@tangle-network/agent-eval 0.126.6 → 0.127.0

package/dist/chunk-ZPUEEM5U.js

@@ -0,0 +1,179 @@
+// src/campaign/run-record.ts
+function campaignCellToRunRecord(cell, options) {
+  const quality = projectCampaignCellQuality(cell);
+  const execution = campaignCellExecutionEvidence(cell);
+  const judgeErrorCount = Math.max(
+    quality.raw.judge_error_count ?? 0,
+    execution.judgeErrorCount ?? 0
+  );
+  const cellCostCaptured = Number.isFinite(cell.costUsd) && cell.costUsd >= 0;
+  const costUsd = cellCostCaptured ? cell.costUsd : options.defaultCostUsd ?? null;
+  const costProvenance = costUsd === null ? { kind: "uncaptured", usd: null } : cellCostCaptured && !cell.costEstimated ? { kind: "observed", usd: costUsd } : { kind: "estimated", usd: costUsd };
+  const raw = {
+    ...finiteMetrics(options.raw),
+    ...quality.raw,
+    rep: cell.rep,
+    duration_ms: cell.durationMs,
+    ...costUsd === null ? {} : { cost_usd: costUsd },
+    cost_estimated: cell.costEstimated ? 1 : 0,
+    tokens_input: cell.tokenUsage.input,
+    tokens_output: cell.tokenUsage.output,
+    latency_ms: cell.durationMs,
+    ...execution.executionErrorCount === void 0 ? {} : { execution_error_count: execution.executionErrorCount },
+    ...judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {},
+    ...execution.unclassifiedErrorCount === void 0 ? {} : { unclassified_error_count: execution.unclassifiedErrorCount }
+  };
+  if (typeof cell.generation === "number") raw.generation = cell.generation;
+  if (cell.tokenUsage.reasoning !== void 0) {
+    raw.tokens_reasoning = cell.tokenUsage.reasoning;
+  }
+  if (cell.tokenUsage.cached !== void 0) raw.tokens_cached = cell.tokenUsage.cached;
+  if (cell.tokenUsage.cacheWrite !== void 0) {
+    raw.tokens_cache_write = cell.tokenUsage.cacheWrite;
+  }
+  if (costUsd !== null && costUsd > 0) {
+    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd;
+  }
+  if (costUsd !== null && quality.score !== void 0 && quality.score > 0.01) {
+    raw.cost_per_quality = costUsd / quality.score;
+  }
+  const outcome = {
+    raw,
+    ...quality.judgeScores ? { judgeScores: quality.judgeScores } : {}
+  };
+  if (quality.score !== void 0) {
+    if (options.splitTag === "holdout") outcome.holdoutScore = quality.score;
+    else outcome.searchScore = quality.score;
+  }
+  return {
+    runId: options.runId,
+    experimentId: options.experimentId,
+    candidateId: options.candidateId,
+    seed: options.seed ?? cell.seed,
+    model: options.model,
+    promptHash: options.promptHash,
+    configHash: options.configHash,
+    commitSha: options.commitSha,
+    wallMs: cell.durationMs,
+    costUsd,
+    costProvenance,
+    tokenUsage: { ...cell.tokenUsage },
+    terminalOutcome: execution.terminalOutcome,
+    ...execution.terminalFailureReason ? { terminalFailureReason: execution.terminalFailureReason } : {},
+    outcome,
+    splitTag: options.splitTag,
+    scenarioId: options.scenarioId ?? cell.scenarioId,
+    ...options.agentProfile ? { agentProfile: options.agentProfile } : {}
+  };
+}
+function campaignCellExecutionEvidence(cell) {
+  if (cell.errorStage === "dispatch") {
+    return {
+      terminalOutcome: "failed",
+      executionErrorCount: 1,
+      ...cell.error ? { terminalFailureReason: cell.error } : {}
+    };
+  }
+  if (cell.errorStage === "judge") {
+    return {
+      terminalOutcome: "succeeded",
+      executionErrorCount: 0,
+      judgeErrorCount: 1
+    };
+  }
+  if (!cell.error) {
+    return { terminalOutcome: "succeeded", executionErrorCount: 0 };
+  }
+  return {
+    terminalOutcome: "unknown",
+    unclassifiedErrorCount: 1
+  };
+}
+function projectCampaignCellQuality(cell) {
+  if (cell.errorStage === "dispatch") {
+    return { successfulJudgeScores: {}, failedJudges: [], raw: {} };
+  }
+  const perJudge = {};
+  const successfulJudgeScores = {};
+  const dimensionValues = /* @__PURE__ */ new Map();
+  const composites = [];
+  const notes = [];
+  const failedJudges = new Set(
+    cell.errorStage === "judge" ? [cell.errorJudge ?? "unknown-judge"] : []
+  );
+  const raw = {};
+  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {
+    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite);
+    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {
+      failedJudges.add(judgeName);
+      continue;
+    }
+    composites.push(score.composite);
+    successfulJudgeScores[judgeName] = score;
+    const dimensions = { ...score.dimensions };
+    perJudge[judgeName] = dimensions;
+    for (const [dimension, value] of Object.entries(dimensions)) {
+      raw[`${judgeName}.${dimension}`] = value;
+      const values = dimensionValues.get(dimension) ?? [];
+      values.push(value);
+      dimensionValues.set(dimension, values);
+    }
+    if (score.notes) notes.push(`${judgeName}: ${score.notes}`);
+    for (const failedJudge of score.failedJudges ?? []) {
+      failedJudges.add(`${judgeName}/${failedJudge}`);
+    }
+  }
+  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size;
+  const sortedFailedJudges = [...failedJudges].sort();
+  if (composites.length === 0) {
+    return {
+      successfulJudgeScores,
+      failedJudges: sortedFailedJudges,
+      raw
+    };
+  }
+  const composite = mean(composites);
+  const perDimMean = Object.fromEntries(
+    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)])
+  );
+  const complete = cell.error === void 0 && cell.errorStage === void 0 && failedJudges.size === 0;
+  if (complete) raw.composite = composite;
+  return {
+    ...complete ? { score: composite } : {},
+    raw,
+    successfulJudgeScores,
+    failedJudges: sortedFailedJudges,
+    judgeScores: {
+      perJudge,
+      perDimMean,
+      composite,
+      ...sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {},
+      ...notes.length > 0 ? { notes: notes.join(" | ") } : {}
+    }
+  };
+}
+function campaignCellTaskScore(cell) {
+  return projectCampaignCellQuality(cell).score;
+}
+function campaignCellJudgeDimensions(cell) {
+  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {};
+}
+function finiteMetrics(metrics) {
+  const finite = {};
+  for (const [key, value] of Object.entries(metrics ?? {})) {
+    if (Number.isFinite(value)) finite[key] = value;
+  }
+  return finite;
+}
+function mean(values) {
+  return values.reduce((sum, value) => sum + value, 0) / values.length;
+}
+
+export {
+  campaignCellToRunRecord,
+  campaignCellExecutionEvidence,
+  projectCampaignCellQuality,
+  campaignCellTaskScore,
+  campaignCellJudgeDimensions
+};
+//# sourceMappingURL=chunk-ZPUEEM5U.js.map

package/dist/chunk-ZPUEEM5U.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/campaign/run-record.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n  runId: string\n  experimentId: string\n  candidateId: string\n  model: string\n  promptHash: string\n  configHash: string\n  commitSha: string\n  splitTag: RunSplitTag\n  seed?: number\n  scenarioId?: string\n  defaultCostUsd?: number\n  agentProfile?: AgentProfileCell\n  raw?: Record<string, number>\n}\n\nexport interface CampaignCellQualityProjection {\n  score?: number\n  judgeScores?: JudgeScoresRecord\n  successfulJudgeScores: Record<string, JudgeScore>\n  failedJudges: string[]\n  raw: Record<string, number>\n}\n\nexport interface CampaignCellExecutionEvidence {\n  terminalOutcome: RunTerminalOutcome\n  executionErrorCount?: number\n  judgeErrorCount?: number\n  unclassifiedErrorCount?: number\n  terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n  options: CampaignCellRunRecordOptions,\n): RunRecord {\n  const quality = projectCampaignCellQuality(cell)\n  const execution = campaignCellExecutionEvidence(cell)\n  const judgeErrorCount = Math.max(\n    quality.raw.judge_error_count ?? 0,\n    execution.judgeErrorCount ?? 0,\n  )\n  const cellCostCaptured = Number.isFinite(cell.costUsd) && cell.costUsd >= 0\n  const costUsd = cellCostCaptured ? cell.costUsd : (options.defaultCostUsd ?? null)\n  const costProvenance =\n    costUsd === null\n      ? ({ kind: 'uncaptured', usd: null } as const)\n      : cellCostCaptured && !cell.costEstimated\n        ? ({ kind: 'observed', usd: costUsd } as const)\n        : ({ kind: 'estimated', usd: costUsd } as const)\n  const raw: Record<string, number> = {\n    ...finiteMetrics(options.raw),\n    ...quality.raw,\n    rep: cell.rep,\n    duration_ms: cell.durationMs,\n    ...(costUsd === null ? {} : { cost_usd: costUsd }),\n    cost_estimated: cell.costEstimated ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    latency_ms: cell.durationMs,\n    ...(execution.executionErrorCount === undefined\n      ? {}\n      : { execution_error_count: execution.executionErrorCount }),\n    ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n    ...(execution.unclassifiedErrorCount === undefined\n      ? {}\n      : { unclassified_error_count: execution.unclassifiedErrorCount }),\n  }\n  if (typeof cell.generation === 'number') raw.generation = cell.generation\n  if (cell.tokenUsage.reasoning !== undefined) {\n    raw.tokens_reasoning = cell.tokenUsage.reasoning\n  }\n  if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n  if (cell.tokenUsage.cacheWrite !== undefined) {\n    raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n  }\n  if (costUsd !== null && costUsd > 0) {\n    raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n  }\n  if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n    raw.cost_per_quality = costUsd / quality.score\n  }\n\n  const outcome: RunOutcome = {\n    raw,\n    ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n  }\n  if (quality.score !== undefined) {\n    if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n    else outcome.searchScore = quality.score\n  }\n\n  return {\n    runId: options.runId,\n    experimentId: options.experimentId,\n    candidateId: options.candidateId,\n    seed: options.seed ?? cell.seed,\n    model: options.model,\n    promptHash: options.promptHash,\n    configHash: options.configHash,\n    commitSha: options.commitSha,\n    wallMs: cell.durationMs,\n    costUsd,\n    costProvenance,\n    tokenUsage: { ...cell.tokenUsage },\n    terminalOutcome: execution.terminalOutcome,\n    ...(execution.terminalFailureReason\n      ? { terminalFailureReason: execution.terminalFailureReason }\n      : {}),\n    outcome,\n    splitTag: options.splitTag,\n    scenarioId: options.scenarioId ?? cell.scenarioId,\n    ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n  }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n  if (cell.errorStage === 'dispatch') {\n    return {\n      terminalOutcome: 'failed',\n      executionErrorCount: 1,\n      ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n    }\n  }\n  if (cell.errorStage === 'judge') {\n    return {\n      terminalOutcome: 'succeeded',\n      executionErrorCount: 0,\n      judgeErrorCount: 1,\n    }\n  }\n  if (!cell.error) {\n    return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n  }\n  return {\n    terminalOutcome: 'unknown',\n    unclassifiedErrorCount: 1,\n  }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n  if (cell.errorStage === 'dispatch') {\n    return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n  }\n\n  const perJudge: Record<string, Record<string, number>> = {}\n  const successfulJudgeScores: Record<string, JudgeScore> = {}\n  const dimensionValues = new Map<string, number[]>()\n  const composites: number[] = []\n  const notes: string[] = []\n  const failedJudges = new Set<string>(\n    cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n  )\n  const raw: Record<string, number> = {}\n\n  for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n    const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n    if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n      failedJudges.add(judgeName)\n      continue\n    }\n\n    composites.push(score.composite)\n    successfulJudgeScores[judgeName] = score\n    const dimensions = { ...score.dimensions }\n    perJudge[judgeName] = dimensions\n    for (const [dimension, value] of Object.entries(dimensions)) {\n      raw[`${judgeName}.${dimension}`] = value\n      const values = dimensionValues.get(dimension) ?? []\n      values.push(value)\n      dimensionValues.set(dimension, values)\n    }\n    if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n    for (const failedJudge of score.failedJudges ?? []) {\n      failedJudges.add(`${judgeName}/${failedJudge}`)\n    }\n  }\n\n  if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n  const sortedFailedJudges = [...failedJudges].sort()\n  if (composites.length === 0) {\n    return {\n      successfulJudgeScores,\n      failedJudges: sortedFailedJudges,\n      raw,\n    }\n  }\n\n  const composite = mean(composites)\n  const perDimMean = Object.fromEntries(\n    [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n  )\n  const complete =\n    cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n  if (complete) raw.composite = composite\n\n  return {\n    ...(complete ? { score: composite } : {}),\n    raw,\n    successfulJudgeScores,\n    failedJudges: sortedFailedJudges,\n    judgeScores: {\n      perJudge,\n      perDimMean,\n      composite,\n      ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n      ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n    },\n  }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n  return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n  cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n  return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n  const finite: Record<string, number> = {}\n  for (const [key, value] of Object.entries(metrics ?? {})) {\n    if (Number.isFinite(value)) finite[key] = value\n  }\n  return finite\n}\n\nfunction mean(values: number[]): number {\n  return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";AAkDO,SAAS,wBACd,MACA,SACW;AACX,QAAM,UAAU,2BAA2B,IAAI;AAC/C,QAAM,YAAY,8BAA8B,IAAI;AACpD,QAAM,kBAAkB,KAAK;AAAA,IAC3B,QAAQ,IAAI,qBAAqB;AAAA,IACjC,UAAU,mBAAmB;AAAA,EAC/B;AACA,QAAM,mBAAmB,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,WAAW;AAC1E,QAAM,UAAU,mBAAmB,KAAK,UAAW,QAAQ,kBAAkB;AAC7E,QAAM,iBACJ,YAAY,OACP,EAAE,MAAM,cAAc,KAAK,KAAK,IACjC,oBAAoB,CAAC,KAAK,gBACvB,EAAE,MAAM,YAAY,KAAK,QAAQ,IACjC,EAAE,MAAM,aAAa,KAAK,QAAQ;AAC3C,QAAM,MAA8B;AAAA,IAClC,GAAG,cAAc,QAAQ,GAAG;AAAA,IAC5B,GAAG,QAAQ;AAAA,IACX,KAAK,KAAK;AAAA,IACV,aAAa,KAAK;AAAA,IAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;AAAA,IAChD,gBAAgB,KAAK,gBAAgB,IAAI;AAAA,IACzC,cAAc,KAAK,WAAW;AAAA,IAC9B,eAAe,KAAK,WAAW;AAAA,IAC/B,YAAY,KAAK;AAAA,IACjB,GAAI,UAAU,wBAAwB,SAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;AAAA,IAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;AAAA,IACpE,GAAI,UAAU,2BAA2B,SACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;AAAA,EACnE;AACA,MAAI,OAAO,KAAK,eAAe,SAAU,KAAI,aAAa,KAAK;AAC/D,MAAI,KAAK,WAAW,cAAc,QAAW;AAC3C,QAAI,mBAAmB,KAAK,WAAW;AAAA,EACzC;AACA,MAAI,KAAK,WAAW,WAAW,OAAW,KAAI,gBAAgB,KAAK,WAAW;AAC9E,MAAI,KAAK,WAAW,eAAe,QAAW;AAC5C,QAAI,qBAAqB,KAAK,WAAW;AAAA,EAC3C;AACA,MAAI,YAAY,QAAQ,UAAU,GAAG;AACnC,QAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;AAAA,EAC7E;AACA,MAAI,YAAY,QAAQ,QAAQ,UAAU,UAAa,QAAQ,QAAQ,MAAM;AAC3E,QAAI,mBAAmB,UAAU,QAAQ;AAAA,EAC3C;AAEA,QAAM,UAAsB;AAAA,IAC1B;AAAA,IACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;AAAA,EACpE;AACA,MAAI,QAAQ,UAAU,QAAW;AAC/B,QAAI,QAAQ,aAAa,UAAW,SAAQ,eAAe,QAAQ;AAAA,QAC9D,SAAQ,cAAc,QAAQ;AAAA,EACrC;AAEA,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf,cAAc,QAAQ;AAAA,IACtB,aAAa,QAAQ;AAAA,IACrB,MAAM,QAAQ,QAAQ,KAAK;AAAA,IAC3B,OAAO,QAAQ;AAAA,IACf,YAAY,QAAQ;AAAA,IACpB,YAAY,QAAQ;AAAA,IACpB,WAAW,QAAQ;AAAA,IACnB,QAAQ,KAAK;AAAA,IACb;AAAA,IACA;AAAA,IACA,YAAY,EAAE,GAAG,KAAK,WAAW;AAAA,IACjC,iBAAiB,UAAU;AAAA,IAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;AAAA,IACL;AAAA,IACA,UAAU,QAAQ;AAAA,IAClB,YAAY,QAAQ,cAAc,KAAK;AAAA,IACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;AAAA,EACvE;AACF;AAEO,SAAS,8BACd,MAC+B;AAC/B,MAAI,KAAK,eAAe,YAAY;AAClC,WAAO;AAAA,MACL,iBAAiB;AAAA,MACjB,qBAAqB;AAAA,MACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;AAAA,IAC5D;AAAA,EACF;AACA,MAAI,KAAK,eAAe,SAAS;AAC/B,WAAO;AAAA,MACL,iBAAiB;AAAA,MACjB,qBAAqB;AAAA,MACrB,iBAAiB;AAAA,IACnB;AAAA,EACF;AACA,MAAI,CAAC,KAAK,OAAO;AACf,WAAO,EAAE,iBAAiB,aAAa,qBAAqB,EAAE;AAAA,EAChE;AACA,SAAO;AAAA,IACL,iBAAiB;AAAA,IACjB,wBAAwB;AAAA,EAC1B;AACF;AASO,SAAS,2BACd,MAC+B;AAC/B,MAAI,KAAK,eAAe,YAAY;AAClC,WAAO,EAAE,uBAAuB,CAAC,GAAG,cAAc,CAAC,GAAG,KAAK,CAAC,EAAE;AAAA,EAChE;AAEA,QAAM,WAAmD,CAAC;AAC1D,QAAM,wBAAoD,CAAC;AAC3D,QAAM,kBAAkB,oBAAI,IAAsB;AAClD,QAAM,aAAuB,CAAC;AAC9B,QAAM,QAAkB,CAAC;AACzB,QAAM,eAAe,IAAI;AAAA,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC;AAAA,EACxE;AACA,QAAM,MAA8B,CAAC;AAErC,aAAW,CAAC,WAAW,KAAK,KAAK,OAAO,QAAQ,KAAK,WAAW,GAAG;AACjE,UAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,EAAE,MAAM,OAAO,QAAQ;AAC9E,QAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;AAC1E,mBAAa,IAAI,SAAS;AAC1B;AAAA,IACF;AAEA,eAAW,KAAK,MAAM,SAAS;AAC/B,0BAAsB,SAAS,IAAI;AACnC,UAAM,aAAa,EAAE,GAAG,MAAM,WAAW;AACzC,aAAS,SAAS,IAAI;AACtB,eAAW,CAAC,WAAW,KAAK,KAAK,OAAO,QAAQ,UAAU,GAAG;AAC3D,UAAI,GAAG,SAAS,IAAI,SAAS,EAAE,IAAI;AACnC,YAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;AAClD,aAAO,KAAK,KAAK;AACjB,sBAAgB,IAAI,WAAW,MAAM;AAAA,IACvC;AACA,QAAI,MAAM,MAAO,OAAM,KAAK,GAAG,SAAS,KAAK,MAAM,KAAK,EAAE;AAC1D,eAAW,eAAe,MAAM,gBAAgB,CAAC,GAAG;AAClD,mBAAa,IAAI,GAAG,SAAS,IAAI,WAAW,EAAE;AAAA,IAChD;AAAA,EACF;AAEA,MAAI,aAAa,OAAO,EAAG,KAAI,oBAAoB,aAAa;AAChE,QAAM,qBAAqB,CAAC,GAAG,YAAY,EAAE,KAAK;AAClD,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO;AAAA,MACL;AAAA,MACA,cAAc;AAAA,MACd;AAAA,IACF;AAAA,EACF;AAEA,QAAM,YAAY,KAAK,UAAU;AACjC,QAAM,aAAa,OAAO;AAAA,IACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,WAAW,MAAM,MAAM,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC;AAAA,EACvF;AACA,QAAM,WACJ,KAAK,UAAU,UAAa,KAAK,eAAe,UAAa,aAAa,SAAS;AACrF,MAAI,SAAU,KAAI,YAAY;AAE9B,SAAO;AAAA,IACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;AAAA,IACvC;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd,aAAa;AAAA,MACX;AAAA,MACA;AAAA,MACA;AAAA,MACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;AAAA,MAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;AAAA,IACzD;AAAA,EACF;AACF;AAGO,SAAS,sBACd,MACoB;AACpB,SAAO,2BAA2B,IAAI,EAAE;AAC1C;AAGO,SAAS,4BACd,MACwC;AACxC,SAAO,2BAA2B,IAAI,EAAE,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;AAC1F,QAAM,SAAiC,CAAC;AACxC,aAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,WAAW,CAAC,CAAC,GAAG;AACxD,QAAI,OAAO,SAAS,KAAK,EAAG,QAAO,GAAG,IAAI;AAAA,EAC5C;AACA,SAAO;AACT;AAEA,SAAS,KAAK,QAA0B;AACtC,SAAO,OAAO,OAAO,CAAC,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE;","names":[]}

package/dist/{chunk-KE2VWPZX.js → chunk-ZUUWPZCV.js}

@@ -1,6 +1,6 @@
 import {
   computeFindingId
-} from "./chunk-LUNF2SEL.js";
+} from "./chunk-DJKY2TSY.js";
 import {
   Mutex,
   aggregateRunScore,
@@ -11,7 +11,7 @@ import {
   costReceiptFromLlm,
   costReceiptFromLlmError,
   maximumChargeForLlmRequest
-} from "./chunk-J5SQWP6Y.js";
+} from "./chunk-PBE2LOSS.js";
 import {
   CostLedger
 } from "./chunk-WS3NZZQQ.js";
@@ -749,4 +749,4 @@ export {
   runSemanticConceptJudge,
   createSemanticConceptJudge
 };
-//# sourceMappingURL=chunk-KE2VWPZX.js.map
+//# sourceMappingURL=chunk-ZUUWPZCV.js.map

package/dist/cli.js

@@ -5,12 +5,12 @@ import {
   runRpcBatch,
   runRpcOnce,
   startServerAsync
-} from "./chunk-FO7HEH76.js";
+} from "./chunk-YJBNWCAA.js";
 import {
   runRolloutReleaseCli
-} from "./chunk-IPYXE555.js";
-import "./chunk-MAX3TN3C.js";
-import "./chunk-J5SQWP6Y.js";
+} from "./chunk-EJGRPCO3.js";
+import "./chunk-UWZZKKU7.js";
+import "./chunk-PBE2LOSS.js";
 import "./chunk-WS3NZZQQ.js";
 import "./chunk-VI2UW6B6.js";
 import "./chunk-PC4UYEBM.js";