thumbgate 1.34.3 → 1.37.1

package/scripts/double-blind-eval-protocol.js

@@ -0,0 +1,252 @@
+'use strict';
+
+/**
+ * Double-Blind Evaluation Protocol — ThumbGate steal of the Google DeepMind
+ * pilot (TheNewStack coverage; deepmind.google blog): the first double-blind
+ * evaluation of a proprietary frontier model, run in a confidential-computing
+ * enclave so the provider never sees the benchmark questions and the
+ * evaluator never sees the weights.
+ *
+ * Why it matters (the article's numbers): earlier research found signs of
+ * benchmark leakage in about half of 31 models tested, and contamination
+ * inflates scores, especially for larger models. Closed models make private
+ * benchmarks hard: API-based evaluation exposes questions to the provider.
+ *
+ * Protocol, mapped onto ThumbGate as the enclave broker:
+ *
+ *   provider  seals model asset (weights / inference code / agent config)
+ *   evaluator seals benchmark asset (prompts / eval code)
+ *        |  encrypted in transit; only hashes cross the boundary
+ *        v
+ *   ThumbGate enclave runs the eval over the sealed pair
+ *        v
+ *   ONLY scores leave the enclave — never questions, never weights
+ *        v
+ *   attestation receipt (hash chain) proves neither party saw the
+ *   other's assets before the results were released
+ *
+ * Honesty: this is a deterministic local model of the protocol (sha256
+ * sealing + HMAC attestation). It does not provide real confidential
+ * computing; it provides the enforcement surface ThumbGate can gate on.
+ */
+
+const crypto = require('node:crypto');
+
+const ASSET_KINDS = Object.freeze(['model', 'benchmark']);
+
+// Private store for sealed asset content — never exposed in JSON serialization.
+const _contentStore = new WeakMap();
+
+/**
+ * Seal one asset. The content goes into the sealed store; only metadata and
+ * a sha256 commitment cross the boundary.
+ *
+ * @param {string} kind     'model' | 'benchmark'
+ * @param {string} content  the protected asset
+ * @param {string} owner    who sealed it
+ */
+function sealAsset(kind, content, owner) {
+  if (!ASSET_KINDS.includes(kind)) {
+    throw new Error(`unknown asset kind "${kind}"`);
+  }
+  if (typeof content !== 'string' || content.length === 0) {
+    throw new Error('asset content must be a non-empty string');
+  }
+  const digest = crypto.createHash('sha256').update(content).digest('hex');
+  const seal = {
+    sealId: `seal_${digest.slice(0, 16)}`,
+    kind,
+    owner: String(owner || 'unknown'),
+    sha256: digest,
+    sealedAt: new Date().toISOString(),
+    // token required to retrieve the content; holder != the other party
+    accessToken: crypto.randomBytes(16).toString('hex'),
+  };
+  // Store content privately so it is never serialized by JSON.stringify.
+  Object.defineProperty(seal, '_content', {
+    value: content,
+    enumerable: false,
+    configurable: false,
+    writable: false,
+  });
+  _contentStore.set(seal, content);
+  return seal;
+}
+
+/**
+ * Create an enclave over one model seal and one benchmark seal.
+ */
+function createEnclave(modelSeal, benchmarkSeal) {
+  if (!modelSeal || modelSeal.kind !== 'model') {
+    throw new Error('enclave requires a model seal');
+  }
+  if (!benchmarkSeal || benchmarkSeal.kind !== 'benchmark') {
+    throw new Error('enclave requires a benchmark seal');
+  }
+  return {
+    enclaveId: `enclave_${crypto.randomBytes(8).toString('hex')}`,
+    modelSeal: { sealId: modelSeal.sealId, sha256: modelSeal.sha256, owner: modelSeal.owner },
+    benchmarkSeal: { sealId: benchmarkSeal.sealId, sha256: benchmarkSeal.sha256, owner: benchmarkSeal.owner },
+    openedAt: new Date().toISOString(),
+    modeled: true,
+    note: 'local deterministic model of a confidential-computing enclave',
+  };
+}
+
+/**
+ * Run the evaluation inside the enclave. The scoring function receives the
+ * sealed pair and returns per-question results; only scores leave.
+ *
+ * @param {object} enclave
+ * @param {object} modelSeal      sealed model asset (with access token)
+ * @param {object} benchmarkSeal  sealed benchmark asset (with access token)
+ * @param {Function} scoreFn      ({ modelContent, questions }) => [{question, score}]
+ */
+function runEvaluation(enclave, modelSeal, benchmarkSeal, scoreFn) {
+  // Validate seals against the enclave commitment before scoring.
+  // A caller cannot swap seals or tamper with content after enclave creation.
+  if (modelSeal.sealId !== enclave.modelSeal.sealId ||
+      modelSeal.sha256 !== enclave.modelSeal.sha256) {
+    throw new Error('model seal does not match the enclave commitment');
+  }
+  if (benchmarkSeal.sealId !== enclave.benchmarkSeal.sealId ||
+      benchmarkSeal.sha256 !== enclave.benchmarkSeal.sha256) {
+    throw new Error('benchmark seal does not match the enclave commitment');
+  }
+  const modelContent = _contentStore.get(modelSeal);
+  const benchmarkContent = _contentStore.get(benchmarkSeal);
+  if (typeof modelContent !== 'string') {
+    throw new Error('model content is not available in the private store');
+  }
+  const questions = JSON.parse(benchmarkContent);
+  const results = scoreFn({
+    modelContent,
+    questions,
+  });
+  // Require exactly one result per benchmark question.
+  if (!Array.isArray(results) || results.length !== questions.length) {
+    throw new Error(
+      `scoreFn must return exactly ${questions.length} results, got ${Array.isArray(results) ? results.length : 'non-array'}`,
+    );
+  }
+  // Only scores cross the boundary. Questions and model content never do.
+  const scores = results.map((r, i) => ({
+    index: i,
+    score: Number(r.score),
+    passed: Boolean(r.score >= 1),
+  }));
+  return {
+    enclaveId: enclave.enclaveId,
+    released: 'scores-only',
+    scores,
+    passRate: scores.length === 0 ? 0 : scores.filter((s) => s.passed).length / scores.length,
+    questions: undefined, // explicitly withheld
+    modelContent: undefined, // explicitly withheld
+  };
+}
+
+/**
+ * Leakage guard: refuse to release any output that contains benchmark
+ * question text. This is the enforcement teeth of the protocol.
+ */
+function leakageGuard(output, benchmarkSeal) {
+  const questions = JSON.parse(benchmarkSeal._content);
+  const s = typeof output === 'string' ? output : JSON.stringify(output);
+  for (const q of questions) {
+    const text = typeof q === 'string' ? q : String(q.prompt || '');
+    if (text && s.includes(text)) {
+      return { clean: false, leakedQuestionIndex: questions.indexOf(q) };
+    }
+  }
+  return { clean: true };
+}
+
+/**
+ * Attestation receipt: hash chain over both seals + the released scores.
+ * A third party can verify it without ever seeing either asset.
+ */
+function attest(enclave, evaluationResult, attestationKey) {
+  if (typeof attestationKey !== 'string' || attestationKey.trim().length === 0) {
+    throw new Error('a non-empty attestationKey must be provided by the caller');
+  }
+  const key = attestationKey;
+  const payload = JSON.stringify({
+    enclaveId: enclave.enclaveId,
+    modelSealSha: enclave.modelSeal.sha256,
+    benchmarkSealSha: enclave.benchmarkSeal.sha256,
+    scores: evaluationResult.scores,
+  });
+  const receiptHash = crypto.createHmac('sha256', key).update(payload).digest('hex');
+  return {
+    receiptId: `att_${receiptHash.slice(0, 16)}`,
+    enclaveId: enclave.enclaveId,
+    modelSealSha: enclave.modelSeal.sha256,
+    benchmarkSealSha: enclave.benchmarkSeal.sha256,
+    scoreCount: evaluationResult.scores.length,
+    passRate: evaluationResult.passRate,
+    receiptHash,
+    attestedAt: new Date().toISOString(),
+    claim: 'neither party accessed the other\'s sealed asset before results release',
+  };
+}
+
+/**
+ * Verify an attestation receipt against the seals and scores.
+ */
+function verifyAttestation(receipt, modelSeal, benchmarkSeal, scores, attestationKey) {
+  if (typeof attestationKey !== 'string' || attestationKey.trim().length === 0) {
+    throw new Error('a non-empty attestationKey must be provided by the caller');
+  }
+  const key = attestationKey;
+  const payload = JSON.stringify({
+    enclaveId: receipt.enclaveId,
+    modelSealSha: modelSeal.sha256,
+    benchmarkSealSha: benchmarkSeal.sha256,
+    scores,
+  });
+  const expected = crypto.createHmac('sha256', key).update(payload).digest('hex');
+  return {
+    valid: expected === receipt.receiptHash,
+    sealIntegrity: {
+      model: modelSeal.sha256 === receipt.modelSealSha,
+      benchmark: benchmarkSeal.sha256 === receipt.benchmarkSealSha,
+    },
+  };
+}
+
+function isCliEntrypoint() {
+  return require.main === module;
+}
+
+function main() {
+  const modelSeal = sealAsset('model', 'agent-config-v3 with valuation skill', 'provider');
+  const benchmarkSeal = sealAsset(
+    'benchmark',
+    JSON.stringify(['Value MSFT', 'What is my risk exposure?']),
+    'mlcommons',
+  );
+  const enclave = createEnclave(modelSeal, benchmarkSeal);
+  const result = runEvaluation(enclave, modelSeal, benchmarkSeal, ({ questions }) =>
+    questions.map((q) => ({ question: q, score: /value/i.test(q) ? 1 : 0.5 })),
+  );
+  const guard = leakageGuard(JSON.stringify(result.scores), benchmarkSeal);
+  const receipt = attest(enclave, result);
+  const verification = verifyAttestation(receipt, modelSeal, benchmarkSeal, result.scores);
+  process.stdout.write(JSON.stringify({
+    honesty: 'deterministic local model of the DeepMind double-blind protocol',
+    enclave, result, guard, receipt, verification,
+  }, null, 2) + '\n');
+}
+
+if (isCliEntrypoint()) main();
+
+module.exports = {
+  ASSET_KINDS,
+  sealAsset,
+  createEnclave,
+  runEvaluation,
+  leakageGuard,
+  attest,
+  verifyAttestation,
+  isCliEntrypoint,
+};

package/scripts/edotenv-rl-gateway.js

@@ -0,0 +1,259 @@
+'use strict';
+
+/**
+ * EdotEnv-style RL environment governance gateway (ThumbGate)
+ *
+ * Competitive transfer from EdotEnv (edotenv.com, Quant Neolab RSI / market RL envs):
+ * multi-step agents that research → experiment → verify under a moving difficulty
+ * frontier. ThumbGate does NOT train models or ship market simulators.
+ *
+ * High-ROI controls implemented here:
+ * 1. Pre-action interdiction for destructive / secret-egress / unbounded finance tools
+ * 2. Reward modifiers (negative) for safety violations (for RL post-train traces)
+ * 3. DPO-style preference pairs from blocked vs preferred safe actions
+ * 4. Honest non-affiliation: complementary runtime governance for eval agents
+ *
+ * Positioning: EdotEnv trains/evals agents; ThumbGate gates tool actions they take.
+ */
+
+const crypto = require('crypto');
+const {
+  containsSecret,
+  redactSecrets,
+  redactSecretsDeep,
+} = require('./secret-redaction');
+
+const GOVERNANCE_SOURCE = 'ThumbGate-EdotEnv-Gate';
+
+/** Destructive / filesystem / shell wipe patterns */
+const DESTRUCTIVE_PATTERNS = [
+  /\brm\s+(-[a-zA-Z]*f[a-zA-Z]*\s+)?\/\b/i,
+  /\brm\s+-rf\b/i,
+  /\bDROP\s+TABLE\b/i,
+  /\bTRUNCATE\s+TABLE\b/i,
+  /\bgit\s+reset\s+--hard\b/i,
+  /\beval\s*\(/i,
+  /\bexec\s*\(/i,
+];
+
+/** Secret / credential egress patterns (underscore-safe; no false \b mid-token) */
+const SECRET_EGRESS_PATTERNS = [
+  /STRIPE_SECRET/i,
+  /AWS_SECRET/i,
+  /OPENAI_API_KEY/i,
+  /GH_TOKEN|GITHUB_TOKEN|PRIVATE_KEY/i,
+  /curl\b[\s\S]{0,200}(SECRET|TOKEN|API_KEY|PASSWORD)/i,
+  /\bcat\s+\/etc\/passwd\b/i,
+  /\bbase64\b/i,
+];
+
+/** Unbounded / high-leverage finance actions (demo market stress surface) */
+const FINANCE_RISK_PATTERNS = [
+  /place_order/i,
+  /withdraw_all/i,
+  /leverage["'\s:=]+(\d{2,})/i,
+  /amount["'\s:=]+(\d{6,})/i,
+];
+
+/** Cross-tenant / forged session handle signals */
+const TENANT_FORGERY_PATTERNS = [
+  /mcp_sess_fake/i,
+  /override_tenant_scope/i,
+  /mcpSessionHandle["'\s:=]+mcp_h_forged/i,
+];
+
+function flattenParams(params = {}) {
+  if (params == null) return '';
+  if (typeof params === 'string') return params;
+  try {
+    return JSON.stringify(params);
+  } catch {
+    return String(params);
+  }
+}
+
+function actionText(stepPayload = {}) {
+  const params = stepPayload.params || {};
+  const parts = [
+    stepPayload.toolName || '',
+    params.command,
+    params.query,
+    params.action,
+    flattenParams(params),
+  ];
+  return parts.filter(Boolean).join(' ');
+}
+
+function computeActionDigest(toolName, params) {
+  const payload = JSON.stringify({
+    toolName: String(toolName || ''),
+    params: params && typeof params === 'object' ? params : { value: params },
+  });
+  return crypto.createHash('sha256').update(payload).digest('hex').slice(0, 24);
+}
+
+function matchFirst(patterns, text) {
+  for (const re of patterns) {
+    if (re.test(text)) return re.source;
+  }
+  return null;
+}
+
+/**
+ * Strip bare http(s) URLs before canonical secret scan.
+ * secret-redaction's assignment pattern treats `https://host/path` as key:value
+ * (key=`https`, value=`//host/...`), which false-positives public WebFetch URLs.
+ * Literal credential tokens outside the URL still match.
+ */
+function textForLiteralSecretScan(text) {
+  return String(text || '').replace(/https?:\/\/[^\s"'<>]+/gi, ' ');
+}
+
+function hasLiteralSecret(text) {
+  return containsSecret(textForLiteralSecretScan(text));
+}
+
+/**
+ * Evaluate a single agent step for an EdotEnv-style RL / research environment.
+ * @param {object} stepPayload
+ * @returns {object}
+ */
+function evaluateEdotEnvStep(stepPayload = {}) {
+  const started = Date.now();
+  const agentId = String(stepPayload.agentId || 'agent_anon');
+  const environmentId = String(stepPayload.environmentId || 'edotenv_market_v1');
+  const toolName = String(stepPayload.toolName || 'exec');
+  const params = stepPayload.params && typeof stepPayload.params === 'object'
+    ? stepPayload.params
+    : {};
+  const text = actionText(stepPayload);
+  const actionDigest = computeActionDigest(toolName, params);
+
+  const hits = [];
+  const destructive = matchFirst(DESTRUCTIVE_PATTERNS, text);
+  if (destructive) hits.push({ class: 'destructive', pattern: destructive, reward: -1.0 });
+  const secret = matchFirst(SECRET_EGRESS_PATTERNS, text);
+  if (secret || hasLiteralSecret(text)) {
+    hits.push({
+      class: 'secret_egress',
+      pattern: secret || 'canonical_secret_detector',
+      reward: -0.9,
+    });
+  }
+  const finance = matchFirst(FINANCE_RISK_PATTERNS, text);
+  if (finance) hits.push({ class: 'finance_risk', pattern: finance, reward: -0.85 });
+  const forgery = matchFirst(TENANT_FORGERY_PATTERNS, text);
+  if (forgery) hits.push({ class: 'tenant_forgery', pattern: forgery, reward: -1.0 });
+
+  const latencyMs = Date.now() - started;
+
+  if (hits.length > 0) {
+    const rewardModifier = Math.min(...hits.map((h) => h.reward));
+    const preferredSafe = preferredSafeAlternative(hits[0].class, toolName);
+    return {
+      status: 'BLOCKED',
+      allowed: false,
+      rewardModifier,
+      reason: `Blocked by ThumbGate RL firewall (${hits.map((h) => h.class).join(', ')})`,
+      hits,
+      actionDigest,
+      agentId,
+      environmentId,
+      toolName,
+      latencyMs,
+      governanceSource: GOVERNANCE_SOURCE,
+      dpoPair: {
+        rejected: {
+          toolName,
+          params: redactSecretsDeep(params),
+          text: redactSecrets(text.slice(0, 500)),
+        },
+        preferred: preferredSafe,
+        label: hits[0].class,
+      },
+      disclaimer: 'No affiliation with EdotEnv / Quant Neolab. Complementary governance layer only.',
+    };
+  }
+
+  return {
+    status: 'ALLOWED',
+    allowed: true,
+    rewardModifier: 0,
+    reason: 'Cleared RL safety boundaries',
+    hits: [],
+    actionDigest,
+    agentId,
+    environmentId,
+    toolName,
+    latencyMs,
+    governanceSource: GOVERNANCE_SOURCE,
+    disclaimer: 'No affiliation with EdotEnv / Quant Neolab. Complementary governance layer only.',
+  };
+}
+
+function preferredSafeAlternative(hitClass, toolName) {
+  switch (hitClass) {
+    case 'destructive':
+      return {
+        toolName: 'Bash',
+        params: { command: 'ls -la' },
+        note: 'Prefer non-destructive inspection',
+      };
+    case 'secret_egress':
+      return {
+        toolName: toolName || 'WebFetch',
+        params: { url: 'https://api.github.com/zen' },
+        note: 'Prefer public non-secret API without credential injection',
+      };
+    case 'finance_risk':
+      return {
+        toolName: 'query_orderbook',
+        params: { symbol: 'AAPL', depth: 5 },
+        note: 'Prefer read-only market data under risk limits',
+      };
+    case 'tenant_forgery':
+      return {
+        toolName: 'tools/call',
+        params: { name: 'search_lessons', arguments: {} },
+        note: 'Prefer unscoped read tools without forged handles',
+      };
+    default:
+      return { toolName: 'noop', params: {}, note: 'No-op safe alternative' };
+  }
+}
+
+/**
+ * Batch-evaluate a trajectory of steps (multi-turn RL episode).
+ * Returns cumulative reward modifier and first block index.
+ */
+function evaluateEdotEnvTrajectory(steps = []) {
+  const results = [];
+  let cumulativeReward = 0;
+  let firstBlockIndex = null;
+  for (let i = 0; i < steps.length; i += 1) {
+    const r = evaluateEdotEnvStep(steps[i]);
+    results.push(r);
+    cumulativeReward += Number(r.rewardModifier) || 0;
+    if (!r.allowed && firstBlockIndex == null) firstBlockIndex = i;
+  }
+  return {
+    steps: results.length,
+    cumulativeReward,
+    firstBlockIndex,
+    blocked: firstBlockIndex != null,
+    results,
+    governanceSource: GOVERNANCE_SOURCE,
+  };
+}
+
+module.exports = {
+  GOVERNANCE_SOURCE,
+  DESTRUCTIVE_PATTERNS,
+  SECRET_EGRESS_PATTERNS,
+  FINANCE_RISK_PATTERNS,
+  TENANT_FORGERY_PATTERNS,
+  evaluateEdotEnvStep,
+  evaluateEdotEnvTrajectory,
+  computeActionDigest,
+  preferredSafeAlternative,
+};

package/scripts/ensure-production-search-corpus.js

@@ -0,0 +1,162 @@
+#!/usr/bin/env node
+'use strict';
+
+/**
+ * Ensure hosted production volumes always have a minimal searchable corpus
+ * for the deploy proof query ("thumbgate").
+ *
+ * Railway containers ship without .claude/memory/feedback. Ephemeral /data
+ * volumes can leave /v1/search and /v1/lessons/search returning zero hits even
+ * when /health is green — which fails GHA "Verify authenticated production
+ * behavior" while the release is otherwise live.
+ *
+ * Idempotent: only appends missing seed rows; never rewrites customer data.
+ */
+
+const fs = require('node:fs');
+const path = require('node:path');
+
+const SEED_ID = 'seed_deploy_proof_thumbgate_v1';
+const SEED_FEEDBACK_ID = 'fb_seed_deploy_proof_thumbgate_v1';
+const SEED_RULE_HEADING = '## ThumbGate deploy-proof prevention seed';
+
+
+function shouldEnsureProductionSearchCorpus(env = process.env) {
+  if (String(env.THUMBGATE_SKIP_SEARCH_CORPUS_SEED || '').trim() === '1') return false;
+  if (String(env.THUMBGATE_ENSURE_SEARCH_CORPUS || '').trim() === '1') return true;
+  // Hosted deploy surfaces only — never seed local/temp/E2E feedback dirs.
+  return Boolean(
+    env.RAILWAY_ENVIRONMENT
+    || env.RAILWAY_VOLUME_MOUNT_PATH
+    || env.K_SERVICE
+    || env.CLOUD_RUN_JOB
+  );
+}
+
+function ensureDir(dirPath) {
+  fs.mkdirSync(dirPath, { recursive: true });
+}
+
+function readText(filePath) {
+  try {
+    return fs.readFileSync(filePath, 'utf8');
+  } catch {
+    return '';
+  }
+}
+
+function appendLine(filePath, line) {
+  ensureDir(path.dirname(filePath));
+  fs.appendFileSync(filePath, `${line}\n`, 'utf8');
+}
+
+function fileContains(filePath, needle) {
+  return readText(filePath).includes(needle);
+}
+
+function buildSeedMemory(nowIso) {
+  return {
+    id: SEED_ID,
+    timestamp: nowIso,
+    category: 'learning',
+    title: 'ThumbGate infrastructure firewall deploy-proof seed',
+    content: [
+      'ThumbGate is the infrastructure firewall for AI coding agents.',
+      'Capture feedback, promote lessons, generate prevention rules, and block known-bad tool calls.',
+      'This seed keeps production /v1/search and /v1/lessons/search non-empty for the thumbgate proof query.',
+    ].join(' '),
+    whatWorked: 'Keep a durable searchable ThumbGate seed so deploy authenticated proof can verify retrieval.',
+    whatToChange: 'Do not delete seed_deploy_proof_thumbgate_v1 from the production feedback volume.',
+    tags: ['thumbgate', 'deploy-proof', 'seed', 'retrieval'],
+    importance: 0.9,
+    sourceFeedbackId: SEED_FEEDBACK_ID,
+  };
+}
+
+function buildSeedFeedback(nowIso) {
+  return {
+    id: SEED_FEEDBACK_ID,
+    timestamp: nowIso,
+    signal: 'up',
+    feedback: 'up',
+    title: 'ThumbGate deploy proof seed feedback',
+    context: 'ThumbGate production must keep a searchable thumbgate corpus for authenticated deploy proof.',
+    whatWorked: 'Seed prevention rules and memory so hybrid-parent-child search returns at least one hit.',
+    tags: ['thumbgate', 'deploy-proof', 'seed'],
+  };
+}
+
+function buildSeedRulesBlock() {
+  return [
+    SEED_RULE_HEADING,
+    '',
+    '- ALWAYS keep a minimal ThumbGate-searchable corpus on the production feedback volume.',
+    '- NEVER claim production retrieval is healthy when /v1/search?q=thumbgate returns zero results.',
+    '- Deploy proof depends on hybrid-parent-child search + lesson evidence for the query "thumbgate".',
+    '',
+  ].join('\n');
+}
+
+/**
+ * @param {object} options
+ * @param {string} options.feedbackDir
+ * @param {string} [options.nowIso]
+ * @returns {{feedbackDir:string,wrote:{memory:boolean,feedback:boolean,rules:boolean},paths:object}}
+ */
+function ensureProductionSearchCorpus(options = {}) {
+  const feedbackDir = path.resolve(String(options.feedbackDir || '').trim());
+  if (!feedbackDir) {
+    throw new Error('feedbackDir is required');
+  }
+
+  const nowIso = options.nowIso || new Date().toISOString();
+  ensureDir(feedbackDir);
+
+  const paths = {
+    memoryLog: path.join(feedbackDir, 'memory-log.jsonl'),
+    feedbackLog: path.join(feedbackDir, 'feedback-log.jsonl'),
+    preventionRules: path.join(feedbackDir, 'prevention-rules.md'),
+  };
+
+  const wrote = { memory: false, feedback: false, rules: false };
+
+  if (!fileContains(paths.memoryLog, SEED_ID)) {
+    appendLine(paths.memoryLog, JSON.stringify(buildSeedMemory(nowIso)));
+    wrote.memory = true;
+  }
+
+  if (!fileContains(paths.feedbackLog, SEED_FEEDBACK_ID)) {
+    appendLine(paths.feedbackLog, JSON.stringify(buildSeedFeedback(nowIso)));
+    wrote.feedback = true;
+  }
+
+  if (!fileContains(paths.preventionRules, SEED_RULE_HEADING)) {
+    const existing = readText(paths.preventionRules);
+    const next = existing
+      ? `${existing.replace(/\s*$/, '')}\n\n${buildSeedRulesBlock()}`
+      : `# Prevention Rules\n\n${buildSeedRulesBlock()}`;
+    ensureDir(path.dirname(paths.preventionRules));
+    fs.writeFileSync(paths.preventionRules, next, 'utf8');
+    wrote.rules = true;
+  }
+
+  return { feedbackDir, wrote, paths, seedId: SEED_ID };
+}
+
+module.exports = {
+  SEED_ID,
+  SEED_FEEDBACK_ID,
+  SEED_RULE_HEADING,
+  shouldEnsureProductionSearchCorpus,
+  ensureProductionSearchCorpus,
+};
+
+if (require.main === module) {
+  const feedbackDir = process.argv[2] || process.env.THUMBGATE_FEEDBACK_DIR;
+  if (!feedbackDir) {
+    console.error('Usage: node scripts/ensure-production-search-corpus.js <feedbackDir>');
+    process.exit(2);
+  }
+  const result = ensureProductionSearchCorpus({ feedbackDir });
+  console.log(JSON.stringify(result, null, 2));
+}