearbyCleanCode/protectedVariantsCode). This
+// command never invents a pattern; it only drafts the IR, checks eligibility and records DRAFT->CANDIDATE.
+async function cmdSynthesize(args, root) {
+  const findingId = args._[2];
+  const spec = readJsonArg(args.flags.spec);
+  if (!findingId || !spec) { console.error('usage: rules synthesize <finding-id> --spec <file.json>'); return 4; }
+  const finding = loadFinding(root, findingId);
+  if (!finding) { console.error(`Finding ${findingId} not found in last scan.`); return 4; }
+  const elig = checkEligibility(finding, { proofState: args.flags['proof-state'] || null });
+  if (!elig.eligible) { console.error(`Not eligible for rule synthesis: ${elig.reasons.join('; ')}`); return 4; }
+
+  const ir = draftRuleIR({ ...spec, originatingFindingId: finding.stableId ?? finding.id, originatingCommit: spec.originatingCommit ?? null, evidenceState: elig.evidenceState });
+  const errs = validateRuleIR(ir);
+  if (errs.length) { console.error(`Invalid rule: ${errs.join('; ')}`); return 4; }
+
+  const ledger = loadLedger(root);
+  const draft = recordTransition(ledger, { ruleId: ir.id, from: null, to: 'DRAFT', reason: `drafted from ${findingId}`, actor: args.flags.actor || 'agent' });
+  if (!draft.ok) { console.error(draft.reason); return 4; }
+  const cand = recordTransition(ledger, { ruleId: ir.id, from: 'DRAFT', to: 'CANDIDATE', reason: 'eligibility passed', evidence: { eligibility: elig }, actor: args.flags.actor || 'agent' });
+  saveLedger(root, ledger);
+  saveIR(root, { ...ir, state: 'CANDIDATE' });
+  console.log(`Rule ${ir.id}: DRAFT -> CANDIDATE (digest ${ruleIRDigest(ir).slice(0, 12)})`);
+  if (!cand.ok) console.error(cand.reason);
+  return 0;
+}
+
+async function cmdEvaluate(args, root) {
+  const id = args._[2];
+  if (!id) { console.error('usage: rules evaluate <rule-id> --corpus <file.json>'); return 4; }
+  const ir = loadIR(root, id);
+  if (!ir) { console.error(`No rule IR found for ${id}. Run rules synthesize first.`); return 4; }
+  const corpusSpec = readJsonArg(args.flags.corpus);
+  if (!corpusSpec) { console.error('No corpus supplied. Pass --corpus <file.json> with originalCode (and optionally patchedCode/nearbyCleanCode/protectedVariantsCode).'); return 4; }
+  const corpus = buildCorpus(ir, corpusSpec);
+  const leakage = checkLeakage(JSON.stringify(corpusSpec), { benchmarkNames: args.flags['benchmark-names'] ? String(args.flags['benchmark-names']).split(',') : [] });
+  const gate = gatePromotion({ ir, corpus, leakageCheck: leakage, thresholds: args.flags.thresholds ? readJsonArg(args.flags.thresholds) : PROMOTION_DEFAULTS });
+
+  const ledger = loadLedger(root);
+  const cur = stateOf(ledger, id);
+  if (gate.pass && cur === 'CANDIDATE') recordTransition(ledger, { ruleId: id, from: 'CANDIDATE', to: 'VALIDATED', reason: 'promotion gate passed', evidence: { score: gate.evaluation.score }, actor: args.flags.actor || 'agent' });
+  saveLedger(root, ledger);
+  saveEvaluation(root, id, gate);
+  if (args.flags.json) console.log(JSON.stringify(gate, null, 2));
+  else {
+    console.log(`Rule ${id}: ${gate.pass ? 'VALIDATED' : 'gate FAILED'}`);
+    console.log(`  precision ${fmtPct(gate.evaluation.score.precision)}  recall ${fmtPct(gate.evaluation.score.recall)}  protected-failures ${gate.evaluation.score.protectedFailures}`);
+    for (const r of gate.reasons) console.log(`  - ${r}`);
+  }
+  if (gate.pass) { const f = writeFixtures(root, ir, corpus); if (f.ok) console.log(`  regression fixtures written: ${f.written.length} file(s) under .agentic-security/rule-compiler/${id}/fixtures/`); }
+  return gate.pass ? 0 : 1;
+}
+const fmtPct = (n) => (n == null ? 'unmeasured' : `${(n * 100).toFixed(1)}%`);
+
+async function cmdPromote(args, root) {
+  const id = args._[2];
+  if (!id) { console.error('usage: rules promote <rule-id>'); return 4; }
+  const ledger = loadLedger(root);
+  if (ledger.tampered) { console.error(`Refusing: rule lifecycle ledger failed its integrity check (${ledger.tamperReason}).`); return 4; }
+  if (stateOf(ledger, id) !== 'VALIDATED') { console.error(`Rule ${id} is not VALIDATED (run rules evaluate first); refusing to promote.`); return 4; }
+  const ir = loadIR(root, id);
+  if (!ir) { console.error(`No rule IR found for ${id}.`); return 4; }
+  const pub = publishCustomRule(root, ir);
+  if (!pub.ok) { console.error(`Promotion refused: ${pub.reason}`); return 4; }
+  const t = recordTransition(ledger, { ruleId: id, from: 'VALIDATED', to: 'PROMOTED', reason: args.flags.reason || 'promoted', actor: args.flags.actor || 'operator' });
+  saveLedger(root, ledger);
+  console.log(`Rule ${id}: VALIDATED -> PROMOTED. Published to ${pub.file}.`);
+  if (!t.ok) console.error(t.reason);
+  return 0;
+}
+
+async function cmdRollback(args, root) {
+  const id = args._[2];
+  if (!id) { console.error('usage: rules rollback <rule-id> [--reason text]'); return 4; }
+  const ledger = loadLedger(root);
+  const cur = stateOf(ledger, id);
+  if (cur !== 'PROMOTED') { console.error(`Rule ${id} is not PROMOTED (current: ${cur ?? 'unknown'}); nothing to roll back.`); return 4; }
+  const un = unpublishCustomRule(root, id);
+  const t = recordTransition(ledger, { ruleId: id, from: 'PROMOTED', to: 'ROLLED_BACK', reason: args.flags.reason || 'operator rollback', actor: args.flags.actor || 'operator' });
+  saveLedger(root, ledger);
+  console.log(`Rule ${id}: PROMOTED -> ROLLED_BACK.${un.ok ? '' : ` (warning: could not remove published rule file: ${un.reason})`}`);
+  return t.ok ? 0 : 1;
+}
+
+async function cmdStatus(args, root) {
+  const ledger = loadLedger(root);
+  const id = args._[2];
+  if (id) {
+    const h = historyFor(ledger, id);
+    if (!h.length) { console.error(`No history for rule ${id}.`); return 4; }
+    if (args.flags.json) console.log(JSON.stringify(h, null, 2));
+    else { console.log(`Rule ${id}: ${stateOf(ledger, id)}`); for (const e of h) console.log(`  ${e.at}  ${e.from ?? '(new)'} -> ${e.to}  (${e.reason || ''})`); }
+    return 0;
+  }
+  const ids = [...new Set(ledger.entries.map((e) => e.ruleId))];
+  if (args.flags.json) { console.log(JSON.stringify(ids.map((rid) => ({ id: rid, state: stateOf(ledger, rid) })), null, 2)); return 0; }
+  if (!ids.length) { console.log('No rules in the lifecycle ledger yet.'); return 0; }
+  for (const rid of ids) console.log(`  ${rid.padEnd(40)} ${stateOf(ledger, rid)}`);
+  return 0;
+}

package/src/posture/rule-compiler/corpus.js

@@ -0,0 +1,61 @@
+// Automatic corpus creation and scoring for ONE candidate rule.
+//
+// Positive corpus: the original vulnerable snippet plus its generated sibling variants (mutation.js).
+// Negative corpus: an explicit patched/clean variant the caller supplies, nearby clean patterns, protected
+// (sanitizer/guard) variants, and the NEGATIVE mutations of those (same mutators, applied to the safe side) — a
+// candidate rule must decline on ALL of them, not just the one hand-picked negative.
+//
+// Scoring reuses the exact runtime matcher (`runRule` in `../custom-rules.js`), never a second implementation —
+// the score a rule gets here is the score it will actually produce once promoted.
+import { generateVariants } from './mutation.js';
+import { compileToCustomRule } from './ir.js';
+import { runRule } from '../custom-rules.js';
+
+function toRuntimeRule(ir) {
+  const { ok, doc, errs } = compileToCustomRule(ir);
+  if (!ok) throw new Error(`cannot build corpus for an invalid rule IR: ${errs.join('; ')}`);
+  const regexes = doc.match.allOf ? doc.match.allOf.map((p) => new RegExp(p, 'g')) : [new RegExp(doc.match.pattern, 'g')];
+  return {
+    id: doc.id, title: doc.title, severity: doc.severity, cwe: doc.cwe, languages: doc.languages, message: doc.message,
+    regexes, requireAll: !!doc.match.allOf, windowLines: doc.match.window ?? 5,
+    notMatch: doc.match.notMatch ? new RegExp(doc.match.notMatch) : null,
+  };
+}
+
+const EXT = { javascript: 'x.js', typescript: 'x.ts', python: 'x.py', go: 'x.go', ruby: 'x.rb', java: 'x.java', rust: 'x.rs', csharp: 'x.cs', php: 'x.php', any: 'x.txt' };
+
+export function buildCorpus(ir, { originalCode, patchedCode = null, nearbyCleanCode = [], protectedVariantsCode = [] } = {}) {
+  const positives = [{ label: 'original', code: originalCode }];
+  for (const v of generateVariants(originalCode)) positives.push({ label: `sibling:${v.mutator}`, code: v.code });
+
+  const negatives = [];
+  if (patchedCode) {
+    negatives.push({ label: 'patched', code: patchedCode });
+    for (const v of generateVariants(patchedCode)) negatives.push({ label: `patched-sibling:${v.mutator}`, code: v.code });
+  }
+  nearbyCleanCode.forEach((code, i) => negatives.push({ label: `nearby-clean-${i}`, code }));
+  protectedVariantsCode.forEach((code, i) => negatives.push({ label: `protected-${i}`, code }));
+
+  return { positives, negatives };
+}
+
+export function scoreRule(ir, corpus) {
+  const rule = toRuntimeRule(ir);
+  const lang = ir.languages[0] || 'javascript';
+  const file = EXT[lang] || 'x.txt';
+  const run = (code) => runRule(rule, file, code).length > 0;
+
+  const posResults = corpus.positives.map((p) => ({ ...p, fired: run(p.code) }));
+  const negResults = corpus.negatives.map((n) => ({ ...n, fired: run(n.code) }));
+
+  const tp = posResults.filter((p) => p.fired).length;
+  const fn = posResults.length - tp;
+  const fp = negResults.filter((n) => n.fired).length;
+  const tn = negResults.length - fp;
+  const precision = tp + fp === 0 ? null : tp / (tp + fp);
+  const recall = corpus.positives.length === 0 ? null : tp / corpus.positives.length;
+  // "Protected negative failures" per PRD 3: a rule firing on a sanitizer/guard-carrying variant specifically.
+  const protectedFailures = negResults.filter((n) => n.label.startsWith('protected-') && n.fired).length;
+
+  return { precision, recall, tp, fp, tn, fn, protectedFailures, positives: posResults, negatives: negResults };
+}

package/src/posture/rule-compiler/eligibility.js

@@ -0,0 +1,25 @@
+// X-eligibility (PRD 3 "Eligibility"): a finding may enter synthesis only with a stable id, a vulnerable span, a
+// CWE/family, code-flow evidence, and deterministic corroboration OR exploit proof. Reuses the existing
+// vocabularies rather than inventing a third: proof-fabric's evidence-state ladder when present, else the older
+// proof-tier (`taint-proven`/`execution-proven`), else the finding's own `parser` (an IR-TAINT/MULTI-SINK parser
+// origin is code-flow evidence even with no execution).
+import { stateRank } from '../proof-fabric/states.js';
+
+const ANALYSED_PARSERS = new Set(['IR-TAINT', 'MULTI-SINK']);
+
+export function checkEligibility(finding, { proofState = null } = {}) {
+  const reasons = [];
+  if (!finding?.stableId && !finding?.id) reasons.push('no stable id');
+  if (finding?.file == null || finding?.line == null) reasons.push('no vulnerable span (file:line)');
+  if (!finding?.cwe && !finding?.family) reasons.push('no CWE or family');
+  const hasFlowEvidence = ANALYSED_PARSERS.has(finding?.parser) || !!finding?.snippet;
+  if (!hasFlowEvidence) reasons.push('no code-flow evidence (parser not IR-TAINT/MULTI-SINK, no snippet)');
+
+  let evidenceState = 'none';
+  if (proofState && stateRank(proofState) >= stateRank('EXPLOIT_CONFIRMED')) evidenceState = 'execution-proven';
+  else if (finding?.proofTier === 'execution-proven') evidenceState = 'execution-proven';
+  else if (finding?.proofTier === 'taint-proven' || ANALYSED_PARSERS.has(finding?.parser)) evidenceState = 'taint-proven';
+  if (evidenceState === 'none') reasons.push('no deterministic corroboration or exploit proof');
+
+  return { eligible: reasons.length === 0, reasons, evidenceState };
+}

package/src/posture/rule-compiler/ir.js

@@ -0,0 +1,87 @@
+// Rule intermediate representation: compiles DOWN into the existing pattern-rule engine
+// (`../custom-rules.js`), never a parallel matcher. "Compile" here means "emit a
+// custom-rules.js-compatible document" — detection at runtime is 100% the already-proven
+// regex/allOf/notMatch/window engine; this module only adds PROVENANCE metadata around it.
+//
+// A rule's match semantics are deliberately the same closed vocabulary custom-rules.js
+// already supports (pattern | allOf+window | notMatch) — no new matching primitive is
+// introduced, so a promoted rule is exactly as auditable as a hand-written one.
+import { sha256 } from '../proof-fabric/reproducer.js';
+
+const RULE_IR_SCHEMA = 'agentic-security/rule-ir';
+const RULE_IR_VERSION = 1;
+export const RULE_STATES = Object.freeze(['DRAFT', 'CANDIDATE', 'VALIDATED', 'PROMOTED', 'DEPRECATED', 'ROLLED_BACK']);
+const VALID_LANGS = new Set(['javascript', 'typescript', 'python', 'go', 'ruby', 'java', 'rust', 'csharp', 'php', 'any']);
+
+const isStr = (v) => typeof v === 'string' && v.length > 0;
+
+/** Problems with a draft rule IR; an empty list means it is well-formed (not yet validated/promoted). */
+export function validateRuleIR(ir) {
+  const errs = [];
+  if (!ir || typeof ir !== 'object') return ['rule IR must be an object'];
+  if (ir.schema !== RULE_IR_SCHEMA) errs.push('unknown schema');
+  if (ir.version !== RULE_IR_VERSION) errs.push('unsupported rule IR version');
+  if (!isStr(ir.id) || !/^[a-z][a-z0-9-]{2,79}$/.test(ir.id)) errs.push('id must be a lowercase kebab-case string');
+  if (!isStr(ir.title)) errs.push('title required');
+  if (!['critical', 'high', 'medium', 'low'].includes(ir.severity)) errs.push('severity must be critical|high|medium|low');
+  if (!isStr(ir.cwe) || !/^CWE-\d+$/.test(ir.cwe)) errs.push('cwe must look like CWE-###');
+  if (!isStr(ir.family)) errs.push('family required');
+  if (!Array.isArray(ir.languages) || !ir.languages.length || !ir.languages.every((l) => VALID_LANGS.has(l))) errs.push('languages must be a non-empty array from the supported set');
+  const m = ir.match;
+  if (!m || typeof m !== 'object') errs.push('match required');
+  else {
+    const hasPattern = isStr(m.pattern);
+    const hasAllOf = Array.isArray(m.allOf) && m.allOf.length > 0 && m.allOf.every(isStr);
+    if (!hasPattern && !hasAllOf) errs.push('match needs a pattern or a non-empty allOf');
+    if (m.notMatch != null && !isStr(m.notMatch)) errs.push('notMatch must be a string');
+    if (m.window != null && !(Number.isInteger(m.window) && m.window > 0 && m.window <= 200)) errs.push('window must be an integer in 1..200');
+    for (const re of [m.pattern, ...(m.allOf || []), m.notMatch].filter(isStr)) {
+      try { new RegExp(re); } catch { errs.push(`invalid regex: ${re}`); }
+    }
+  }
+  // Declarative-only hint fields (sink/source/propagator/sanitizer/dominatingGuard/frameworkConstraint): free text,
+  // never executable, carried for provenance/review only and NOT consulted by the matcher.
+  for (const k of ['sink', 'source', 'propagator', 'sanitizer', 'dominatingGuard', 'frameworkConstraint']) {
+    if (ir[k] != null && !isStr(ir[k])) errs.push(`${k} must be a string when present`);
+  }
+  if (ir.exclusions != null && !(Array.isArray(ir.exclusions) && ir.exclusions.every(isStr))) errs.push('exclusions must be an array of strings');
+  if (!ir.provenance || typeof ir.provenance !== 'object') errs.push('provenance required');
+  else {
+    if (!isStr(ir.provenance.originatingFindingId)) errs.push('provenance.originatingFindingId required');
+    if (ir.provenance.originatingCommit != null && !isStr(ir.provenance.originatingCommit)) errs.push('provenance.originatingCommit must be a string');
+    if (!['execution-proven', 'taint-proven', 'deterministic-corroboration'].includes(ir.provenance.evidenceState)) errs.push('provenance.evidenceState must be execution-proven|taint-proven|deterministic-corroboration');
+  }
+  return errs;
+}
+
+export function ruleIRDigest(ir) {
+  const { state: _s, ...rest } = ir;
+  return sha256(JSON.stringify(rest, Object.keys(rest).sort()));
+}
+
+/** Build a DRAFT rule IR. Caller supplies the candidate pattern — this module never invents one. */
+export function draftRuleIR({ id, title, severity, cwe, family, languages, match, remediation, originatingFindingId, originatingCommit, evidenceState, model = null, promptConfigVersion = null, ...hints }) {
+  const ir = {
+    schema: RULE_IR_SCHEMA, version: RULE_IR_VERSION, id, title, severity, cwe, family, languages, match,
+    remediation: remediation ?? null, state: 'DRAFT',
+    ...hints,
+    provenance: { originatingFindingId, originatingCommit: originatingCommit ?? null, evidenceState, model, promptConfigVersion },
+  };
+  return ir;
+}
+
+/** Compile a rule IR into a custom-rules.js-compatible document. Caller writes the YAML. */
+export function compileToCustomRule(ir) {
+  const errs = validateRuleIR(ir);
+  if (errs.length) return { ok: false, errs };
+  const doc = {
+    id: `rule-compiler/${ir.id}`,
+    title: ir.title,
+    severity: ir.severity,
+    cwe: ir.cwe,
+    languages: ir.languages,
+    match: { ...(ir.match.pattern ? { pattern: ir.match.pattern } : {}), ...(ir.match.allOf ? { allOf: ir.match.allOf, window: ir.match.window ?? 5 } : {}), ...(ir.match.notMatch ? { notMatch: ir.match.notMatch } : {}) },
+    message: ir.remediation || `${ir.title} (${ir.cwe}, compiled from ${ir.provenance.originatingFindingId})`,
+  };
+  return { ok: true, doc };
+}

package/src/posture/rule-compiler/lifecycle.js

@@ -0,0 +1,57 @@
+// Rule lifecycle ledger: DRAFT -> CANDIDATE -> VALIDATED -> PROMOTED -> DEPRECATED/ROLLED_BACK.
+//
+// Append-only, hash-chained (same idiom as remediation-ledger.js / invariants/lifecycle.js): each entry links the
+// previous entry's hash, so a dropped or edited middle record breaks the chain and `verifyLedger` catches it. A
+// rule cannot promote itself: `transition` only records a transition a caller already decided; `advance`-style
+// guarding on VALID edges lives here so no caller can jump DRAFT straight to PROMOTED.
+import { createHash } from 'node:crypto';
+import { RULE_STATES } from './ir.js';
+
+export const LEDGER_SCHEMA = 'agentic-security/rule-lifecycle-ledger';
+const EDGES = Object.freeze({
+  DRAFT: ['CANDIDATE'], CANDIDATE: ['VALIDATED', 'DRAFT'], VALIDATED: ['PROMOTED', 'DRAFT'],
+  PROMOTED: ['DEPRECATED', 'ROLLED_BACK'], DEPRECATED: [], ROLLED_BACK: ['CANDIDATE'],
+});
+
+const hashOf = (entry) => createHash('sha256').update(JSON.stringify(entry)).digest('hex');
+
+export function newLedger() { return { schema: LEDGER_SCHEMA, entries: [] }; }
+
+function currentState(ledger, ruleId) {
+  for (let i = ledger.entries.length - 1; i >= 0; i--) if (ledger.entries[i].ruleId === ruleId) return ledger.entries[i].to;
+  return null;
+}
+
+/**
+ * Record a transition. Refuses an edge not in EDGES, and refuses DRAFT->DRAFT re-entry for a rule that has never
+ * existed (the first entry for a rule id must be `from: null, to: 'DRAFT'`).
+ */
+export function recordTransition(ledger, { ruleId, from, to, reason, evidence = null, actor }) {
+  if (!RULE_STATES.includes(to)) return { ok: false, reason: `unknown target state: ${to}` };
+  const cur = currentState(ledger, ruleId);
+  if (cur === null) {
+    if (from !== null || to !== 'DRAFT') return { ok: false, reason: 'first transition for a rule must be null -> DRAFT' };
+  } else {
+    if (from !== cur) return { ok: false, reason: `stale transition: current state is ${cur}, not ${from}` };
+    if (!EDGES[cur].includes(to)) return { ok: false, reason: `cannot go from ${cur} to ${to}` };
+  }
+  const prevHash = ledger.entries.length ? ledger.entries[ledger.entries.length - 1].entryHash : null;
+  const entry = { ruleId, from: cur, to, reason: reason ?? null, evidence, actor: actor ?? 'unknown', at: new Date().toISOString(), prevHash };
+  entry.entryHash = hashOf({ ...entry, entryHash: undefined });
+  ledger.entries.push(entry);
+  return { ok: true, entry };
+}
+
+export function verifyLedger(ledger) {
+  let prevHash = null;
+  for (const e of ledger.entries) {
+    if (e.prevHash !== prevHash) return { ok: false, reason: 'chain broken: prevHash mismatch' };
+    const recomputed = hashOf({ ...e, entryHash: undefined });
+    if (recomputed !== e.entryHash) return { ok: false, reason: `entry hash mismatch for ${e.ruleId} at ${e.at}` };
+    prevHash = e.entryHash;
+  }
+  return { ok: true };
+}
+
+export function historyFor(ledger, ruleId) { return ledger.entries.filter((e) => e.ruleId === ruleId); }
+export function stateOf(ledger, ruleId) { return currentState(ledger, ruleId); }

package/src/posture/rule-compiler/mutation.js

@@ -0,0 +1,60 @@
+// Bounded structural mutation for ONE candidate rule's own corpus — not bench/mutation's engine-wide verdict-flip
+// harness (a different scope: the whole deterministic+taint engine against curated adversarial cases). This module
+// exists only to answer PRD 3's "unseen sibling" question for a single rule: does the candidate still fire on a
+// structurally different variant of the same vulnerable shape, and does it still decline on a structurally
+// different variant of a mitigated shape.
+//
+// Mutations are SYNTACTIC, deterministic, seed-free and language-agnostic: variable rename, wrap-in-function,
+// reorder-adjacent-statements, insert-indirection (a pass-through local variable). No semantics are changed on
+// purpose — a mutation that silently changed behavior would make a "same vulnerability, different shape" claim
+// false.
+const RESERVED = new Set(['if', 'else', 'for', 'while', 'return', 'function', 'const', 'let', 'var', 'import', 'from', 'def', 'class', 'req', 'res']);
+
+function identifiers(code) {
+  const seen = new Set();
+  for (const m of code.matchAll(/\b([A-Za-z_][A-Za-z0-9_]*)\b/g)) if (!RESERVED.has(m[1]) && !/^[A-Z_]+$/.test(m[1])) seen.add(m[1]);
+  return [...seen];
+}
+
+function renameOne(code, from, to) {
+  if (!from || code.includes(to)) return null;
+  return code.replace(new RegExp(`\\b${from}\\b`, 'g'), to);
+}
+
+export function variableRename(code, { suffix = '_v2' } = {}) {
+  const ids = identifiers(code);
+  const target = ids.find((id) => id.length > 1);
+  if (!target) return null;
+  return renameOne(code, target, `${target}${suffix}`);
+}
+
+export function wrapInFunction(code, { name = 'wrapped_v2' } = {}) {
+  const indented = code.split('\n').map((l) => (l ? `  ${l}` : l)).join('\n');
+  return `function ${name}() {\n${indented}\n}\n${name}();\n`;
+}
+
+/** Insert a pass-through local for the first simple right-hand-side identifier assigned with `=` (indirection,
+ * no behavior change): `x = y;` becomes `const indirectN = y; x = indirectN;`. */
+function insertIndirection(code) {
+  const m = /^(\s*)([A-Za-z_][A-Za-z0-9_.]*)\s*=\s*([A-Za-z_][A-Za-z0-9_.]*)\s*;/m.exec(code);
+  if (!m) return null;
+  const [whole, indent, lhs, rhs] = m;
+  const varName = `indirect_${Math.abs(_hash(rhs)) % 1000}`;
+  if (code.includes(varName)) return null;
+  return code.replace(whole, `${indent}const ${varName} = ${rhs};\n${indent}${lhs} = ${varName};`);
+}
+
+function _hash(s) { let h = 0; for (const c of s) h = (h * 31 + c.charCodeAt(0)) | 0; return h; }
+
+const MUTATORS = Object.freeze({ 'variable-rename': variableRename, 'wrap-in-function': wrapInFunction, 'insert-indirection': insertIndirection });
+
+/** Every mutator applied to `code`; mutators that don't apply (return null) are omitted, never silently skipped without saying so. */
+export function generateVariants(code) {
+  const out = [];
+  for (const [name, fn] of Object.entries(MUTATORS)) {
+    let variant = null;
+    try { variant = fn(code); } catch { variant = null; }
+    if (variant && variant !== code) out.push({ mutator: name, code: variant });
+  }
+  return out;
+}

package/src/posture/rule-compiler/promotion.js

@@ -0,0 +1,47 @@
+// Promotion gate (PRD 3 "Promotion Gate"): precision/recall thresholds, zero protected-negative failures, zero
+// benchmark leakage. A candidate rule does NOT automatically become a production rule — `evaluateRule` only
+// scores; `gatePromotion` only judges; the lifecycle ledger (`lifecycle.js`) is the one place a PROMOTED state is
+// actually recorded, and nothing in this module writes to it.
+import { scoreRule } from './corpus.js';
+
+export const PROMOTION_DEFAULTS = Object.freeze({ minPrecision: 0.95, minRecall: 0.80, maxProtectedFailures: 0 });
+
+// Deliberately independent of posture/evaluation/custody.js: that module's term list and guard are reserved for
+// the evaluation custodian's sealed-label boundary (QA-002.AC03), and a test enforces that nothing outside
+// posture/evaluation/ may import it. The anti-cheating need here is smaller and different in kind — a rule
+// synthesis prompt/corpus must not embed a benchmark name or an advisory id — so it gets its own small, independent
+// checker rather than reaching across that isolation boundary.
+const ADVISORY_ID = /\b(?:CVE|GHSA|CWE)-[A-Za-z0-9-]+\b/i;
+
+/** Anti-cheating: the synthesis agent must never see benchmark answer keys, hidden labels, CWE-bearing hidden
+ * filenames or vulnerability-comment metadata. */
+export function checkLeakage(promptText, { benchmarkNames = [] } = {}) {
+  const hay = String(promptText ?? '');
+  const offending = [];
+  for (const name of benchmarkNames) if (name && hay.toLowerCase().includes(String(name).toLowerCase())) offending.push(name);
+  const m = ADVISORY_ID.exec(hay);
+  if (m) offending.push(m[0]);
+  return { ok: offending.length === 0, offendingTerms: offending };
+}
+
+export function evaluateRule(ir, corpus, thresholds = PROMOTION_DEFAULTS) {
+  const score = scoreRule(ir, corpus);
+  const t = { ...PROMOTION_DEFAULTS, ...thresholds };
+  const reasons = [];
+  if (score.precision === null) reasons.push('precision unmeasured (no positive or negative corpus fired)');
+  else if (score.precision < t.minPrecision) reasons.push(`precision ${(score.precision * 100).toFixed(1)}% below ${(t.minPrecision * 100).toFixed(0)}%`);
+  if (score.recall === null) reasons.push('recall unmeasured (empty positive corpus)');
+  else if (score.recall < t.minRecall) reasons.push(`recall ${(score.recall * 100).toFixed(1)}% below ${(t.minRecall * 100).toFixed(0)}%`);
+  if (score.protectedFailures > t.maxProtectedFailures) reasons.push(`${score.protectedFailures} protected-negative failure(s) (max ${t.maxProtectedFailures})`);
+  return { pass: reasons.length === 0, reasons, score, thresholds: t };
+}
+
+/** The one promotion decision. Never auto-applied: a caller still has to act on `pass:true` by recording a
+ * lifecycle transition. Localization ("acceptable") is a human/caller judgment this function cannot make from
+ * line/column alone; it is surfaced for review, never silently assumed. */
+export function gatePromotion({ ir, corpus, leakageCheck, thresholds } = {}) {
+  const evalResult = evaluateRule(ir, corpus, thresholds);
+  const reasons = [...evalResult.reasons];
+  if (!leakageCheck || leakageCheck.ok !== true) reasons.push(leakageCheck ? `leakage check failed: ${leakageCheck.offendingTerms.join(', ')}` : 'leakage check was not run');
+  return { pass: reasons.length === 0, reasons, evaluation: evalResult };
+}