@tangle-network/agent-eval 0.79.0 → 0.81.0

package/dist/chunk-4DIJWVUT.js

@@ -0,0 +1,131 @@
+import {
+  ValidationError
+} from "./chunk-3BFEG2F6.js";
+
+// src/rl/off-policy.ts
+function inverseProbabilityWeighting(trajectories, opts = {}) {
+  const cap = opts.weightCap ?? Infinity;
+  const clip = opts.rewardClip ?? { low: 0, high: 1 };
+  if (trajectories.length === 0) {
+    return zeroEstimate();
+  }
+  const weights = [];
+  const weightedRewards = [];
+  let maxW = 0;
+  for (const t of trajectories) {
+    if (t.behaviorProb <= 0) {
+      throw new ValidationError(
+        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`
+      );
+    }
+    const w = Math.min(cap, t.targetProb / t.behaviorProb);
+    const r = clamp(t.reward, clip.low, clip.high);
+    weights.push(w);
+    weightedRewards.push(w * r);
+    if (w > maxW) maxW = w;
+  }
+  const n = weights.length;
+  const value = weightedRewards.reduce((s, x) => s + x, 0) / n;
+  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
+  const sumW = weights.reduce((s, w) => s + w, 0);
+  const sumW2 = weights.reduce((s, w) => s + w * w, 0);
+  const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+  return {
+    value,
+    standardError: Math.sqrt(variance / n),
+    effectiveSampleSize: effN,
+    n,
+    maxImportanceWeight: maxW
+  };
+}
+function selfNormalizedImportanceWeighting(trajectories, opts = {}) {
+  const cap = opts.weightCap ?? Infinity;
+  const clip = opts.rewardClip ?? { low: 0, high: 1 };
+  if (trajectories.length === 0) return zeroEstimate();
+  const weights = [];
+  const rewards = [];
+  let maxW = 0;
+  for (const t of trajectories) {
+    if (t.behaviorProb <= 0) {
+      throw new ValidationError(
+        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`
+      );
+    }
+    const w = Math.min(cap, t.targetProb / t.behaviorProb);
+    weights.push(w);
+    rewards.push(clamp(t.reward, clip.low, clip.high));
+    if (w > maxW) maxW = w;
+  }
+  const sumW = weights.reduce((s, w) => s + w, 0);
+  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i], 0);
+  const value = sumW === 0 ? 0 : sumWR / sumW;
+  const sumW2 = weights.reduce((s, w) => s + w * w, 0);
+  const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+  const phi = weights.map((w, i) => w * (rewards[i] - value));
+  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW);
+  return {
+    value,
+    standardError: Math.sqrt(variance),
+    effectiveSampleSize: effN,
+    n: trajectories.length,
+    maxImportanceWeight: maxW
+  };
+}
+function doublyRobust(trajectories, opts = {}) {
+  const cap = opts.weightCap ?? Infinity;
+  const clip = opts.rewardClip ?? { low: 0, high: 1 };
+  if (trajectories.length === 0) return zeroEstimate();
+  const contributions = [];
+  let maxW = 0;
+  let sumW = 0;
+  let sumW2 = 0;
+  for (const t of trajectories) {
+    if (t.behaviorProb <= 0) {
+      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`);
+    }
+    const w = Math.min(cap, t.targetProb / t.behaviorProb);
+    const r = clamp(t.reward, clip.low, clip.high);
+    const q = typeof t.qHat === "number" && Number.isFinite(t.qHat) ? clamp(t.qHat, clip.low, clip.high) : null;
+    if (q === null) {
+      contributions.push(w * r);
+    } else {
+      contributions.push(q + w * (r - q));
+    }
+    if (w > maxW) maxW = w;
+    sumW += w;
+    sumW2 += w * w;
+  }
+  const n = contributions.length;
+  const value = contributions.reduce((s, x) => s + x, 0) / n;
+  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1);
+  const effN = sumW === 0 ? 0 : sumW * sumW / sumW2;
+  return {
+    value,
+    standardError: Math.sqrt(variance / n),
+    effectiveSampleSize: effN,
+    n,
+    maxImportanceWeight: maxW
+  };
+}
+function offPolicyEstimateAll(trajectories, opts = {}) {
+  return {
+    ips: inverseProbabilityWeighting(trajectories, opts),
+    snips: selfNormalizedImportanceWeighting(trajectories, opts),
+    dr: doublyRobust(trajectories, opts)
+  };
+}
+function zeroEstimate() {
+  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 };
+}
+function clamp(x, lo, hi) {
+  if (!Number.isFinite(x)) return lo;
+  return Math.max(lo, Math.min(hi, x));
+}
+
+export {
+  inverseProbabilityWeighting,
+  selfNormalizedImportanceWeighting,
+  doublyRobust,
+  offPolicyEstimateAll
+};
+//# sourceMappingURL=chunk-4DIJWVUT.js.map

package/dist/chunk-4DIJWVUT.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rl/off-policy.ts"],"sourcesContent":["/**\n * Off-policy evaluation primitives.\n *\n * Standard inverse-probability-weighted (IPS), self-normalized\n * importance-weighted (SNIPS), and doubly-robust (DR) estimators for the\n * value of a *target* policy given trajectories collected under a\n * *behavior* policy. This is the canonical RL eval task: \"we have last\n * week's runs, we changed the policy — how would the new one do without\n * re-running?\"\n *\n * The math here is textbook (Dudík, Langford, Li 2011 for DR; Swaminathan\n * & Joachims 2015 for SNIPS) but the *application* to LLM-agent\n * evaluation needs care:\n *\n *   - The \"policy\" is the (prompt, tool config, model snapshot) triple.\n *     Two policies have the same probability over an action *iff* their\n *     LLM call would emit the same token with the same probability —\n *     which is generally unknowable without the model log-probs.\n *   - For LLM agents, propensity scores must be supplied by the caller\n *     (logged in the trace, recovered from token log-probs, or estimated\n *     via a learned propensity model). We do NOT estimate propensity here.\n *   - Doubly-robust requires a Q-function (model-based reward predictor).\n *     We accept any callable; consumers pass either a tabular average,\n *     a regression fit, or a learned reward model.\n *\n * Bias / variance tradeoffs:\n *   - IPS: unbiased; high variance for small overlap, infinite variance\n *     when target has support outside behavior.\n *   - SNIPS: lower variance, slight bias; usually preferred in practice.\n *   - DR: doubly-robust — unbiased if either propensity OR Q-function is\n *     correct. Lowest practical variance when Q is decent. Use this.\n *\n * Caveat the panel will land: on the LLM-agent setting, propensity scores\n * recovered from token log-probs are noisy, the action space is enormous,\n * and overlap is often poor. These estimators are useful but not magic;\n * complement with `replayCampaign` (exact replay where the request hashes\n * match) for high-confidence answers and OPE for the gap.\n */\n\nimport { ValidationError } from '../errors'\n\nexport interface OffPolicyTrajectory {\n  /** Stable id, for traceability through the dataset. */\n  runId: string\n  /** Reward observed under the behavior policy (the realized outcome). */\n  reward: number\n  /**\n   * Behavior-policy probability of the action that was taken. For LLM\n   * agents this is typically `exp(sum(token_log_probs))` over the chosen\n   * trajectory. Must be in (0, 1].\n   */\n  behaviorProb: number\n  /**\n   * Target-policy probability of the same action. For replay-style\n   * counterfactual evaluation this is what the *new* policy would have\n   * assigned to the *old* trajectory. Must be in [0, 1].\n   */\n  targetProb: number\n  /**\n   * Optional model-based reward prediction at the same context. Used by\n   * `doublyRobust`. Set to `null` for IPS-only evaluation.\n   */\n  qHat?: number | null\n}\n\nexport interface OffPolicyEstimate {\n  /** Estimated value of the target policy. */\n  value: number\n  /** Standard error of the estimate. */\n  standardError: number\n  /** Effective sample size (Kong 1992). Lower = more reliance on a few high-weight samples. */\n  effectiveSampleSize: number\n  /** Number of trajectories used. */\n  n: number\n  /**\n   * Diagnostic: maximum importance weight observed. Large values (>>10x\n   * mean) are a red flag — variance is dominated by a few outliers.\n   */\n  maxImportanceWeight: number\n}\n\nexport interface OffPolicyOptions {\n  /**\n   * Cap importance weights at this value (Ionides 2008 truncated IS) to\n   * trade unbiasedness for variance reduction. Default `Infinity` (no cap).\n   * Set e.g. `10` for stable estimates when the policies are close.\n   */\n  weightCap?: number\n  /** Reward clipping range. Default `[0, 1]`. */\n  rewardClip?: { low: number; high: number }\n}\n\n/**\n * Inverse Probability Weighting (Horvitz-Thompson). Unbiased estimator\n * of E[reward under target policy]. Variance scales with the spread of\n * target/behavior ratios.\n */\nexport function inverseProbabilityWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n\n  if (trajectories.length === 0) {\n    return zeroEstimate()\n  }\n\n  const weights: number[] = []\n  const weightedRewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `inverseProbabilityWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    weights.push(w)\n    weightedRewards.push(w * r)\n    if (w > maxW) maxW = w\n  }\n  const n = weights.length\n  const value = weightedRewards.reduce((s, x) => s + x, 0) / n\n  const variance = weightedRewards.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Self-Normalized Importance Sampling. Lower variance than vanilla IPS at\n * the cost of small bias (vanishing as N grows). The right default for\n * LLM-agent evaluation where overlap is often poor.\n */\nexport function selfNormalizedImportanceWeighting(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const weights: number[] = []\n  const rewards: number[] = []\n  let maxW = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(\n        `selfNormalizedImportanceWeighting: behaviorProb must be > 0 (runId=${t.runId})`,\n      )\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    weights.push(w)\n    rewards.push(clamp(t.reward, clip.low, clip.high))\n    if (w > maxW) maxW = w\n  }\n  const sumW = weights.reduce((s, w) => s + w, 0)\n  const sumWR = weights.reduce((s, w, i) => s + w * rewards[i]!, 0)\n  const value = sumW === 0 ? 0 : sumWR / sumW\n  const sumW2 = weights.reduce((s, w) => s + w * w, 0)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  // Influence-function-based SE for SNIPS (Owen 2013, Ch. 9).\n  const phi = weights.map((w, i) => w * (rewards[i]! - value))\n  const variance = phi.reduce((s, x) => s + x * x, 0) / Math.max(1, sumW * sumW)\n  return {\n    value,\n    standardError: Math.sqrt(variance),\n    effectiveSampleSize: effN,\n    n: trajectories.length,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Doubly-robust off-policy estimator (Dudík, Langford, Li 2011).\n *\n *     V_DR = (1/N) * sum_i [ q_hat_i + (target_prob_i / behavior_prob_i) * (r_i - q_hat_i) ]\n *\n * Unbiased if EITHER:\n *   - the importance ratios are correct (IPS-style validity), OR\n *   - the Q-hat function is correct (model-based validity).\n *\n * In practice both are imperfect, but the residual bias is the *product*\n * of both errors — much smaller than either alone. This is why DR is the\n * default in production OPE pipelines.\n *\n * Requires `qHat` on every trajectory. If any are `null`, the estimator\n * falls back to SNIPS for those entries (loud-fallback behavior; the\n * report's `n` reflects the full set but `effectiveSampleSize` accounts\n * for the lost variance reduction).\n */\nexport function doublyRobust(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): OffPolicyEstimate {\n  const cap = opts.weightCap ?? Infinity\n  const clip = opts.rewardClip ?? { low: 0, high: 1 }\n  if (trajectories.length === 0) return zeroEstimate()\n\n  const contributions: number[] = []\n  let maxW = 0\n  let sumW = 0\n  let sumW2 = 0\n  for (const t of trajectories) {\n    if (t.behaviorProb <= 0) {\n      throw new ValidationError(`doublyRobust: behaviorProb must be > 0 (runId=${t.runId})`)\n    }\n    const w = Math.min(cap, t.targetProb / t.behaviorProb)\n    const r = clamp(t.reward, clip.low, clip.high)\n    const q =\n      typeof t.qHat === 'number' && Number.isFinite(t.qHat)\n        ? clamp(t.qHat, clip.low, clip.high)\n        : null\n    if (q === null) {\n      contributions.push(w * r) // fallback: IPS for this entry\n    } else {\n      contributions.push(q + w * (r - q))\n    }\n    if (w > maxW) maxW = w\n    sumW += w\n    sumW2 += w * w\n  }\n  const n = contributions.length\n  const value = contributions.reduce((s, x) => s + x, 0) / n\n  const variance = contributions.reduce((s, x) => s + (x - value) ** 2, 0) / Math.max(1, n - 1)\n  const effN = sumW === 0 ? 0 : (sumW * sumW) / sumW2\n  return {\n    value,\n    standardError: Math.sqrt(variance / n),\n    effectiveSampleSize: effN,\n    n,\n    maxImportanceWeight: maxW,\n  }\n}\n\n/**\n * Convenience: run all three estimators and return them side-by-side.\n * The recommended diagnostic — agreement across estimators is a much\n * stronger signal than any single one.\n */\nexport function offPolicyEstimateAll(\n  trajectories: OffPolicyTrajectory[],\n  opts: OffPolicyOptions = {},\n): { ips: OffPolicyEstimate; snips: OffPolicyEstimate; dr: OffPolicyEstimate } {\n  return {\n    ips: inverseProbabilityWeighting(trajectories, opts),\n    snips: selfNormalizedImportanceWeighting(trajectories, opts),\n    dr: doublyRobust(trajectories, opts),\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction zeroEstimate(): OffPolicyEstimate {\n  return { value: 0, standardError: 0, effectiveSampleSize: 0, n: 0, maxImportanceWeight: 0 }\n}\n\nfunction clamp(x: number, lo: number, hi: number): number {\n  if (!Number.isFinite(x)) return lo\n  return Math.max(lo, Math.min(hi, x))\n}\n"],"mappings":";;;;;AAiGO,SAAS,4BACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAElD,MAAI,aAAa,WAAW,GAAG;AAC7B,WAAO,aAAa;AAAA,EACtB;AAEA,QAAM,UAAoB,CAAC;AAC3B,QAAM,kBAA4B,CAAC;AACnC,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,gEAAgE,EAAE,KAAK;AAAA,MACzE;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,YAAQ,KAAK,CAAC;AACd,oBAAgB,KAAK,IAAI,CAAC;AAC1B,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,IAAI,QAAQ;AAClB,QAAM,QAAQ,gBAAgB,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC3D,QAAM,WAAW,gBAAgB,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC9F,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,kCACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,UAAoB,CAAC;AAC3B,QAAM,UAAoB,CAAC;AAC3B,MAAI,OAAO;AACX,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI;AAAA,QACR,sEAAsE,EAAE,KAAK;AAAA,MAC/E;AAAA,IACF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,YAAQ,KAAK,CAAC;AACd,YAAQ,KAAK,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI,CAAC;AACjD,QAAI,IAAI,KAAM,QAAO;AAAA,EACvB;AACA,QAAM,OAAO,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC;AAC9C,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,GAAG,MAAM,IAAI,IAAI,QAAQ,CAAC,GAAI,CAAC;AAChE,QAAM,QAAQ,SAAS,IAAI,IAAI,QAAQ;AACvC,QAAM,QAAQ,QAAQ,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC;AACnD,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAE9C,QAAM,MAAM,QAAQ,IAAI,CAAC,GAAG,MAAM,KAAK,QAAQ,CAAC,IAAK,MAAM;AAC3D,QAAM,WAAW,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,IAAI,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,OAAO,IAAI;AAC7E,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,QAAQ;AAAA,IACjC,qBAAqB;AAAA,IACrB,GAAG,aAAa;AAAA,IAChB,qBAAqB;AAAA,EACvB;AACF;AAoBO,SAAS,aACd,cACA,OAAyB,CAAC,GACP;AACnB,QAAM,MAAM,KAAK,aAAa;AAC9B,QAAM,OAAO,KAAK,cAAc,EAAE,KAAK,GAAG,MAAM,EAAE;AAClD,MAAI,aAAa,WAAW,EAAG,QAAO,aAAa;AAEnD,QAAM,gBAA0B,CAAC;AACjC,MAAI,OAAO;AACX,MAAI,OAAO;AACX,MAAI,QAAQ;AACZ,aAAW,KAAK,cAAc;AAC5B,QAAI,EAAE,gBAAgB,GAAG;AACvB,YAAM,IAAI,gBAAgB,iDAAiD,EAAE,KAAK,GAAG;AAAA,IACvF;AACA,UAAM,IAAI,KAAK,IAAI,KAAK,EAAE,aAAa,EAAE,YAAY;AACrD,UAAM,IAAI,MAAM,EAAE,QAAQ,KAAK,KAAK,KAAK,IAAI;AAC7C,UAAM,IACJ,OAAO,EAAE,SAAS,YAAY,OAAO,SAAS,EAAE,IAAI,IAChD,MAAM,EAAE,MAAM,KAAK,KAAK,KAAK,IAAI,IACjC;AACN,QAAI,MAAM,MAAM;AACd,oBAAc,KAAK,IAAI,CAAC;AAAA,IAC1B,OAAO;AACL,oBAAc,KAAK,IAAI,KAAK,IAAI,EAAE;AAAA,IACpC;AACA,QAAI,IAAI,KAAM,QAAO;AACrB,YAAQ;AACR,aAAS,IAAI;AAAA,EACf;AACA,QAAM,IAAI,cAAc;AACxB,QAAM,QAAQ,cAAc,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACzD,QAAM,WAAW,cAAc,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC,IAAI,KAAK,IAAI,GAAG,IAAI,CAAC;AAC5F,QAAM,OAAO,SAAS,IAAI,IAAK,OAAO,OAAQ;AAC9C,SAAO;AAAA,IACL;AAAA,IACA,eAAe,KAAK,KAAK,WAAW,CAAC;AAAA,IACrC,qBAAqB;AAAA,IACrB;AAAA,IACA,qBAAqB;AAAA,EACvB;AACF;AAOO,SAAS,qBACd,cACA,OAAyB,CAAC,GACmD;AAC7E,SAAO;AAAA,IACL,KAAK,4BAA4B,cAAc,IAAI;AAAA,IACnD,OAAO,kCAAkC,cAAc,IAAI;AAAA,IAC3D,IAAI,aAAa,cAAc,IAAI;AAAA,EACrC;AACF;AAIA,SAAS,eAAkC;AACzC,SAAO,EAAE,OAAO,GAAG,eAAe,GAAG,qBAAqB,GAAG,GAAG,GAAG,qBAAqB,EAAE;AAC5F;AAEA,SAAS,MAAM,GAAW,IAAY,IAAoB;AACxD,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,KAAK,IAAI,IAAI,KAAK,IAAI,IAAI,CAAC,CAAC;AACrC;","names":[]}

package/dist/{chunk-RPLZ4OIB.js → chunk-BABOZOSN.js}

@@ -1,10 +1,10 @@
 import {
   runCampaign,
   summarizeBackendIntegrity
-} from "./chunk-ZPSKPT3V.js";
+} from "./chunk-ZZ2HOPME.js";
 import {
   callLlm
-} from "./chunk-IHDHUN2X.js";
+} from "./chunk-CVVHBFGN.js";
 import {
   DEFAULT_REDACTION_RULES
 } from "./chunk-GGE4NNQT.js";
@@ -1252,6 +1252,9 @@ function labelTrustRank(trust) {
 import { createHash } from "crypto";
 async function runOptimization(opts) {
   const promoteTopK = opts.promoteTopK ?? 2;
+  if (typeof opts.runDir !== "string" || opts.runDir.trim().length === 0) {
+    throw new Error("runOptimization: runDir is required and must be a non-empty string");
+  }
   const baselineCampaign = await runCampaign({
     ...opts,
     dispatch: (scenario, ctx) => opts.dispatchWithSurface(opts.baselineSurface, scenario, ctx),
@@ -1440,7 +1443,7 @@ async function runImprovementLoop(opts) {
   const dispatchTimeoutMs = opts.dispatchTimeoutMs ?? DEFAULT_DISPATCH_TIMEOUT_MS;
   const optimization = await runOptimization({ ...opts, dispatchTimeoutMs });
   const winnerIsBaseline = optimization.winnerSurfaceHash === surfaceHash(opts.baselineSurface);
-  const { runCampaign: runCampaign2 } = await import("./run-campaign-OVEZF24D.js");
+  const { runCampaign: runCampaign2 } = await import("./run-campaign-4Y5V5CN3.js");
   const baselineOnHoldout = await runCampaign2({
     ...opts,
     dispatchTimeoutMs,
@@ -1841,4 +1844,4 @@ export {
   provenanceSpansPath,
   emitLoopProvenance
 };
-//# sourceMappingURL=chunk-RPLZ4OIB.js.map
+//# sourceMappingURL=chunk-BABOZOSN.js.map