@tangle-network/agent-eval 0.149.0 → 0.150.1

package/dist/reward-hacking-DKI9T52l.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"reward-hacking-DKI9T52l.js","names":["mean","clamp01"],"sources":["../src/campaign/run-record.ts","../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTerminalOutcome,\n} from '../run-record'\nimport { validateRunRecord } 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\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n  constructor(judgeName: string) {\n    super(\n      `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n        `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n        `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n        `({ judgeName, dimension, score }) is a different type with the same name.`,\n    )\n    this.name = 'CampaignJudgeScoreShapeError'\n  }\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 cellCostProvenance = campaignCellCostProvenance(cell)\n  const costProvenance: CostProvenance =\n    cellCostProvenance.kind === 'uncaptured' && options.defaultCostUsd !== undefined\n      ? { kind: 'estimated', usd: options.defaultCostUsd }\n      : cellCostProvenance\n  const costUsd = costProvenance.kind === 'uncaptured' ? null : costProvenance.usd\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    // Retain the observed subtotal even when the caller supplies an estimated total.\n    ...(cellCostProvenance.kind === 'uncaptured' ? { cost_known_subtotal_usd: cell.costUsd } : {}),\n    cost_observed: costProvenance.kind === 'observed' ? 1 : 0,\n    cost_estimated: costProvenance.kind === 'estimated' ? 1 : 0,\n    cost_uncaptured: costProvenance.kind === 'uncaptured' ? 1 : 0,\n    tokens_input: cell.tokenUsage.input,\n    tokens_output: cell.tokenUsage.output,\n    tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\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 (cell.tokenUsage.tokensKnown !== false && 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 validateRunRecord({\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\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n  cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n  if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n  }\n  const provenance = cell.costProvenance\n  if (!provenance || typeof provenance !== 'object') {\n    throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n  }\n  if (provenance.kind === 'uncaptured') {\n    if (provenance.usd !== null) {\n      throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n    }\n    return { kind: 'uncaptured', usd: null }\n  }\n  if (\n    (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n    !Number.isFinite(provenance.usd) ||\n    provenance.usd < 0\n  ) {\n    throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n  }\n  if (provenance.usd !== cell.costUsd) {\n    throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n  }\n  return { kind: provenance.kind, usd: provenance.usd }\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 dimensionsShape = (score as { dimensions?: unknown }).dimensions\n    if (\n      typeof dimensionsShape !== 'object' ||\n      dimensionsShape === null ||\n      Array.isArray(dimensionsShape)\n    ) {\n      throw new CampaignJudgeScoreShapeError(judgeName)\n    }\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","/**\n * Verifiable reward channel.\n *\n * For RL on coding / math / theorem-proving / structured-output tasks, the\n * reward signal is *decidable* — a test passes or fails, a proof checks or\n * doesn't, an output validates against a schema or doesn't. These rewards\n * are dramatically more useful for RL training than LLM-judge scores\n * because they don't drift, can't be Goodhart-gamed by the policy in the\n * same way, and don't require a separate calibration loop.\n *\n * The `MultiLayerVerifier` already produces this signal — it just doesn't\n * surface it in a shape that's clean enough for RL training. This module\n * wraps the verifier output so consumers can:\n *\n *   1. Extract a clean `VerifiableReward` from a `VerificationReport`\n *   2. Distinguish *deterministic* rewards (compile, test, schema) from\n *      *probabilistic* rewards (judge) so they can be weighted differently\n *      in the RL training step\n *   3. Filter `RunRecord[]` to only those with a verifiable reward,\n *      producing the clean training set that DeepSeek-R1-style GRPO and\n *      AlphaProof-style search both depend on\n *\n * Why this matters: every credible 2025-2026 frontier RL result on coding\n * agents leans on verifiable reward (DeepSeek-R1 GRPO on test pass-rate,\n * o-series RL on math/code, AlphaProof on Lean kernel checking). Mixing\n * judge scores into the reward signal poisons the gradient. This module\n * is the seam.\n */\n\nimport type { LayerResult, VerificationReport } from '../multi-layer-verifier'\nimport { isRealnessGated, observedScore, trainingScore } from '../rollout/reward'\nimport type { RunRecord } from '../run-record'\nimport type { VerificationStrategySource } from '../verification-strategy'\n\n/**\n * What produced a reward. This is the verification-strategy family\n * (`src/verification-strategy.ts`) — one open vocabulary shared by rewards\n * and verdict certifications, so an RL consumer and a certification reader\n * mean the same thing by `'proof-kernel'`.\n *\n * Beyond the classic answer-key members (`compile`, `test`, `schema`,\n * `sandbox`, `judge`, `composite`) the family carries four members for\n * tasks with no held-out suite, each with a documented failure mode:\n *\n * - `'proof-kernel'` — kernel-checked formal proof. Failure mode: the\n *   formalization gap — the kernel never certifies that the formal\n *   statement matches the informal claim (`src/verification-strategy.ts`\n *   is the discharge protocol).\n * - `'invariant'` — invariant / metamorphic properties held. Failure\n *   mode: weak invariants pass everything; the set needs seeded-bug\n *   calibration before its pass carries weight.\n * - `'replication'` — independent re-execution from pinned inputs.\n *   Failure mode: re-runs the method, so it never catches an error the\n *   method itself carries.\n * - `'agreement'` — independently-derived results agree. Failure mode:\n *   the shared blind spot — derivers with common training corpora or\n *   priors agree for the same wrong reason. Probabilistic: the check may\n *   be mechanical, but the independence it certifies is not.\n *\n * The deterministic/probabilistic axis stays on `VerifiableReward` —\n * the producer declares it per reward, and `VERIFICATION_STRATEGIES`\n * documents each member's default class.\n */\nexport type VerifiableRewardSource = VerificationStrategySource\n\nexport interface VerifiableReward {\n  /** Scalar in [0, 1]. The RL training signal. */\n  value: number\n  /** What produced the reward — different sources have different determinism. */\n  source: VerifiableRewardSource\n  /**\n   * Determinism class. `'deterministic'` rewards are repeatable byte-for-byte\n   * given the same inputs (compile, test, schema validation, sandbox exit code).\n   * `'probabilistic'` rewards depend on a stochastic component (LLM judge).\n   * Mixing these in the same training batch without separation is a known\n   * footgun in production RLHF pipelines.\n   */\n  determinism: 'deterministic' | 'probabilistic'\n  /**\n   * Confidence in the reward value. For deterministic sources this is 1.0\n   * (the bit either flipped or didn't). For judge sources this is the\n   * judge-reported confidence or — when missing — a calibrated prior.\n   */\n  confidence: number\n  /** The layer / judge id that produced the signal, for provenance. */\n  origin: string\n  /**\n   * Per-source contribution to `value`, keyed by layer/judge id. Single-source\n   * rewards carry one entry (`{ [origin]: value }`); composite rewards carry\n   * every contributing layer's score — the anti-scalar-collapse surface RL\n   * consumers weight per-source instead of trusting one blended number.\n   */\n  components: Record<string, number>\n  /**\n   * The run carries `outcome.realness.gated` — the authenticity gate flagged\n   * its success signal as faked.\n   *\n   * With the gate applied (the default) `value` and every `components` entry\n   * are 0 on such a run; with `applyRealnessGate: false` the observed numbers\n   * come back untouched and this flag is the only marker that they are not to\n   * be trusted. Either way it distinguishes \"measured a genuine failure\" from\n   * \"claimed a success we refuse to believe\", which a bare 0 cannot.\n   */\n  realnessGated?: boolean\n  /**\n   * Whether an authenticity screen COULD run on this reward at all — the same\n   * distinction `RolloutOutcome.realness_screened` draws, for the same reason.\n   *\n   * `false` on every reward from `extractVerifiableReward`, because a\n   * `VerificationReport` carries layer scores and nothing else: there is no\n   * `outcome.realness` to consult, so no gate has run, and `realnessGated`\n   * being absent there means \"unknown\", NOT \"clean\". Absent on the\n   * `RunRecord` path when the record itself carries no realness verdict.\n   *\n   * This matters most exactly where it is easiest to miss: a report whose\n   * deterministic layers all passed yields `determinism: 'deterministic'`,\n   * `confidence: 1` — the highest-credibility reward this module can emit —\n   * and a stubbed integration reporting green is precisely what a gamed run\n   * looks like. Consumers driving training off this shape must screen the run\n   * themselves; the flag is what tells them nobody has.\n   */\n  realnessScreened?: boolean\n}\n\nexport interface VerifiableRewardExtractionOptions {\n  /**\n   * Which layers count as deterministic-reward sources. The verifier doesn't\n   * tag layers as \"this is verifiable\"; the caller declares it via this list\n   * (or via the layer name → source mapping). Default treats common names\n   * (`install`, `typecheck`, `build`, `lint`, `test`, `compile`, `schema`,\n   * `sandbox`) as deterministic.\n   */\n  deterministicLayers?: string[]\n  /**\n   * Map layer name → reward source. Defaults to a sensible string-match.\n   */\n  sourceFor?: (layerName: string) => VerifiableRewardSource\n  /**\n   * Whether to fall back to a probabilistic (judge) reward when no\n   * deterministic layer produced a numeric score. Default `true`. Set to\n   * `false` for \"deterministic-only\" training pipelines that should\n   * discard runs without a verifiable signal.\n   */\n  fallbackToJudge?: boolean\n  /**\n   * Default confidence for probabilistic (judge) rewards when the judge\n   * doesn't report one. Default `0.7`.\n   */\n  judgeConfidenceFloor?: number\n  /**\n   * Whether the anti-Goodhart realness gate applies. Default `true`, and the\n   * default is the one every training path must keep.\n   *\n   * Set `false` ONLY for detection and analysis. `rl/reward-hacking.ts` does,\n   * for the same reason it reads `observedScore` for its proxy: it measures the\n   * DIVERGENCE between the judge signal and the deterministic one, and a\n   * deterministic reward that another gate already forced to 0 manufactures\n   * exactly that divergence on exactly the gamed population. The detector would\n   * then be re-reporting a verdict it was supposed to reach independently.\n   */\n  applyRealnessGate?: boolean\n}\n\n// 'agreement' is deliberately absent: agreement between fixed artifacts is\n// mechanical, but the independent derivation it certifies is stochastic, so\n// a producer that wants a deterministic agreement reward must declare it\n// (via `deterministicLayers` or by constructing the reward directly).\nconst DEFAULT_DETERMINISTIC_LAYERS = new Set([\n  'install',\n  'typecheck',\n  'build',\n  'lint',\n  'test',\n  'compile',\n  'schema',\n  'sandbox',\n  'unit_tests',\n  'integration_tests',\n  'proof_kernel',\n  'proof-kernel',\n  'invariant',\n  'metamorphic',\n  'replication',\n])\n\nconst DEFAULT_SOURCE_FOR = (name: string): VerifiableRewardSource => {\n  const lower = name.toLowerCase()\n  if (lower.includes('test')) return 'test'\n  if (\n    lower.includes('compile') ||\n    lower.includes('build') ||\n    lower.includes('typecheck') ||\n    lower.includes('lint')\n  )\n    return 'compile'\n  if (lower.includes('schema')) return 'schema'\n  if (lower.includes('sandbox')) return 'sandbox'\n  if (lower.includes('judge') || lower.includes('semantic')) return 'judge'\n  // Answer-key names above take precedence: a name like 'proof_test' maps\n  // to 'test'; only names with no answer-key match reach the open-family\n  // members.\n  if (lower.includes('proof') || lower.includes('kernel') || lower.includes('lean')) {\n    return 'proof-kernel'\n  }\n  if (lower.includes('invariant') || lower.includes('metamorphic')) return 'invariant'\n  if (lower.includes('replicat')) return 'replication'\n  if (lower.includes('agreement') || lower.includes('consensus')) return 'agreement'\n  return 'composite'\n}\n\n/**\n * Extract a `VerifiableReward` from a `VerificationReport`.\n *\n * Strategy: prefer the deterministic layers (in order: test → compile →\n * schema → sandbox), fall back to the judge layer if `fallbackToJudge` is\n * true, return `null` if no signal qualifies. When multiple deterministic\n * layers contribute, return a `'composite'` source with a weighted blend.\n *\n * NO realness gate is applied and none can be: a `VerificationReport` carries\n * layer scores and nothing about whether the run faked them — `realness` lives\n * on the `RunRecord`. Use `extractVerifiableRewardsFromRecords` for anything\n * that becomes training data; this signature is for scoring a report in hand.\n */\nexport function extractVerifiableReward(\n  report: VerificationReport,\n  opts: VerifiableRewardExtractionOptions = {},\n): VerifiableReward | null {\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n\n  const deterministic = report.layers.filter(\n    (layer) => deterministicSet.has(layer.layer) && isMeasuredLayer(layer),\n  )\n\n  if (deterministic.length === 1) {\n    const layer = deterministic[0]!\n    const value = clamp01(layer.score!)\n    return {\n      value,\n      source: sourceFor(layer.layer),\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: layer.layer,\n      components: { [layer.layer]: value },\n      realnessScreened: false,\n    }\n  }\n\n  if (deterministic.length > 1) {\n    // Composite: weighted blend by `Layer.weight` if present, else equal.\n    let num = 0\n    let denom = 0\n    const components: Record<string, number> = {}\n    for (const l of deterministic) {\n      const w = (l.detail?.weight as number | undefined) ?? 1\n      num += w * (l.score ?? 0)\n      denom += w\n      components[l.layer] = l.score!\n    }\n    return {\n      value: denom === 0 ? 0 : clamp01(num / denom),\n      source: 'composite',\n      determinism: 'deterministic',\n      confidence: 1,\n      origin: deterministic.map((l) => l.layer).join('+'),\n      components,\n      realnessScreened: false,\n    }\n  }\n\n  if (!fallbackToJudge) return null\n\n  const judge =\n    report.layers.find((layer) => isMeasuredLayer(layer) && sourceFor(layer.layer) === 'judge') ??\n    report.layers.find(isMeasuredLayer)\n\n  if (!judge) return null\n\n  const confFromDetail = judge.detail?.confidence as number | undefined\n  const judgeValue = clamp01(judge.score!)\n  return {\n    value: judgeValue,\n    source: 'judge',\n    determinism: 'probabilistic',\n    confidence: typeof confFromDetail === 'number' ? confFromDetail : judgeFloor,\n    origin: judge.layer,\n    components: { [judge.layer]: judgeValue },\n    realnessScreened: false,\n  }\n}\n\nfunction isMeasuredLayer(\n  layer: LayerResult,\n): layer is LayerResult & { status: 'pass' | 'fail'; score: number } {\n  return (\n    (layer.status === 'pass' || layer.status === 'fail') &&\n    typeof layer.score === 'number' &&\n    Number.isFinite(layer.score) &&\n    layer.score >= 0 &&\n    layer.score <= 1\n  )\n}\n\n/**\n * Extract verifiable rewards from `RunRecord[]` produced via the\n * `verificationReportToRunRecord` adapter (which encodes per-layer scores\n * in `outcome.raw['layer.<name>']`). For records that don't carry layer\n * scores, returns `null` for that record.\n *\n * This is the canonical bridge from \"campaign-shaped artifacts\" to\n * \"RL-training-ready reward signals\": every record that has a clean\n * verifiable reward becomes a training datum, every record that doesn't\n * gets filtered out (or kept with `'probabilistic'` determinism for\n * separate downstream handling).\n *\n * The realness gate applies to EVERY channel here, and to the deterministic one\n * MOST. It is tempting to reason that a decidable signal cannot be gamed, so\n * the gate is redundant on it — that reasoning is backwards. `realness.gated`\n * means the run's success signal was FAKED, and a test suite reporting green on\n * a stubbed integration is precisely what that looks like: the deterministic\n * layer is the thing that got faked. Exporting it ungated hands a trainer the\n * highest-credibility reward the module can emit (`determinism: 'deterministic'`,\n * `confidence: 1`) for the one population the gate exists to catch. Pass\n * `applyRealnessGate: false` only to look at the ungated numbers for detection.\n */\nexport function extractVerifiableRewardsFromRecords(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ runId: string; reward: VerifiableReward | null }> {\n  const sourceFor = opts.sourceFor ?? DEFAULT_SOURCE_FOR\n  const deterministicSet = new Set(opts.deterministicLayers ?? [...DEFAULT_DETERMINISTIC_LAYERS])\n  const fallbackToJudge = opts.fallbackToJudge ?? true\n  const judgeFloor = opts.judgeConfidenceFloor ?? 0.7\n  const applyGate = opts.applyRealnessGate ?? true\n\n  return runs.map((run) => {\n    const flagged = isRealnessGated(run)\n    // Present only when the record carries an actual realness verdict. Absent\n    // is the honest \"unknown\"; `false` is reserved for a producer that\n    // declares it HAS no screen, which is the report-shaped path above.\n    const screened = run.outcome.realness === undefined ? {} : ({ realnessScreened: true } as const)\n    // Zeroed with `value`, never left at the measured number: `components`\n    // exists so an RL consumer can re-weight per source, and a raw layer score\n    // surviving there would let that re-weighting reconstruct the very reward\n    // the gate just refused. The measured layer scores stay on\n    // `run.outcome.raw['layer.*']`, which is where analysis reads them.\n    const gate = (value: number): number => (applyGate && flagged ? 0 : value)\n    // Recover per-layer scores from outcome.raw['layer.<name>']\n    const layerScores: Array<{ name: string; score: number }> = []\n    for (const [k, v] of Object.entries(run.outcome.raw)) {\n      if (\n        k.startsWith('layer.') &&\n        !k.includes('.', 6) &&\n        typeof v === 'number' &&\n        Number.isFinite(v)\n      ) {\n        layerScores.push({ name: k.slice('layer.'.length), score: v })\n      }\n    }\n    const det = layerScores.filter((l) => deterministicSet.has(l.name))\n\n    if (det.length === 1) {\n      const layer = det[0]!\n      const value = gate(clamp01(layer.score))\n      return {\n        runId: run.runId,\n        reward: {\n          value,\n          source: sourceFor(layer.name),\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: layer.name,\n          components: { [layer.name]: value },\n          realnessGated: flagged,\n          ...screened,\n        },\n      }\n    }\n    if (det.length > 1) {\n      const value = gate(clamp01(det.reduce((s, l) => s + l.score, 0) / det.length))\n      // Same clamp as the headline value: a producer writing layer.score 1.5\n      // into outcome.raw must not propagate 1.5 through a component either.\n      const components: Record<string, number> = Object.fromEntries(\n        det.map((l) => [l.name, gate(clamp01(l.score))]),\n      )\n      return {\n        runId: run.runId,\n        reward: {\n          value,\n          source: 'composite',\n          determinism: 'deterministic',\n          confidence: 1,\n          origin: det.map((l) => l.name).join('+'),\n          components,\n          realnessGated: flagged,\n          ...screened,\n        },\n      }\n    }\n    if (!fallbackToJudge) return { runId: run.runId, reward: null }\n\n    // Probabilistic fallback: the run's primary score. `trainingScore` already\n    // carries the gate, so a gamed run falls to 0 rather than earning the\n    // judge's number; `observedScore` is the ungated reader the detection\n    // opt-out asks for. Either way an unscored run stays a labeled gap\n    // (`reward: null`), never a fabricated 0.\n    const primary = applyGate ? trainingScore(run) : observedScore(run)\n    if (typeof primary !== 'number' || !Number.isFinite(primary)) {\n      return { runId: run.runId, reward: null }\n    }\n    const primaryValue = clamp01(primary)\n    return {\n      runId: run.runId,\n      reward: {\n        value: primaryValue,\n        source: 'judge',\n        determinism: 'probabilistic',\n        confidence: judgeFloor,\n        origin: 'run.outcome.score',\n        components: { 'run.outcome.score': primaryValue },\n        realnessGated: flagged,\n        ...screened,\n      },\n    }\n  })\n}\n\n/**\n * Filter `RunRecord[]` to those with deterministic verifiable rewards.\n *\n * A realness-gated run is KEPT, at reward 0 with `realnessGated: true` — the\n * same rule GRPO uses on a gated line. 0 is the honest label for a faked\n * success and is usable signal, whereas dropping the run would move a group\n * baseline without saying so. (SFT differs: there every row is a target to\n * imitate, so a gated row is removed outright.)\n */\nexport function filterDeterministicallyRewarded(\n  runs: RunRecord[],\n  opts: VerifiableRewardExtractionOptions = {},\n): Array<{ run: RunRecord; reward: VerifiableReward }> {\n  const rewarded = extractVerifiableRewardsFromRecords(runs, { ...opts, fallbackToJudge: false })\n  const out: Array<{ run: RunRecord; reward: VerifiableReward }> = []\n  for (let i = 0; i < runs.length; i++) {\n    const r = rewarded[i]!\n    if (r.reward && r.reward.determinism === 'deterministic') {\n      out.push({ run: runs[i]!, reward: r.reward })\n    }\n  }\n  return out\n}\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n","/**\n * Reward hacking / Goodhart detection.\n *\n * Goodhart's Law says: when a measure becomes a target, it ceases to be\n * a good measure. In RLHF and agentic-RL settings this is the dominant\n * failure mode — the policy learns to produce outputs that score well on\n * the proxy reward (judge, rubric, test pass-rate) without producing\n * the underlying capability the proxy was meant to track.\n *\n * Krakovna et al. (2020, \"Specification Gaming Examples in AI\") and the\n * subsequent RLHF reward-hacking literature (Skalse et al. 2022, Kim et al.\n * 2023) converge on a few diagnostic signatures:\n *\n *   1. **Reward divergence:** the proxy reward grows while the held-out\n *      ground-truth signal stagnates or drops. Predictive validity over\n *      time captures this.\n *   2. **Distributional shift in outputs:** after RL, the policy produces\n *      outputs that no longer match the reference distribution — usually\n *      because it found a high-reward attractor that's degenerate (e.g.\n *      one-token responses, repetition, formatting tricks).\n *   3. **Disagreement between independent rewards:** if you train on\n *      reward A and a held-out independent reward B drops sharply, you're\n *      probably hacking A.\n *   4. **Calibration drift:** the verifiable / deterministic component of\n *      the reward is stable; the probabilistic / judge component drifts up\n *      while the deterministic component doesn't. The judge is being\n *      gamed.\n *\n * This module ships explicit detectors for all four signatures, plus a\n * combined verdict. The output is diagnostic — actionable signals,\n * not autoreject — because each signature has known false positives\n * (e.g., a policy that genuinely improves can show distributional shift).\n *\n * Differs from `rubricPredictiveValidity` (which is a *standing* check on\n * whether rubrics correlate with deployment outcomes) — this is a\n * *temporal* check on whether the reward-vs-truth gap is *widening over\n * time during a training run*.\n */\n\nimport { observedScore } from '../rollout/reward'\nimport type { RunRecord } from '../run-record'\nimport { pearsonR } from '../statistics'\nimport {\n  filterDeterministicallyRewarded,\n  type VerifiableRewardExtractionOptions,\n} from './verifiable-reward'\n\nexport type RewardHackingSignal =\n  | 'reward_divergence'\n  | 'distribution_shift'\n  | 'reward_disagreement'\n  | 'judge_drift'\n\nexport interface RewardHackingFinding {\n  signal: RewardHackingSignal\n  /** Severity in [0, 1]. >0.5 = strong signal. */\n  severity: number\n  message: string\n  /** Numeric evidence the consumer can render. */\n  detail: Record<string, number>\n}\n\nexport interface RewardHackingReport {\n  findings: RewardHackingFinding[]\n  /** Signals with enough usable observations to produce a finding. */\n  evaluatedSignals: RewardHackingSignal[]\n  /**\n   * Composite verdict. `'insufficient_evidence'` when fewer than four scored\n   * runs exist; otherwise `'clean'` if every signal severity < 0.3,\n   * `'suspect'` if at least one ≥ 0.3 but none ≥ 0.6, and `'gaming'` if any ≥ 0.6.\n   */\n  verdict: 'insufficient_evidence' | 'clean' | 'suspect' | 'gaming'\n  /** Rationale for the verdict, ready to paste into an audit log. */\n  rationale: string[]\n  /** Number of runs with a usable proxy reward. */\n  n: number\n}\n\nexport interface DetectRewardHackingInput {\n  /**\n   * Run records ordered by recency (oldest first). The detector segments\n   * them into prefix/suffix windows to compute \"did the gap widen.\"\n   */\n  runs: RunRecord[]\n  /**\n   * The metric the policy was trained to optimize. Should be present on\n   * `outcome.raw` or `outcome.holdoutScore`. Default reads `outcome.holdoutScore`.\n   */\n  proxyOf?: (run: RunRecord) => number | null\n  /**\n   * The held-out ground-truth metric. For RL on coding, this is typically\n   * test pass-rate. For RLHF, it's downstream task performance or human\n   * preference. For knowledge tasks, it's an independently-graded score.\n   */\n  truthOf?: (run: RunRecord) => number | null\n  /**\n   * Independent secondary reward. Used for the `reward_disagreement`\n   * signal. Default uses the verifiable reward extractor (deterministic\n   * sources only).\n   */\n  secondaryRewardOf?: (run: RunRecord) => number | null\n  /**\n   * Window size — how many of the most recent runs count as the \"after\"\n   * cohort. Default min(50, half the runs).\n   */\n  windowSize?: number\n  /**\n   * Severity threshold to flag a signal. Default 0.3 (suspect) and 0.6\n   * (gaming).\n   */\n  thresholds?: { suspect?: number; gaming?: number }\n  /**\n   * Verifiable-reward options used for the secondary-reward fallback.\n   */\n  verifiableRewardOptions?: VerifiableRewardExtractionOptions\n}\n\nconst DEFAULT_PROXY = (r: RunRecord): number | null => {\n  // DELIBERATELY UNGATED. This is the proxy reward the detector tests for\n  // Goodharting, and gated runs are exactly the gamed population. Forcing them\n  // to 0 would collapse the proxy toward the deterministic secondary signal —\n  // `reward_disagreement`'s correlation would rise, `judge_drift`'s gap would\n  // shrink, and `reward_divergence`'s proxy-up/truth-flat fingerprint would be\n  // erased — so the detector would report \"clean\" on the very runs it exists to\n  // catch. `null` (never 0) is also load-bearing: it sets the n-denominator via\n  // the filter below.\n  const v = observedScore(r)\n  return typeof v === 'number' && Number.isFinite(v) ? v : null\n}\n\nexport function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport {\n  const proxyOf = input.proxyOf ?? DEFAULT_PROXY\n  const truthOf = input.truthOf\n  const sus = input.thresholds?.suspect ?? 0.3\n  const gam = input.thresholds?.gaming ?? 0.6\n\n  const runs = input.runs.filter((run) => finiteNumber(proxyOf(run)))\n  const n = runs.length\n  if (n < 4) {\n    return {\n      findings: [],\n      evaluatedSignals: [],\n      verdict: 'insufficient_evidence',\n      n,\n      rationale: [`fewer than 4 runs with proxy reward (n=${n}); insufficient evidence`],\n    }\n  }\n  const windowSize = Math.max(1, input.windowSize ?? Math.min(50, Math.floor(n / 2)))\n  const before = runs.slice(0, n - windowSize)\n  const after = runs.slice(n - windowSize)\n\n  const findings: RewardHackingFinding[] = []\n\n  // ── Signal 1: reward divergence (proxy ↑ while truth flat or ↓) ──────\n  if (truthOf) {\n    const beforeProxy = before.map(proxyOf).filter(finiteNumber)\n    const afterProxy = after.map(proxyOf).filter(finiteNumber)\n    const beforeTruth = before.map(truthOf).filter(finiteNumber)\n    const afterTruth = after.map(truthOf).filter(finiteNumber)\n    if (\n      beforeProxy.length >= 2 &&\n      afterProxy.length >= 2 &&\n      beforeTruth.length >= 2 &&\n      afterTruth.length >= 2\n    ) {\n      const proxyDelta = mean(afterProxy) - mean(beforeProxy)\n      const truthDelta = mean(afterTruth) - mean(beforeTruth)\n      // Divergence: proxy goes up while truth goes flat or down.\n      // Severity = max(0, (proxyDelta - truthDelta)) — bigger gap = bigger signal.\n      const gap = Math.max(0, proxyDelta - truthDelta)\n      const severity = clamp01(gap * 5) // scale: 0.2 absolute gap → severity 1.0\n      findings.push({\n        signal: 'reward_divergence',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy reward rose by ${proxyDelta.toFixed(3)} while truth changed by ${truthDelta.toFixed(3)} — potential Goodhart`\n            : `proxy and truth moved together (proxy ${proxyDelta.toFixed(3)}, truth ${truthDelta.toFixed(3)})`,\n        detail: {\n          proxyDelta,\n          truthDelta,\n          gap,\n          beforeN: beforeProxy.length,\n          afterN: afterProxy.length,\n        },\n      })\n    }\n  }\n\n  // ── Signal 2: distributional shift in outputs (KS on score distributions) ──\n  {\n    const beforeP = before.map(proxyOf).filter(finiteNumber)\n    const afterP = after.map(proxyOf).filter(finiteNumber)\n    if (beforeP.length >= 4 && afterP.length >= 4) {\n      const ks = ksStatistic(beforeP, afterP)\n      // KS statistic: bigger = more shift. We're agnostic about direction;\n      // genuine improvement ALSO produces shift, so this signal is\n      // contributory rather than load-bearing.\n      const severity = clamp01(ks - 0.2)\n      findings.push({\n        signal: 'distribution_shift',\n        severity,\n        message:\n          severity >= sus\n            ? `KS=${ks.toFixed(3)} between before/after windows — distributional shift large`\n            : `KS=${ks.toFixed(3)} between before/after windows — within-distribution drift`,\n        detail: { ks, beforeN: beforeP.length, afterN: afterP.length },\n      })\n    }\n  }\n\n  // ── Signal 3: reward disagreement (proxy vs independent secondary) ────\n  {\n    const secondaryOf = input.secondaryRewardOf ?? defaultSecondary(input.verifiableRewardOptions)\n    const aligned = runs\n      .map((r) => ({ p: proxyOf(r), s: secondaryOf(r) }))\n      .filter((x): x is { p: number; s: number } => finiteNumber(x.p) && finiteNumber(x.s))\n    if (aligned.length >= 4) {\n      const ps = aligned.map((x) => x.p)\n      const ss = aligned.map((x) => x.s)\n      const r = pearsonR(ps, ss)\n      // Disagreement: low or negative correlation between primary proxy\n      // reward and an independent secondary signal.\n      const severity = clamp01(0.5 - Math.max(0, r))\n      findings.push({\n        signal: 'reward_disagreement',\n        severity,\n        message:\n          severity >= sus\n            ? `proxy and independent secondary reward correlate ρ=${r.toFixed(3)} — possibly hacking proxy`\n            : `proxy and secondary reward correlate ρ=${r.toFixed(3)}`,\n        detail: { pearson: r, n: aligned.length },\n      })\n    }\n  }\n\n  // ── Signal 4: judge drift (probabilistic up while deterministic flat) ─\n  {\n    // Ungated on purpose, exactly like `DEFAULT_PROXY` above. This signal is\n    // the GAP between the judge reward and the deterministic one; a\n    // deterministic reward another gate already forced to 0 would open that gap\n    // by construction on the gamed population, so the detector would fire on\n    // its own input rather than on evidence it found.\n    const detRuns = filterDeterministicallyRewarded(runs, {\n      ...(input.verifiableRewardOptions ?? {}),\n      applyRealnessGate: false,\n    })\n    if (detRuns.length >= 4) {\n      const detBefore = detRuns.slice(0, Math.floor(detRuns.length / 2))\n      const detAfter = detRuns.slice(Math.floor(detRuns.length / 2))\n      const detDelta =\n        mean(detAfter.map((r) => r.reward.value)) - mean(detBefore.map((r) => r.reward.value))\n      const proxyDelta =\n        mean(after.map(proxyOf).filter(finiteNumber)) -\n        mean(before.map(proxyOf).filter(finiteNumber))\n      const driftGap = Math.max(0, proxyDelta - detDelta)\n      const severity = clamp01(driftGap * 5)\n      findings.push({\n        signal: 'judge_drift',\n        severity,\n        message:\n          severity >= sus\n            ? `judge proxy +${proxyDelta.toFixed(3)} while deterministic reward +${detDelta.toFixed(3)} — judge drifting up without verifiable backing`\n            : `judge and deterministic rewards move in step (judge ${proxyDelta.toFixed(3)}, det ${detDelta.toFixed(3)})`,\n        detail: { proxyDelta, detDelta, driftGap, n: detRuns.length },\n      })\n    }\n  }\n\n  const maxSev = findings.reduce((m, f) => Math.max(m, f.severity), 0)\n  if (findings.length === 0) {\n    return {\n      findings,\n      evaluatedSignals: [],\n      verdict: 'insufficient_evidence',\n      rationale: [`no reward-hacking signal had enough paired evidence (n=${n})`],\n      n,\n    }\n  }\n  const verdict: RewardHackingReport['verdict'] =\n    maxSev >= gam ? 'gaming' : maxSev >= sus ? 'suspect' : 'clean'\n  const rationale = findings\n    .filter((f) => f.severity >= sus)\n    .map((f) => `${f.signal}: severity ${f.severity.toFixed(2)} — ${f.message}`)\n  if (rationale.length === 0) rationale.push('no signals fired above suspect threshold')\n\n  return {\n    findings,\n    evaluatedSignals: findings.map((finding) => finding.signal),\n    verdict,\n    rationale,\n    n,\n  }\n}\n\n// ── Helpers ──────────────────────────────────────────────────────────────\n\nfunction mean(xs: number[]): number {\n  if (xs.length === 0) return 0\n  return xs.reduce((s, x) => s + x, 0) / xs.length\n}\n\nfunction finiteNumber(value: number | null): value is number {\n  return typeof value === 'number' && Number.isFinite(value)\n}\n\nfunction clamp01(x: number): number {\n  if (!Number.isFinite(x)) return 0\n  return Math.max(0, Math.min(1, x))\n}\n\nfunction ksStatistic(a: number[], b: number[]): number {\n  // Two-sample Kolmogorov-Smirnov statistic.\n  const sortedA = [...a].sort((x, y) => x - y)\n  const sortedB = [...b].sort((x, y) => x - y)\n  const all = [...new Set([...sortedA, ...sortedB])].sort((x, y) => x - y)\n  let max = 0\n  for (const v of all) {\n    const fa = sortedA.filter((x) => x <= v).length / sortedA.length\n    const fb = sortedB.filter((x) => x <= v).length / sortedB.length\n    max = Math.max(max, Math.abs(fa - fb))\n  }\n  return max\n}\n\nfunction defaultSecondary(\n  verifiableOpts?: VerifiableRewardExtractionOptions,\n): (run: RunRecord) => number | null {\n  return (run: RunRecord) => {\n    // Ungated for the same reason as signal 4: this is the INDEPENDENT\n    // secondary reward whose correlation with the proxy is the evidence.\n    // Zeroing it on gated runs would drive that correlation down mechanically.\n    const filtered = filterDeterministicallyRewarded([run], {\n      ...(verifiableOpts ?? {}),\n      applyRealnessGate: false,\n    })\n    return filtered.length === 1 ? filtered[0]!.reward.value : null\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;AA8CA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,qBAAqB,2BAA2B,IAAI;CAC1D,MAAM,iBACJ,mBAAmB,SAAS,gBAAgB,QAAQ,mBAAmB,KAAA,IACnE;EAAE,MAAM;EAAa,KAAK,QAAQ;CAAe,IACjD;CACN,MAAM,UAAU,eAAe,SAAS,eAAe,OAAO,eAAe;CAC7E,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAEhD,GAAI,mBAAmB,SAAS,eAAe,EAAE,yBAAyB,KAAK,QAAQ,IAAI,CAAC;EAC5F,eAAe,eAAe,SAAS,aAAa,IAAI;EACxD,gBAAgB,eAAe,SAAS,cAAc,IAAI;EAC1D,iBAAiB,eAAe,SAAS,eAAe,IAAI;EAC5D,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;CACvE,CAAC;AACH;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAYA,OAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAWA,OAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAASA,OAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE;;;ACtKA,MAAM,+CAA+B,IAAI,IAAI;CAC3C;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;AAED,MAAM,sBAAsB,SAAyC;CACnE,MAAM,QAAQ,KAAK,YAAY;CAC/B,IAAI,MAAM,SAAS,MAAM,GAAG,OAAO;CACnC,IACE,MAAM,SAAS,SAAS,KACxB,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,MAAM,GAErB,OAAO;CACT,IAAI,MAAM,SAAS,QAAQ,GAAG,OAAO;CACrC,IAAI,MAAM,SAAS,SAAS,GAAG,OAAO;CACtC,IAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,UAAU,GAAG,OAAO;CAIlE,IAAI,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,QAAQ,KAAK,MAAM,SAAS,MAAM,GAC9E,OAAO;CAET,IAAI,MAAM,SAAS,WAAW,KAAK,MAAM,SAAS,aAAa,GAAG,OAAO;CACzE,IAAI,MAAM,SAAS,UAAU,GAAG,OAAO;CACvC,IAAI,MAAM,SAAS,WAAW,KAAK,MAAM,SAAS,WAAW,GAAG,OAAO;CACvE,OAAO;AACT;;;;;;;;;;;;;;AAeA,SAAgB,wBACd,QACA,OAA0C,CAAC,GAClB;CACzB,MAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;CAC9F,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,aAAa,KAAK,wBAAwB;CAEhD,MAAM,gBAAgB,OAAO,OAAO,QACjC,UAAU,iBAAiB,IAAI,MAAM,KAAK,KAAK,gBAAgB,KAAK,CACvE;CAEA,IAAI,cAAc,WAAW,GAAG;EAC9B,MAAM,QAAQ,cAAc;EAC5B,MAAM,QAAQC,UAAQ,MAAM,KAAM;EAClC,OAAO;GACL;GACA,QAAQ,UAAU,MAAM,KAAK;GAC7B,aAAa;GACb,YAAY;GACZ,QAAQ,MAAM;GACd,YAAY,GAAG,MAAM,QAAQ,MAAM;GACnC,kBAAkB;EACpB;CACF;CAEA,IAAI,cAAc,SAAS,GAAG;EAE5B,IAAI,MAAM;EACV,IAAI,QAAQ;EACZ,MAAM,aAAqC,CAAC;EAC5C,KAAK,MAAM,KAAK,eAAe;GAC7B,MAAM,IAAK,EAAE,QAAQ,UAAiC;GACtD,OAAO,KAAK,EAAE,SAAS;GACvB,SAAS;GACT,WAAW,EAAE,SAAS,EAAE;EAC1B;EACA,OAAO;GACL,OAAO,UAAU,IAAI,IAAIA,UAAQ,MAAM,KAAK;GAC5C,QAAQ;GACR,aAAa;GACb,YAAY;GACZ,QAAQ,cAAc,KAAK,MAAM,EAAE,KAAK,CAAC,CAAC,KAAK,GAAG;GAClD;GACA,kBAAkB;EACpB;CACF;CAEA,IAAI,CAAC,iBAAiB,OAAO;CAE7B,MAAM,QACJ,OAAO,OAAO,MAAM,UAAU,gBAAgB,KAAK,KAAK,UAAU,MAAM,KAAK,MAAM,OAAO,KAC1F,OAAO,OAAO,KAAK,eAAe;CAEpC,IAAI,CAAC,OAAO,OAAO;CAEnB,MAAM,iBAAiB,MAAM,QAAQ;CACrC,MAAM,aAAaA,UAAQ,MAAM,KAAM;CACvC,OAAO;EACL,OAAO;EACP,QAAQ;EACR,aAAa;EACb,YAAY,OAAO,mBAAmB,WAAW,iBAAiB;EAClE,QAAQ,MAAM;EACd,YAAY,GAAG,MAAM,QAAQ,WAAW;EACxC,kBAAkB;CACpB;AACF;AAEA,SAAS,gBACP,OACmE;CACnE,QACG,MAAM,WAAW,UAAU,MAAM,WAAW,WAC7C,OAAO,MAAM,UAAU,YACvB,OAAO,SAAS,MAAM,KAAK,KAC3B,MAAM,SAAS,KACf,MAAM,SAAS;AAEnB;;;;;;;;;;;;;;;;;;;;;;;AAwBA,SAAgB,oCACd,MACA,OAA0C,CAAC,GACgB;CAC3D,MAAM,YAAY,KAAK,aAAa;CACpC,MAAM,mBAAmB,IAAI,IAAI,KAAK,uBAAuB,CAAC,GAAG,4BAA4B,CAAC;CAC9F,MAAM,kBAAkB,KAAK,mBAAmB;CAChD,MAAM,aAAa,KAAK,wBAAwB;CAChD,MAAM,YAAY,KAAK,qBAAqB;CAE5C,OAAO,KAAK,KAAK,QAAQ;EACvB,MAAM,UAAU,gBAAgB,GAAG;EAInC,MAAM,WAAW,IAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAK,EAAE,kBAAkB,KAAK;EAMrF,MAAM,QAAQ,UAA2B,aAAa,UAAU,IAAI;EAEpE,MAAM,cAAsD,CAAC;EAC7D,KAAK,MAAM,CAAC,GAAG,MAAM,OAAO,QAAQ,IAAI,QAAQ,GAAG,GACjD,IACE,EAAE,WAAW,QAAQ,KACrB,CAAC,EAAE,SAAS,KAAK,CAAC,KAClB,OAAO,MAAM,YACb,OAAO,SAAS,CAAC,GAEjB,YAAY,KAAK;GAAE,MAAM,EAAE,MAAM,CAAe;GAAG,OAAO;EAAE,CAAC;EAGjE,MAAM,MAAM,YAAY,QAAQ,MAAM,iBAAiB,IAAI,EAAE,IAAI,CAAC;EAElE,IAAI,IAAI,WAAW,GAAG;GACpB,MAAM,QAAQ,IAAI;GAClB,MAAM,QAAQ,KAAKA,UAAQ,MAAM,KAAK,CAAC;GACvC,OAAO;IACL,OAAO,IAAI;IACX,QAAQ;KACN;KACA,QAAQ,UAAU,MAAM,IAAI;KAC5B,aAAa;KACb,YAAY;KACZ,QAAQ,MAAM;KACd,YAAY,GAAG,MAAM,OAAO,MAAM;KAClC,eAAe;KACf,GAAG;IACL;GACF;EACF;EACA,IAAI,IAAI,SAAS,GAAG;GAClB,MAAM,QAAQ,KAAKA,UAAQ,IAAI,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,IAAI,MAAM,CAAC;GAG7E,MAAM,aAAqC,OAAO,YAChD,IAAI,KAAK,MAAM,CAAC,EAAE,MAAM,KAAKA,UAAQ,EAAE,KAAK,CAAC,CAAC,CAAC,CACjD;GACA,OAAO;IACL,OAAO,IAAI;IACX,QAAQ;KACN;KACA,QAAQ;KACR,aAAa;KACb,YAAY;KACZ,QAAQ,IAAI,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,GAAG;KACvC;KACA,eAAe;KACf,GAAG;IACL;GACF;EACF;EACA,IAAI,CAAC,iBAAiB,OAAO;GAAE,OAAO,IAAI;GAAO,QAAQ;EAAK;EAO9D,MAAM,UAAU,YAAY,cAAc,GAAG,IAAI,cAAc,GAAG;EAClE,IAAI,OAAO,YAAY,YAAY,CAAC,OAAO,SAAS,OAAO,GACzD,OAAO;GAAE,OAAO,IAAI;GAAO,QAAQ;EAAK;EAE1C,MAAM,eAAeA,UAAQ,OAAO;EACpC,OAAO;GACL,OAAO,IAAI;GACX,QAAQ;IACN,OAAO;IACP,QAAQ;IACR,aAAa;IACb,YAAY;IACZ,QAAQ;IACR,YAAY,EAAE,qBAAqB,aAAa;IAChD,eAAe;IACf,GAAG;GACL;EACF;CACF,CAAC;AACH;;;;;;;;;;AAWA,SAAgB,gCACd,MACA,OAA0C,CAAC,GACU;CACrD,MAAM,WAAW,oCAAoC,MAAM;EAAE,GAAG;EAAM,iBAAiB;CAAM,CAAC;CAC9F,MAAM,MAA2D,CAAC;CAClE,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,IAAI,SAAS;EACnB,IAAI,EAAE,UAAU,EAAE,OAAO,gBAAgB,iBACvC,IAAI,KAAK;GAAE,KAAK,KAAK;GAAK,QAAQ,EAAE;EAAO,CAAC;CAEhD;CACA,OAAO;AACT;AAEA,SAASA,UAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;ACnVA,MAAM,iBAAiB,MAAgC;CASrD,MAAM,IAAI,cAAc,CAAC;CACzB,OAAO,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,IAAI,IAAI;AAC3D;AAEA,SAAgB,oBAAoB,OAAsD;CACxF,MAAM,UAAU,MAAM,WAAW;CACjC,MAAM,UAAU,MAAM;CACtB,MAAM,MAAM,MAAM,YAAY,WAAW;CACzC,MAAM,MAAM,MAAM,YAAY,UAAU;CAExC,MAAM,OAAO,MAAM,KAAK,QAAQ,QAAQ,aAAa,QAAQ,GAAG,CAAC,CAAC;CAClE,MAAM,IAAI,KAAK;CACf,IAAI,IAAI,GACN,OAAO;EACL,UAAU,CAAC;EACX,kBAAkB,CAAC;EACnB,SAAS;EACT;EACA,WAAW,CAAC,0CAA0C,EAAE,yBAAyB;CACnF;CAEF,MAAM,aAAa,KAAK,IAAI,GAAG,MAAM,cAAc,KAAK,IAAI,IAAI,KAAK,MAAM,IAAI,CAAC,CAAC,CAAC;CAClF,MAAM,SAAS,KAAK,MAAM,GAAG,IAAI,UAAU;CAC3C,MAAM,QAAQ,KAAK,MAAM,IAAI,UAAU;CAEvC,MAAM,WAAmC,CAAC;CAG1C,IAAI,SAAS;EACX,MAAM,cAAc,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EAC3D,MAAM,aAAa,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACzD,MAAM,cAAc,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EAC3D,MAAM,aAAa,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACzD,IACE,YAAY,UAAU,KACtB,WAAW,UAAU,KACrB,YAAY,UAAU,KACtB,WAAW,UAAU,GACrB;GACA,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;GACtD,MAAM,aAAa,KAAK,UAAU,IAAI,KAAK,WAAW;GAGtD,MAAM,MAAM,KAAK,IAAI,GAAG,aAAa,UAAU;GAC/C,MAAM,WAAW,QAAQ,MAAM,CAAC;GAChC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,wBAAwB,WAAW,QAAQ,CAAC,EAAE,0BAA0B,WAAW,QAAQ,CAAC,EAAE,yBAC9F,yCAAyC,WAAW,QAAQ,CAAC,EAAE,UAAU,WAAW,QAAQ,CAAC,EAAE;IACrG,QAAQ;KACN;KACA;KACA;KACA,SAAS,YAAY;KACrB,QAAQ,WAAW;IACrB;GACF,CAAC;EACH;CACF;CAGA;EACE,MAAM,UAAU,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACvD,MAAM,SAAS,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY;EACrD,IAAI,QAAQ,UAAU,KAAK,OAAO,UAAU,GAAG;GAC7C,MAAM,KAAK,YAAY,SAAS,MAAM;GAItC,MAAM,WAAW,QAAQ,KAAK,EAAG;GACjC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,MAAM,GAAG,QAAQ,CAAC,EAAE,8DACpB,MAAM,GAAG,QAAQ,CAAC,EAAE;IAC1B,QAAQ;KAAE;KAAI,SAAS,QAAQ;KAAQ,QAAQ,OAAO;IAAO;GAC/D,CAAC;EACH;CACF;CAGA;EACE,MAAM,cAAc,MAAM,qBAAqB,iBAAiB,MAAM,uBAAuB;EAC7F,MAAM,UAAU,KACb,KAAK,OAAO;GAAE,GAAG,QAAQ,CAAC;GAAG,GAAG,YAAY,CAAC;EAAE,EAAE,CAAC,CAClD,QAAQ,MAAqC,aAAa,EAAE,CAAC,KAAK,aAAa,EAAE,CAAC,CAAC;EACtF,IAAI,QAAQ,UAAU,GAAG;GAGvB,MAAM,IAAI,SAFC,QAAQ,KAAK,MAAM,EAAE,CAEZ,GADT,QAAQ,KAAK,MAAM,EAAE,CACR,CAAC;GAGzB,MAAM,WAAW,QAAQ,KAAM,KAAK,IAAI,GAAG,CAAC,CAAC;GAC7C,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,sDAAsD,EAAE,QAAQ,CAAC,EAAE,6BACnE,0CAA0C,EAAE,QAAQ,CAAC;IAC3D,QAAQ;KAAE,SAAS;KAAG,GAAG,QAAQ;IAAO;GAC1C,CAAC;EACH;CACF;CAGA;EAME,MAAM,UAAU,gCAAgC,MAAM;GACpD,GAAI,MAAM,2BAA2B,CAAC;GACtC,mBAAmB;EACrB,CAAC;EACD,IAAI,QAAQ,UAAU,GAAG;GACvB,MAAM,YAAY,QAAQ,MAAM,GAAG,KAAK,MAAM,QAAQ,SAAS,CAAC,CAAC;GAEjE,MAAM,WACJ,KAFe,QAAQ,MAAM,KAAK,MAAM,QAAQ,SAAS,CAAC,CAE9C,CAAC,CAAC,KAAK,MAAM,EAAE,OAAO,KAAK,CAAC,IAAI,KAAK,UAAU,KAAK,MAAM,EAAE,OAAO,KAAK,CAAC;GACvF,MAAM,aACJ,KAAK,MAAM,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY,CAAC,IAC5C,KAAK,OAAO,IAAI,OAAO,CAAC,CAAC,OAAO,YAAY,CAAC;GAC/C,MAAM,WAAW,KAAK,IAAI,GAAG,aAAa,QAAQ;GAClD,MAAM,WAAW,QAAQ,WAAW,CAAC;GACrC,SAAS,KAAK;IACZ,QAAQ;IACR;IACA,SACE,YAAY,MACR,gBAAgB,WAAW,QAAQ,CAAC,EAAE,+BAA+B,SAAS,QAAQ,CAAC,EAAE,mDACzF,uDAAuD,WAAW,QAAQ,CAAC,EAAE,QAAQ,SAAS,QAAQ,CAAC,EAAE;IAC/G,QAAQ;KAAE;KAAY;KAAU;KAAU,GAAG,QAAQ;IAAO;GAC9D,CAAC;EACH;CACF;CAEA,MAAM,SAAS,SAAS,QAAQ,GAAG,MAAM,KAAK,IAAI,GAAG,EAAE,QAAQ,GAAG,CAAC;CACnE,IAAI,SAAS,WAAW,GACtB,OAAO;EACL;EACA,kBAAkB,CAAC;EACnB,SAAS;EACT,WAAW,CAAC,0DAA0D,EAAE,EAAE;EAC1E;CACF;CAEF,MAAM,UACJ,UAAU,MAAM,WAAW,UAAU,MAAM,YAAY;CACzD,MAAM,YAAY,SACf,QAAQ,MAAM,EAAE,YAAY,GAAG,CAAC,CAChC,KAAK,MAAM,GAAG,EAAE,OAAO,aAAa,EAAE,SAAS,QAAQ,CAAC,EAAE,KAAK,EAAE,SAAS;CAC7E,IAAI,UAAU,WAAW,GAAG,UAAU,KAAK,0CAA0C;CAErF,OAAO;EACL;EACA,kBAAkB,SAAS,KAAK,YAAY,QAAQ,MAAM;EAC1D;EACA;EACA;CACF;AACF;AAIA,SAAS,KAAK,IAAsB;CAClC,IAAI,GAAG,WAAW,GAAG,OAAO;CAC5B,OAAO,GAAG,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,GAAG;AAC5C;AAEA,SAAS,aAAa,OAAuC;CAC3D,OAAO,OAAO,UAAU,YAAY,OAAO,SAAS,KAAK;AAC3D;AAEA,SAAS,QAAQ,GAAmB;CAClC,IAAI,CAAC,OAAO,SAAS,CAAC,GAAG,OAAO;CAChC,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,CAAC,CAAC;AACnC;AAEA,SAAS,YAAY,GAAa,GAAqB;CAErD,MAAM,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,MAAM,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CAC3C,MAAM,MAAM,CAAC,mBAAG,IAAI,IAAI,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC;CACvE,IAAI,MAAM;CACV,KAAK,MAAM,KAAK,KAAK;EACnB,MAAM,KAAK,QAAQ,QAAQ,MAAM,KAAK,CAAC,CAAC,CAAC,SAAS,QAAQ;EAC1D,MAAM,KAAK,QAAQ,QAAQ,MAAM,KAAK,CAAC,CAAC,CAAC,SAAS,QAAQ;EAC1D,MAAM,KAAK,IAAI,KAAK,KAAK,IAAI,KAAK,EAAE,CAAC;CACvC;CACA,OAAO;AACT;AAEA,SAAS,iBACP,gBACmC;CACnC,QAAQ,QAAmB;EAIzB,MAAM,WAAW,gCAAgC,CAAC,GAAG,GAAG;GACtD,GAAI,kBAAkB,CAAC;GACvB,mBAAmB;EACrB,CAAC;EACD,OAAO,SAAS,WAAW,IAAI,SAAS,EAAE,CAAE,OAAO,QAAQ;CAC7D;AACF"}

package/dist/rl.d.ts

@@ -2,13 +2,13 @@ import { s as VerificationReport } from "./multi-layer-verifier-BUaQ4C17.js";
 import { _ as Span } from "./schema-BtVldJ3T.js";
 import { a as RunRecord, s as RunSplitTag } from "./run-record-DVV82Gwh.js";
 import { s as TraceStore } from "./store-CT9YIIve.js";
-import { a as CampaignResult } from "./types-C1Bmeb8X.js";
+import { a as CampaignResult } from "./types-DdFNuyxQ.js";
 import { b as RolloutSplit, o as MintedRolloutLine } from "./schema-Cef2cFmb.js";
 import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-62tojkQd.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "./outcome-store-BYHIuO0e.js";
 import { t as InterimReleaseConfidence } from "./sequential-BhsrMupG.js";
 import { t as AdversarialMutation } from "./adversarial-smnADNFS.js";
-import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-BySlpla_.js";
+import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-DJnoUE8c.js";
 import { r as RubricPredictiveValidityReport } from "./rubric-predictive-validity-C7LnNvF2.js";
 //#region src/rl/active-curriculum.d.ts
 interface CellObservation {

package/dist/rl.js

@@ -3,9 +3,9 @@ import { o as benjaminiHochberg } from "./power-and-mde-CHIrXJll.js";
 import { l as wilcoxonSignedRank } from "./paired-arms-D-XRF_fy.js";
 import { r as observedSplitScore, s as trainingScore } from "./reward-nw2xZGZG.js";
 import { s as runTaskScore } from "./run-record-D2lDdSAz.js";
-import { i as filterDeterministicallyRewarded, l as campaignCellToRunRecord, n as extractVerifiableReward, r as extractVerifiableRewardsFromRecords, t as detectRewardHacking } from "./reward-hacking-DFo2FU5J.js";
+import { i as filterDeterministicallyRewarded, l as campaignCellToRunRecord, n as extractVerifiableReward, r as extractVerifiableRewardsFromRecords, t as detectRewardHacking } from "./reward-hacking-DKI9T52l.js";
 import { n as InMemoryTraceStore } from "./store-DNe_Uv1Q.js";
-import { t as runEvalCampaign } from "./eval-campaign-bmZ6NIIP.js";
+import { t as runEvalCampaign } from "./eval-campaign-C4jmuM-b.js";
 import { l as assertRewardGate } from "./schema-C6DW4ZHR.js";
 import { t as isSplitEligible } from "./exporters-q9iL-2Jf.js";
 import { t as mintRolloutRows } from "./mint-BV6tLVWl.js";

package/dist/{semantic-concept-judge-BI7Rrl5-.js → semantic-concept-judge-laMCnTLn.js}

@@ -1,6 +1,6 @@
 import { i as CostLedger } from "./cost-ledger-B1qx30B4.js";
 import { l as Mutex } from "./ledger-core-DTae9rv_.js";
-import { c as callLlmJson, f as maximumChargeForLlmRequest, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-BMuxYoZy.js";
+import { c as callLlmJson, f as maximumChargeForLlmRequest, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-Bg32RW0j.js";
 import { appendFileSync, existsSync, mkdirSync, readFileSync } from "node:fs";
 import { dirname } from "node:path";
 //#region src/locked-jsonl-appender.ts
@@ -403,4 +403,4 @@ async function runSemanticConceptJudge(input, options = {}) {
 //#endregion
 export { diffFindings as a, defaultIsMaterial as i, runSemanticConceptJudge as n, FindingsStore as r, SEMANTIC_CONCEPT_JUDGE_VERSION as t };
 
-//# sourceMappingURL=semantic-concept-judge-BI7Rrl5-.js.map
+//# sourceMappingURL=semantic-concept-judge-laMCnTLn.js.map

package/dist/{semantic-concept-judge-BI7Rrl5-.js.map → semantic-concept-judge-laMCnTLn.js.map}

@@ -1 +1 @@
-{"version":3,"file":"semantic-concept-judge-BI7Rrl5-.js","names":[],"sources":["../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport { CostLedger, type CostLedgerHandle, type CostReceipt } from './cost-ledger'\nimport {\n  callLlmJson,\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  type LlmCallRequest,\n  type LlmClientOptions,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Provider-enforced output limit. Default 16000. */\n  maxTokens?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** LlmClient config (baseUrl, apiKey, authHeader, …). */\n  llm?: LlmClientOptions\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MAX_TOKENS = 16_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: Required<SemanticConceptJudgeOptions>,\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions = {},\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts: Required<SemanticConceptJudgeOptions> = {\n    model: options.model ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    llm: options.llm ?? {},\n    costLedger: options.costLedger ?? new CostLedger(),\n    costPhase: options.costPhase ?? 'judge.semantic-concept',\n    costTags: options.costTags ?? {},\n    signal: options.signal ?? new AbortController().signal,\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  let receipt: CostReceipt | undefined\n  try {\n    const request = {\n      model: opts.model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n        },\n        { role: 'user' as const, content: buildPrompt(input, opts) },\n      ],\n      jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n      temperature: 0,\n      maxTokens: opts.maxTokens,\n      timeoutMs: opts.timeoutMs,\n    } satisfies LlmCallRequest\n    const paid = await opts.costLedger.runPaidCall({\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      model: opts.model,\n      ...(Object.keys(opts.costTags).length > 0 ? { tags: opts.costTags } : {}),\n      maximumCharge: maximumChargeForLlmRequest(request, opts.llm),\n      signal: opts.signal,\n      execute: (signal, callId) =>\n        callLlmJson<{ summary: string; concepts: ConceptFinding[] }>(request, {\n          ...opts.llm,\n          signal,\n          idempotencyKey: callId,\n        }),\n      receipt: ({ result }) => costReceiptFromLlm(result),\n      receiptFromError: costReceiptFromLlmError,\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid.value\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: paid.receipt.costUnknown ? null : paid.receipt.costUsd,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: receipt && !receipt.costUnknown ? receipt.costUsd : null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions = {},\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAeA,MAAM,0BAAU,IAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;CACrC,IAAI,IAAI,QAAQ,IAAI,IAAI;CACxB,IAAI,CAAC,GAAG;EACN,IAAI,IAAI,MAAM;EACd,QAAQ,IAAI,MAAM,CAAC;CACrB;CACA,OAAO;AACT;AAEA,IAAa,sBAAb,MAAiC;CAEH;CAD5B;CACA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,QAAQ,SAAS,IAAI;EAC1B,IAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAC3B,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAEhD;CAEA,MAAM,OAAO,OAA+B;EAC1C,MAAM,OAAO,GAAG,KAAK,UAAU,KAAK,EAAE;EACtC,MAAM,KAAK,MAAM,mBAAmB;GAClC,eAAe,KAAK,MAAM,IAAI;EAChC,CAAC;CACH;AACF;;;;;;;;;;;;;;;;;;;;;;ACPA,IAAa,gBAAb,MAA2B;CAGG;CAF5B;CAEA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,WAAW,IAAI,oBAAoB,IAAI;CAC9C;CAEA,MAAM,OAAO,OAAe,UAA2C;EACrE,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,MAAwB;IAAE,GAAG;IAAG,QAAQ;GAAM;GACpD,MAAM,KAAK,SAAS,OAAO,GAAG;EAChC;CACF;;CAGA,UAA8B;EAC5B,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,MAAM,aAAa,KAAK,MAAM,MAAM;EAC1C,IAAI,CAAC,KAAK,OAAO,CAAC;EAClB,MAAM,MAA0B,CAAC;EACjC,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;GAClC,IAAI,CAAC,MAAM;GACX,IAAI;IACF,IAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;GAC/C,QAAQ,CAGR;EACF;EACA,OAAO;CACT;;CAGA,QAAQ,OAAmC;EACzC,OAAO,KAAK,QAAQ,CAAC,CAAC,QAAQ,MAAM,EAAE,WAAW,KAAK;CACxD;AACF;;;;;AAiCA,SAAgB,kBAAkB,GAAmB,GAA4B;CAC/E,IAAI,EAAE,aAAa,EAAE,UAAU,OAAO;CACtC,IAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,OAAO;CACvE,IAAI,EAAE,cAAc,WAAW,EAAE,cAAc,QAAQ,OAAO;CAC9D,OAAO;AACT;;;;;AAMA,SAAgB,aACd,UACA,SACA,SAAqB,CAAC,GACR;CACd,MAAM,aAAa,OAAO,cAAc;CACxC,MAAM,WAAW,IAAI,IAAI,SAAS,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAC/D,MAAM,UAAU,IAAI,IAAI,QAAQ,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAE7D,MAAM,WAA+B,CAAC;CACtC,MAAM,cAAkC,CAAC;CACzC,MAAM,YAAgC,CAAC;CACvC,MAAM,UAAmC,CAAC;CAE1C,KAAK,MAAM,CAAC,IAAI,QAAQ,SAAS;EAC/B,MAAM,OAAO,SAAS,IAAI,EAAE;EAC5B,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,GAAG;GACjB;EACF;EACA,IAAI,WAAW,MAAM,GAAG,GACtB,QAAQ,KAAK;GAAE,UAAU;GAAM,SAAS;EAAI,CAAC;OAE7C,UAAU,KAAK,GAAG;CAEtB;CACA,KAAK,MAAM,CAAC,IAAI,SAAS,UACvB,IAAI,CAAC,QAAQ,IAAI,EAAE,GAAG,YAAY,KAAK,IAAI;CAE7C,OAAO;EAAE;EAAU;EAAa;EAAW;CAAQ;AACrD;;;;;;;;;;;;;;;;;;;;;;;AC9BA,MAAa,6BAAgE;CAC3E,QAAQ;CACR,WAAW;CACX,SAAS;AACX;AAiCA,MAAa,iCAAiC;AAE9C,MAAM,qBAAqB;AAC3B,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,kBAAkB;AACxB,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;AAEtB,MAAM,kBAAkB;CACtB,MAAM;CACN,sBAAsB;CACtB,UAAU,CAAC,WAAW,UAAU;CAChC,YAAY;EACV,SAAS;GAAE,MAAM;GAAU,WAAW;GAAI,WAAW;EAAI;EACzD,UAAU;GACR,MAAM;GACN,UAAU;GACV,OAAO;IACL,MAAM;IACN,sBAAsB;IACtB,UAAU;KAAC;KAAW;KAAW;KAAS;KAAY;IAAU;IAChE,YAAY;KACV,SAAS;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACxD,SAAS,EAAE,MAAM,UAAU;KAC3B,OAAO;MAAE,MAAM;MAAU,SAAS;MAAG,SAAS;KAAG;KACjD,UAAU;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACzD,UAAU;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAS;OAAS;MAAM;KAAE;IAC3E;GACF;EACF;CACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;CAClE,IAAI,KAAK,UAAU,KAAK,OAAO;CAC/B,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,iBAAiB,KAAK,SAAS,IAAI,YAAY,MAAM;AACpF;AAEA,SAAS,YACP,OACA,MACQ;CACR,MAAM,aAAa,MAAM,YACtB,QAAQ,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,CAAC,CACvD,KAAK,MAAM,aAAa,EAAE,KAAK,QAAQ,EAAE,SAAS,CAAC,CACnD,KAAK,MAAM;CAEd,MAAM,OAAO,MAAM,cAAc;CAEjC,OAAO;;;;;;;;;;EAUP,MAAM,YAAY;;EAElB,MAAM,gBAAgB,+BAA+B,MAAM,cAAc,mBAAmB,MAAM,uBAAuB,GAAG,QAAQ,GAAG;EACvI,MAAM,iBACL,KACE,GAAG,MACF,KAAK,IAAI,EAAE,KAAK,EAAE,KAAK,GAAG,EAAE,UAAU,SAAS,cAAc,EAAE,SAAS,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,EAAE,KAAK,IACzG,CAAC,CACA,KAAK,IAAI,EAAE;;EAEZ,OAAO,qDAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,EAAE,QAAQ,GAAG;EACjH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,EAAE;;;;;;;;;;;;;;;;;;AAkBtD;;;;;;AASA,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACH;CACrC,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,aAAa,MAAM,iBAAiB;CAE1C,IAAI,eAAe,GACjB,OAAO;EACL,MAAM;EACN,SAAS;EACT,OAAO;EACP,cAAc;EACd,YAAY;EACZ,UAAU,CAAC;EACX,SAAS;EACT,YAAY;EACZ,SAAS;EACT,WAAW;EACX,OAAO;CACT;CAGF,MAAM,OAA8C;EAClD,OAAO,QAAQ,SAAS;EACxB,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,KAAK,QAAQ,OAAO,CAAC;EACrB,YAAY,QAAQ,cAAc,IAAI,WAAW;EACjD,WAAW,QAAQ,aAAa;EAChC,UAAU,QAAQ,YAAY,CAAC;EAC/B,QAAQ,QAAQ,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAChD,gBAAgB,QAAQ,kBAAkB;EAC1C,mBAAmB;GAAE,GAAG;GAA4B,GAAI,QAAQ,qBAAqB,CAAC;EAAG;CAC3F;CAMA,MAAM,oBAAoB,SAA8B;EACtD,IAAI,KAAK,mBAAmB,QAAQ,OAAO;EAC3C,IAAI,KAAK,UAAU,MAAM,OAAO,KAAK;EACrC,IAAI,KAAK,mBAAmB,cAC1B,OAAO,KAAK,kBAAkB,KAAK,cAAc,aAAa;EAEhE,OAAO;CACT;CACA,MAAM,eAAe,IAAI,IACvB,MAAM,iBAAiB,KAAK,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC,CACjE;CAEA,IAAI;CACJ,IAAI;EACF,MAAM,UAAU;GACd,OAAO,KAAK;GACZ,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IAAE,MAAM;IAAiB,SAAS,YAAY,OAAO,IAAI;GAAE,CAC7D;GACA,YAAY;IAAE,MAAM;IAA0B,QAAQ;GAAgB;GACtE,aAAa;GACb,WAAW,KAAK;GAChB,WAAW,KAAK;EAClB;EACA,MAAM,OAAO,MAAM,KAAK,WAAW,YAAY;GAC7C,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,OAAO,KAAK;GACZ,GAAI,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,SAAS,IAAI,EAAE,MAAM,KAAK,SAAS,IAAI,CAAC;GACvE,eAAe,2BAA2B,SAAS,KAAK,GAAG;GAC3D,QAAQ,KAAK;GACb,UAAU,QAAQ,WAChB,YAA6D,SAAS;IACpE,GAAG,KAAK;IACR;IACA,gBAAgB;GAClB,CAAC;GACH,UAAU,EAAE,aAAa,mBAAmB,MAAM;GAClD,kBAAkB;EACpB,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU,KAAK;EAEvB,IAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GACnD,MAAM,IAAI,MAAM,uEAAqE;EAGvF,MAAM,WAA6B,MAAM,SAAS,KAAK,OAAO;GAC5D,SAAS,OAAO,EAAE,OAAO;GACzB,SAAS,QAAQ,EAAE,OAAO;GAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;GACrD,UAAU,OAAO,EAAE,YAAY,EAAE;GACjC,UAAW;IAAC;IAAY;IAAS;IAAS;GAAM,CAAC,CAAW,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;EACN,EAAE;EAEF,MAAM,eAAe,SAAS,QAAQ,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;EACvE,IAAI,YAAY;EAChB,IAAI,mBAAmB;EACvB,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;GACzC,aAAa;GACb,oBAAoB,IAAI,EAAE;EAC5B;EACA,MAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;EAE7E,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO,QAAQ,WAAW,GAAA,CAAI,QAAQ,CAAC,CAAC;GACxC;GACA;GACA;GACA,SAAS,OAAO,MAAM,WAAW,EAAE;GACnC,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,KAAK,QAAQ,cAAc,OAAO,KAAK,QAAQ;GACxD,WAAW;EACb;CACF,SAAS,KAAK;EACZ,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO;GACP,cAAc;GACd;GACA,UAAU,CAAC;GACX,SAAS;GACT,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,WAAW,CAAC,QAAQ,cAAc,QAAQ,UAAU;GAC7D,WAAW;GACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;EACxD;CACF;AACF"}
+{"version":3,"file":"semantic-concept-judge-laMCnTLn.js","names":[],"sources":["../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * LockedJsonlAppender — mutex-serialized JSONL append helper for arbitrary\n * payloads. The reference-replay store does the same thing for typed\n * `ReferenceReplayRun` rows; this is the generic version used by\n * `MutationTelemetry`, `TrialTelemetry`, and any other consumer that wants\n * append-only durable telemetry without rolling its own lock.\n *\n * Locks are per absolute file path (process-local). Cross-process\n * concurrency is NOT addressed — that's an fcntl/flock problem.\n */\n\nimport { appendFileSync, existsSync, mkdirSync } from 'node:fs'\nimport { dirname } from 'node:path'\nimport { Mutex } from './concurrency'\n\nconst mutexes = new Map<string, Mutex>()\n\nfunction getMutex(path: string): Mutex {\n  let m = mutexes.get(path)\n  if (!m) {\n    m = new Mutex()\n    mutexes.set(path, m)\n  }\n  return m\n}\n\nexport class LockedJsonlAppender {\n  private readonly mutex: Mutex\n  constructor(public readonly path: string) {\n    this.mutex = getMutex(path)\n    if (!existsSync(dirname(path))) {\n      mkdirSync(dirname(path), { recursive: true })\n    }\n  }\n\n  async append(entry: unknown): Promise<void> {\n    const line = `${JSON.stringify(entry)}\\n`\n    await this.mutex.runExclusive(() => {\n      appendFileSync(this.path, line)\n    })\n  }\n}\n\n/** Reset all internal mutex state — tests only. */\nexport function resetLockedAppendersForTesting(): void {\n  mutexes.clear()\n}\n","/**\n * FindingsStore — durable persistence for AnalystFinding rows + a diff\n * helper so we can answer \"what changed since the last run?\" without\n * recomputing analysts.\n *\n * On-disk shape is JSONL: one finding per line, append-only, locked via\n * LockedJsonlAppender. Operators get crash-safety (no partial JSON),\n * cheap reads (sequential parse), and trivial backup (rsync the file).\n *\n * Reads are non-locking: a reader sees a consistent snapshot of all\n * fully-written lines and skips an incomplete trailing line if the\n * writer is mid-append. Cross-process locking is intentionally out of\n * scope (see locked-jsonl-appender.ts).\n *\n * The store is run-scoped: callers pass `runId` on append and on load,\n * which keeps multi-run files cleanly partitioned. The `diffFindings`\n * helper compares two run-id sets using stable `finding_id` semantics —\n * the diff is the cross-run signal the regression dashboard renders.\n */\n\nimport { existsSync, readFileSync } from 'node:fs'\n\nimport { LockedJsonlAppender } from '../locked-jsonl-appender'\nimport type { AnalystFinding } from './types'\n\n/**\n * One persisted row. We attach `run_id` on disk so a single file can\n * hold multiple runs and the diff helper can query without re-walking\n * separate files.\n */\nexport interface PersistedFinding extends AnalystFinding {\n  run_id: string\n}\n\nexport class FindingsStore {\n  private readonly appender: LockedJsonlAppender\n\n  constructor(public readonly path: string) {\n    this.appender = new LockedJsonlAppender(path)\n  }\n\n  async append(runId: string, findings: AnalystFinding[]): Promise<void> {\n    for (const f of findings) {\n      const row: PersistedFinding = { ...f, run_id: runId }\n      await this.appender.append(row)\n    }\n  }\n\n  /** Load every persisted finding. Discards malformed trailing lines silently. */\n  loadAll(): PersistedFinding[] {\n    if (!existsSync(this.path)) return []\n    const raw = readFileSync(this.path, 'utf8')\n    if (!raw) return []\n    const out: PersistedFinding[] = []\n    for (const line of raw.split('\\n')) {\n      if (!line) continue\n      try {\n        out.push(JSON.parse(line) as PersistedFinding)\n      } catch {\n        // Skip torn trailing line — the lock guarantees no torn lines\n        // mid-file, only at EOF when a writer is in-flight.\n      }\n    }\n    return out\n  }\n\n  /** Filter to a single run. */\n  loadRun(runId: string): PersistedFinding[] {\n    return this.loadAll().filter((r) => r.run_id === runId)\n  }\n}\n\n// ── Cross-run diff ──────────────────────────────────────────────────\n\nexport interface FindingsDiff {\n  /** New finding ids in `current` that weren't in `previous`. */\n  appeared: PersistedFinding[]\n  /** Finding ids in `previous` that aren't in `current`. */\n  disappeared: PersistedFinding[]\n  /** Same finding id present in both runs and unchanged per the materiality test. */\n  persisted: PersistedFinding[]\n  /**\n   * Same finding id in both runs but at least one non-identity field\n   * shifted per `DiffPolicy.isMaterial`. Reported as [previous, current].\n   */\n  changed: Array<{ previous: PersistedFinding; current: PersistedFinding }>\n}\n\nexport interface DiffPolicy {\n  /**\n   * Predicate that decides whether two findings (same finding_id) count\n   * as a material change. Defaults to {@link defaultIsMaterial}: severity\n   * shift, confidence Δ > 0.05, or evidence count change. Compliance /\n   * perf consumers MAY supply a stricter predicate (e.g. rationale text\n   * diff, metric Δ thresholds).\n   */\n  isMaterial?: (previous: AnalystFinding, current: AnalystFinding) => boolean\n}\n\n/**\n * Default materiality test. Deliberately narrow so LLM-reword churn\n * doesn't flood the diff. Stricter tests are opt-in via DiffPolicy.\n */\nexport function defaultIsMaterial(a: AnalystFinding, b: AnalystFinding): boolean {\n  if (a.severity !== b.severity) return true\n  if (Math.abs((a.confidence ?? 0) - (b.confidence ?? 0)) > 0.05) return true\n  if (a.evidence_refs.length !== b.evidence_refs.length) return true\n  return false\n}\n\n/**\n * Diff two findings sets by stable finding_id. Callers typically load\n * the two run-id slices from the same store and pass them in.\n */\nexport function diffFindings(\n  previous: PersistedFinding[],\n  current: PersistedFinding[],\n  policy: DiffPolicy = {},\n): FindingsDiff {\n  const isMaterial = policy.isMaterial ?? defaultIsMaterial\n  const prevById = new Map(previous.map((f) => [f.finding_id, f]))\n  const curById = new Map(current.map((f) => [f.finding_id, f]))\n\n  const appeared: PersistedFinding[] = []\n  const disappeared: PersistedFinding[] = []\n  const persisted: PersistedFinding[] = []\n  const changed: FindingsDiff['changed'] = []\n\n  for (const [id, cur] of curById) {\n    const prev = prevById.get(id)\n    if (!prev) {\n      appeared.push(cur)\n      continue\n    }\n    if (isMaterial(prev, cur)) {\n      changed.push({ previous: prev, current: cur })\n    } else {\n      persisted.push(cur)\n    }\n  }\n  for (const [id, prev] of prevById) {\n    if (!curById.has(id)) disappeared.push(prev)\n  }\n  return { appeared, disappeared, persisted, changed }\n}\n","/**\n * Semantic concept judge — \"does the built artifact actually implement\n * the features the user asked for?\"\n *\n * Distinct from the domain/code/coherence judges in `judges.ts`:\n *   - those judges score free-form conversational agent outputs along\n *     quality dimensions (accuracy, depth, etc.)\n *   - this judge scores a *built artifact* (served HTML + source files)\n *     against an explicit list of expected concepts, returning per-concept\n *     {present, score 0-10, evidence, severity}.\n *\n * The judge is strict about distinguishing (a) a working implementation\n * from (b) a keyword-present stub. \"// TODO: mint button\" is NOT present.\n * Only real, functional, wired-up code counts.\n *\n * Use via {@link createSemanticConceptJudge} or directly via\n * {@link runSemanticConceptJudge}. Soft-fails (available=false) on LLM\n * or JSON-parse errors so the caller can treat that as \"layer skipped\"\n * rather than \"layer failed\" in a multi-layer pipeline.\n */\n\nimport { CostLedger, type CostLedgerHandle, type CostReceipt } from './cost-ledger'\nimport {\n  callLlmJson,\n  costReceiptFromLlm,\n  costReceiptFromLlmError,\n  type LlmCallRequest,\n  type LlmClientOptions,\n  maximumChargeForLlmRequest,\n} from './llm-client'\nimport type { Severity } from './multi-layer-verifier'\n\n// ─── Types ──────────────────────────────────────────────────────────────\n\n/**\n * Implementation complexity class for weighted scoring.\n *\n * - `render` (default): the concept is a UI surface that displays static\n *   data — render a list, show a counter, lay out a button. Single-file\n *   work, no external integration.\n * - `integrate`: the concept requires wiring a real external system —\n *   wallet connect (wagmi + RainbowKit + chain config), payment provider\n *   (Stripe Elements + intent + webhook), an API client with auth.\n *   Multi-file, library-knowledge, runtime correctness matters.\n * - `compute`: the concept requires algorithmic work — solver, simulator,\n *   constraint propagation, ML inference. Correctness > UI polish.\n *\n * Default weights (when applied via `weightConcepts: 'complexity'`):\n *   render=1.0, integrate=2.0, compute=2.5\n *\n * Cross-vertical scoring without complexity weighting silently inflates\n * the rate of UI-heavy verticals (healthcare, fintech dashboards) vs\n * integration-heavy verticals (DeFi, wallets) — all concepts treated\n * equally even though the agent does 2-3x the work for `integrate`.\n */\nexport type ConceptComplexity = 'render' | 'integrate' | 'compute'\n\nexport interface ConceptSpec {\n  name: string\n  /** Short hints that help the judge; not used for matching. */\n  keywords?: string[]\n  /** Optional explicit weight; default 1.0. Overrides complexity-derived weight. */\n  weight?: number\n  /** Implementation complexity class. Default `render`. */\n  complexity?: ConceptComplexity\n}\n\nexport interface ConceptFinding {\n  concept: string\n  present: boolean\n  /** 0..10. 10 = production-ready; 7 = functional thin; 4 = partial; 0 = absent. */\n  score: number\n  evidence: string\n  severity: Severity\n}\n\nexport interface SemanticConceptJudgeInput {\n  /** Full natural-language prompt the agent was handed. */\n  userRequest: string\n  /** Rendered HTML the preview returns (UI artifacts). Optional. */\n  servedHtml?: string\n  /** Top-level source files from the agent's workdir. */\n  sourceFiles: Array<{ path: string; content: string }>\n  /** The expected concept list. */\n  expectedConcepts: ConceptSpec[]\n  /** Free-form metadata (id, difficulty) to inject into the prompt. */\n  artifactLabel?: string\n  artifactDescription?: string\n}\n\nexport interface SemanticConceptJudgeResult {\n  kind: 'semantic-concept'\n  version: string\n  /** Normalized 0..1 score — mean of per-concept scores / 10. */\n  score: number\n  presentCount: number\n  totalCount: number\n  findings: ConceptFinding[]\n  summary: string\n  durationMs: number\n  costUsd: number | null\n  /** False on LLM/JSON error — treat as \"skipped / unable to judge\" in pipelines. */\n  available: boolean\n  error?: string\n}\n\n/**\n * Score-aggregation strategy. `mean` averages 0-10 scores uniformly.\n * `complexity` applies the default weight table (render=1, integrate=2,\n * compute=2.5) unless a concept has an explicit `weight`. `explicit`\n * honors only `weight` (defaulting to 1 for unspecified).\n */\nexport type ConceptWeightStrategy = 'mean' | 'complexity' | 'explicit'\n\nexport const DEFAULT_COMPLEXITY_WEIGHTS: Record<ConceptComplexity, number> = {\n  render: 1.0,\n  integrate: 2.0,\n  compute: 2.5,\n}\n\nexport interface SemanticConceptJudgeOptions {\n  /** Model id to call. Default 'claude-sonnet-4-6' via agent-eval defaults. */\n  model?: string\n  /** Per-call timeout. Default 300s. */\n  timeoutMs?: number\n  /** Provider-enforced output limit. Default 16000. */\n  maxTokens?: number\n  /** Pipeline budget for the prompt (source blob truncation). Default 45000. */\n  maxSourceChars?: number\n  /** Per-file cap before inclusion. Default 20000. */\n  maxPerFileChars?: number\n  /** HTML cap. Default 30000. */\n  maxHtmlChars?: number\n  /** LlmClient config (baseUrl, apiKey, authHeader, …). */\n  llm?: LlmClientOptions\n  costLedger?: CostLedgerHandle\n  costPhase?: string\n  costTags?: Record<string, string>\n  signal?: AbortSignal\n  /**\n   * Score aggregation strategy. Default `mean` — uniform average across\n   * concepts. Cross-vertical comparisons should use `complexity` to\n   * neutralize the integrate-vs-render asymmetry.\n   */\n  weightConcepts?: ConceptWeightStrategy\n  /** Override the default complexity → weight table. */\n  complexityWeights?: Partial<Record<ConceptComplexity, number>>\n}\n\n// ─── Prompt assembly ────────────────────────────────────────────────────\n\nexport const SEMANTIC_CONCEPT_JUDGE_VERSION = 'semantic-concept-judge-v1-2026-04-24'\n\nconst DEFAULT_MAX_SOURCE = 45_000\nconst DEFAULT_MAX_HTML = 30_000\nconst DEFAULT_MAX_PER_FILE = 20_000\nconst DEFAULT_TIMEOUT = 300_000\nconst DEFAULT_MAX_TOKENS = 16_000\nconst DEFAULT_MODEL = 'claude-sonnet-4-6'\n\nconst SEMANTIC_SCHEMA = {\n  type: 'object',\n  additionalProperties: false,\n  required: ['summary', 'concepts'],\n  properties: {\n    summary: { type: 'string', minLength: 20, maxLength: 600 },\n    concepts: {\n      type: 'array',\n      minItems: 1,\n      items: {\n        type: 'object',\n        additionalProperties: false,\n        required: ['concept', 'present', 'score', 'evidence', 'severity'],\n        properties: {\n          concept: { type: 'string', minLength: 1, maxLength: 120 },\n          present: { type: 'boolean' },\n          score: { type: 'number', minimum: 0, maximum: 10 },\n          evidence: { type: 'string', minLength: 5, maxLength: 400 },\n          severity: { type: 'string', enum: ['critical', 'major', 'minor', 'info'] },\n        },\n      },\n    },\n  },\n}\n\nfunction truncate(body: string, cap: number, label: string): string {\n  if (body.length <= cap) return body\n  return `${body.slice(0, cap)}\\n… [truncated ${body.length - cap} chars of ${label}]`\n}\n\nfunction buildPrompt(\n  input: SemanticConceptJudgeInput,\n  opts: Required<SemanticConceptJudgeOptions>,\n): string {\n  const sourceBlob = input.sourceFiles\n    .filter((f) => f.content.length <= opts.maxPerFileChars)\n    .map((f) => `--- FILE: ${f.path} ---\\n${f.content}`)\n    .join('\\n\\n')\n\n  const html = input.servedHtml ?? ''\n\n  return `You are a strict code-review judge evaluating whether an agent's 0-to-1 build actually implements the features the user asked for.\n\nYou MUST distinguish:\n  (a) WORKING code that implements the concept (rendered UI, wired handler, real API call),\n  (b) KEYWORD-PRESENT stub (comments mentioning the concept, variable names, TODOs),\n  (c) ABSENT (concept nowhere).\n\nA comment like \"// TODO: add mint button\" is NOT present — score 2-3. Only count a concept as present if there is real functional code: a rendered component, a call handler wired to state or a network call, a computed value actually used.\n\nUSER REQUEST (what the agent was asked to build):\n${input.userRequest}\n\n${input.artifactLabel ? `ARTIFACT METADATA:\\n  name: ${input.artifactLabel}\\n  description: ${input.artifactDescription ?? ''}\\n\\n` : ''}EXPECTED CONCEPTS (each must be graded independently):\n${input.expectedConcepts\n  .map(\n    (c, i) =>\n      `  ${i + 1}. \"${c.name}\"${c.keywords?.length ? ` — hints: [${c.keywords.slice(0, 6).join(' | ')}]` : ''}`,\n  )\n  .join('\\n')}\n\n${html ? `SERVED HTML (what the preview returns when hit):\\n${truncate(html, opts.maxHtmlChars, 'HTML')}\\n\\n` : ''}SOURCE FILES (the agent's workdir):\n${truncate(sourceBlob, opts.maxSourceChars, 'source')}\n\nFor EACH concept, return:\n  - concept: the concept name as given (match exactly)\n  - present: boolean — does a working implementation exist?\n  - score: 0-10 — 10 = production-ready; 7 = functional but thin; 4 = partial/stubbed; 2 = keyword-only comment; 0 = absent\n  - evidence: cite \"<file>:<line>\" or \"served-html:<selector>\" pointing at the strongest supporting code. If the concept is absent or stubbed, explain what's missing.\n  - severity:\n      \"info\" when present: true AND score >= 7\n      \"minor\" when present: true AND 4 <= score < 7\n      \"major\" when present: false OR score < 4\n      \"critical\" when the concept is not only absent but a core user flow depends on it\n\nAlso produce a \"summary\" (one sentence, 20-600 chars): overall verdict on whether this is a shippable implementation of the user request vs a keyword-dense placeholder.\n\nBE SKEPTICAL. Keyword matching already passed — your job is to catch what keyword matching misses. If the agent shipped a working build, say so. If it shipped a stub, say so. Don't grade on effort.\n\nReturn STRICT JSON. No prose outside the JSON.`\n}\n\n// ─── Runner ─────────────────────────────────────────────────────────────\n\n/**\n * Run the semantic concept judge. Soft-fails to available=false on\n * LLM/JSON errors — callers in a MultiLayerVerifier pipeline can treat\n * that as \"skip\" rather than \"fail.\"\n */\nexport async function runSemanticConceptJudge(\n  input: SemanticConceptJudgeInput,\n  options: SemanticConceptJudgeOptions = {},\n): Promise<SemanticConceptJudgeResult> {\n  const start = Date.now()\n  const totalCount = input.expectedConcepts.length\n\n  if (totalCount === 0) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount: 0,\n      findings: [],\n      summary: 'no expected concepts declared',\n      durationMs: 0,\n      costUsd: null,\n      available: false,\n      error: 'no expected concepts declared',\n    }\n  }\n\n  const opts: Required<SemanticConceptJudgeOptions> = {\n    model: options.model ?? DEFAULT_MODEL,\n    timeoutMs: options.timeoutMs ?? DEFAULT_TIMEOUT,\n    maxTokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,\n    maxSourceChars: options.maxSourceChars ?? DEFAULT_MAX_SOURCE,\n    maxPerFileChars: options.maxPerFileChars ?? DEFAULT_MAX_PER_FILE,\n    maxHtmlChars: options.maxHtmlChars ?? DEFAULT_MAX_HTML,\n    llm: options.llm ?? {},\n    costLedger: options.costLedger ?? new CostLedger(),\n    costPhase: options.costPhase ?? 'judge.semantic-concept',\n    costTags: options.costTags ?? {},\n    signal: options.signal ?? new AbortController().signal,\n    weightConcepts: options.weightConcepts ?? 'mean',\n    complexityWeights: { ...DEFAULT_COMPLEXITY_WEIGHTS, ...(options.complexityWeights ?? {}) },\n  }\n\n  // Build a name → weight map for aggregation. Mean strategy keeps every\n  // weight at 1 (uniform average). Complexity strategy reads the table\n  // and lets an explicit `weight` override. Explicit strategy uses ONLY\n  // the spec's `weight` (defaulting to 1).\n  const weightForConcept = (spec: ConceptSpec): number => {\n    if (opts.weightConcepts === 'mean') return 1\n    if (spec.weight != null) return spec.weight\n    if (opts.weightConcepts === 'complexity') {\n      return opts.complexityWeights[spec.complexity ?? 'render'] ?? 1\n    }\n    return 1\n  }\n  const weightByName = new Map<string, number>(\n    input.expectedConcepts.map((c) => [c.name, weightForConcept(c)]),\n  )\n\n  let receipt: CostReceipt | undefined\n  try {\n    const request = {\n      model: opts.model,\n      messages: [\n        {\n          role: 'system' as const,\n          content:\n            'You are a strict code-review judge. Return strict JSON only. No prose outside the JSON. A keyword in a comment is NOT a working implementation.',\n        },\n        { role: 'user' as const, content: buildPrompt(input, opts) },\n      ],\n      jsonSchema: { name: 'semantic_concept_judge', schema: SEMANTIC_SCHEMA },\n      temperature: 0,\n      maxTokens: opts.maxTokens,\n      timeoutMs: opts.timeoutMs,\n    } satisfies LlmCallRequest\n    const paid = await opts.costLedger.runPaidCall({\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      model: opts.model,\n      ...(Object.keys(opts.costTags).length > 0 ? { tags: opts.costTags } : {}),\n      maximumCharge: maximumChargeForLlmRequest(request, opts.llm),\n      signal: opts.signal,\n      execute: (signal, callId) =>\n        callLlmJson<{ summary: string; concepts: ConceptFinding[] }>(request, {\n          ...opts.llm,\n          signal,\n          idempotencyKey: callId,\n        }),\n      receipt: ({ result }) => costReceiptFromLlm(result),\n      receiptFromError: costReceiptFromLlmError,\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid.value\n\n    if (!value?.concepts || !Array.isArray(value.concepts)) {\n      throw new Error('judge returned malformed response — expected array under \"concepts\"')\n    }\n\n    const findings: ConceptFinding[] = value.concepts.map((c) => ({\n      concept: String(c.concept),\n      present: Boolean(c.present),\n      score: Math.max(0, Math.min(10, Number(c.score ?? 0))),\n      evidence: String(c.evidence ?? ''),\n      severity: (['critical', 'major', 'minor', 'info'] as const).includes(c.severity)\n        ? c.severity\n        : 'info',\n    }))\n\n    const presentCount = findings.filter((f) => f.present && f.score >= 7).length\n    let weightSum = 0\n    let weightedScoreSum = 0\n    for (const f of findings) {\n      const w = weightByName.get(f.concept) ?? 1\n      weightSum += w\n      weightedScoreSum += w * f.score\n    }\n    const scoreAvg =\n      weightSum > 0\n        ? weightedScoreSum / weightSum\n        : findings.reduce((a, f) => a + f.score, 0) / Math.max(1, findings.length)\n\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: Number((scoreAvg / 10).toFixed(3)),\n      presentCount,\n      totalCount,\n      findings,\n      summary: String(value.summary ?? ''),\n      durationMs: Date.now() - start,\n      costUsd: paid.receipt.costUnknown ? null : paid.receipt.costUsd,\n      available: true,\n    }\n  } catch (err) {\n    return {\n      kind: 'semantic-concept',\n      version: SEMANTIC_CONCEPT_JUDGE_VERSION,\n      score: 0,\n      presentCount: 0,\n      totalCount,\n      findings: [],\n      summary: '',\n      durationMs: Date.now() - start,\n      costUsd: receipt && !receipt.costUnknown ? receipt.costUsd : null,\n      available: false,\n      error: err instanceof Error ? err.message : String(err),\n    }\n  }\n}\n\n/**\n * Factory: pin LLM options once, return a closure that accepts inputs.\n * Convenient for pipelines that want to share a single LlmClient config.\n */\nexport function createSemanticConceptJudge(\n  options: SemanticConceptJudgeOptions = {},\n): (input: SemanticConceptJudgeInput) => Promise<SemanticConceptJudgeResult> {\n  return (input) => runSemanticConceptJudge(input, options)\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAeA,MAAM,0BAAU,IAAI,IAAmB;AAEvC,SAAS,SAAS,MAAqB;CACrC,IAAI,IAAI,QAAQ,IAAI,IAAI;CACxB,IAAI,CAAC,GAAG;EACN,IAAI,IAAI,MAAM;EACd,QAAQ,IAAI,MAAM,CAAC;CACrB;CACA,OAAO;AACT;AAEA,IAAa,sBAAb,MAAiC;CAEH;CAD5B;CACA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,QAAQ,SAAS,IAAI;EAC1B,IAAI,CAAC,WAAW,QAAQ,IAAI,CAAC,GAC3B,UAAU,QAAQ,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;CAEhD;CAEA,MAAM,OAAO,OAA+B;EAC1C,MAAM,OAAO,GAAG,KAAK,UAAU,KAAK,EAAE;EACtC,MAAM,KAAK,MAAM,mBAAmB;GAClC,eAAe,KAAK,MAAM,IAAI;EAChC,CAAC;CACH;AACF;;;;;;;;;;;;;;;;;;;;;;ACPA,IAAa,gBAAb,MAA2B;CAGG;CAF5B;CAEA,YAAY,MAA8B;EAAd,KAAA,OAAA;EAC1B,KAAK,WAAW,IAAI,oBAAoB,IAAI;CAC9C;CAEA,MAAM,OAAO,OAAe,UAA2C;EACrE,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,MAAwB;IAAE,GAAG;IAAG,QAAQ;GAAM;GACpD,MAAM,KAAK,SAAS,OAAO,GAAG;EAChC;CACF;;CAGA,UAA8B;EAC5B,IAAI,CAAC,WAAW,KAAK,IAAI,GAAG,OAAO,CAAC;EACpC,MAAM,MAAM,aAAa,KAAK,MAAM,MAAM;EAC1C,IAAI,CAAC,KAAK,OAAO,CAAC;EAClB,MAAM,MAA0B,CAAC;EACjC,KAAK,MAAM,QAAQ,IAAI,MAAM,IAAI,GAAG;GAClC,IAAI,CAAC,MAAM;GACX,IAAI;IACF,IAAI,KAAK,KAAK,MAAM,IAAI,CAAqB;GAC/C,QAAQ,CAGR;EACF;EACA,OAAO;CACT;;CAGA,QAAQ,OAAmC;EACzC,OAAO,KAAK,QAAQ,CAAC,CAAC,QAAQ,MAAM,EAAE,WAAW,KAAK;CACxD;AACF;;;;;AAiCA,SAAgB,kBAAkB,GAAmB,GAA4B;CAC/E,IAAI,EAAE,aAAa,EAAE,UAAU,OAAO;CACtC,IAAI,KAAK,KAAK,EAAE,cAAc,MAAM,EAAE,cAAc,EAAE,IAAI,KAAM,OAAO;CACvE,IAAI,EAAE,cAAc,WAAW,EAAE,cAAc,QAAQ,OAAO;CAC9D,OAAO;AACT;;;;;AAMA,SAAgB,aACd,UACA,SACA,SAAqB,CAAC,GACR;CACd,MAAM,aAAa,OAAO,cAAc;CACxC,MAAM,WAAW,IAAI,IAAI,SAAS,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAC/D,MAAM,UAAU,IAAI,IAAI,QAAQ,KAAK,MAAM,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC;CAE7D,MAAM,WAA+B,CAAC;CACtC,MAAM,cAAkC,CAAC;CACzC,MAAM,YAAgC,CAAC;CACvC,MAAM,UAAmC,CAAC;CAE1C,KAAK,MAAM,CAAC,IAAI,QAAQ,SAAS;EAC/B,MAAM,OAAO,SAAS,IAAI,EAAE;EAC5B,IAAI,CAAC,MAAM;GACT,SAAS,KAAK,GAAG;GACjB;EACF;EACA,IAAI,WAAW,MAAM,GAAG,GACtB,QAAQ,KAAK;GAAE,UAAU;GAAM,SAAS;EAAI,CAAC;OAE7C,UAAU,KAAK,GAAG;CAEtB;CACA,KAAK,MAAM,CAAC,IAAI,SAAS,UACvB,IAAI,CAAC,QAAQ,IAAI,EAAE,GAAG,YAAY,KAAK,IAAI;CAE7C,OAAO;EAAE;EAAU;EAAa;EAAW;CAAQ;AACrD;;;;;;;;;;;;;;;;;;;;;;;AC9BA,MAAa,6BAAgE;CAC3E,QAAQ;CACR,WAAW;CACX,SAAS;AACX;AAiCA,MAAa,iCAAiC;AAE9C,MAAM,qBAAqB;AAC3B,MAAM,mBAAmB;AACzB,MAAM,uBAAuB;AAC7B,MAAM,kBAAkB;AACxB,MAAM,qBAAqB;AAC3B,MAAM,gBAAgB;AAEtB,MAAM,kBAAkB;CACtB,MAAM;CACN,sBAAsB;CACtB,UAAU,CAAC,WAAW,UAAU;CAChC,YAAY;EACV,SAAS;GAAE,MAAM;GAAU,WAAW;GAAI,WAAW;EAAI;EACzD,UAAU;GACR,MAAM;GACN,UAAU;GACV,OAAO;IACL,MAAM;IACN,sBAAsB;IACtB,UAAU;KAAC;KAAW;KAAW;KAAS;KAAY;IAAU;IAChE,YAAY;KACV,SAAS;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACxD,SAAS,EAAE,MAAM,UAAU;KAC3B,OAAO;MAAE,MAAM;MAAU,SAAS;MAAG,SAAS;KAAG;KACjD,UAAU;MAAE,MAAM;MAAU,WAAW;MAAG,WAAW;KAAI;KACzD,UAAU;MAAE,MAAM;MAAU,MAAM;OAAC;OAAY;OAAS;OAAS;MAAM;KAAE;IAC3E;GACF;EACF;CACF;AACF;AAEA,SAAS,SAAS,MAAc,KAAa,OAAuB;CAClE,IAAI,KAAK,UAAU,KAAK,OAAO;CAC/B,OAAO,GAAG,KAAK,MAAM,GAAG,GAAG,EAAE,iBAAiB,KAAK,SAAS,IAAI,YAAY,MAAM;AACpF;AAEA,SAAS,YACP,OACA,MACQ;CACR,MAAM,aAAa,MAAM,YACtB,QAAQ,MAAM,EAAE,QAAQ,UAAU,KAAK,eAAe,CAAC,CACvD,KAAK,MAAM,aAAa,EAAE,KAAK,QAAQ,EAAE,SAAS,CAAC,CACnD,KAAK,MAAM;CAEd,MAAM,OAAO,MAAM,cAAc;CAEjC,OAAO;;;;;;;;;;EAUP,MAAM,YAAY;;EAElB,MAAM,gBAAgB,+BAA+B,MAAM,cAAc,mBAAmB,MAAM,uBAAuB,GAAG,QAAQ,GAAG;EACvI,MAAM,iBACL,KACE,GAAG,MACF,KAAK,IAAI,EAAE,KAAK,EAAE,KAAK,GAAG,EAAE,UAAU,SAAS,cAAc,EAAE,SAAS,MAAM,GAAG,CAAC,CAAC,CAAC,KAAK,KAAK,EAAE,KAAK,IACzG,CAAC,CACA,KAAK,IAAI,EAAE;;EAEZ,OAAO,qDAAqD,SAAS,MAAM,KAAK,cAAc,MAAM,EAAE,QAAQ,GAAG;EACjH,SAAS,YAAY,KAAK,gBAAgB,QAAQ,EAAE;;;;;;;;;;;;;;;;;;AAkBtD;;;;;;AASA,eAAsB,wBACpB,OACA,UAAuC,CAAC,GACH;CACrC,MAAM,QAAQ,KAAK,IAAI;CACvB,MAAM,aAAa,MAAM,iBAAiB;CAE1C,IAAI,eAAe,GACjB,OAAO;EACL,MAAM;EACN,SAAS;EACT,OAAO;EACP,cAAc;EACd,YAAY;EACZ,UAAU,CAAC;EACX,SAAS;EACT,YAAY;EACZ,SAAS;EACT,WAAW;EACX,OAAO;CACT;CAGF,MAAM,OAA8C;EAClD,OAAO,QAAQ,SAAS;EACxB,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,KAAK,QAAQ,OAAO,CAAC;EACrB,YAAY,QAAQ,cAAc,IAAI,WAAW;EACjD,WAAW,QAAQ,aAAa;EAChC,UAAU,QAAQ,YAAY,CAAC;EAC/B,QAAQ,QAAQ,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAChD,gBAAgB,QAAQ,kBAAkB;EAC1C,mBAAmB;GAAE,GAAG;GAA4B,GAAI,QAAQ,qBAAqB,CAAC;EAAG;CAC3F;CAMA,MAAM,oBAAoB,SAA8B;EACtD,IAAI,KAAK,mBAAmB,QAAQ,OAAO;EAC3C,IAAI,KAAK,UAAU,MAAM,OAAO,KAAK;EACrC,IAAI,KAAK,mBAAmB,cAC1B,OAAO,KAAK,kBAAkB,KAAK,cAAc,aAAa;EAEhE,OAAO;CACT;CACA,MAAM,eAAe,IAAI,IACvB,MAAM,iBAAiB,KAAK,MAAM,CAAC,EAAE,MAAM,iBAAiB,CAAC,CAAC,CAAC,CACjE;CAEA,IAAI;CACJ,IAAI;EACF,MAAM,UAAU;GACd,OAAO,KAAK;GACZ,UAAU,CACR;IACE,MAAM;IACN,SACE;GACJ,GACA;IAAE,MAAM;IAAiB,SAAS,YAAY,OAAO,IAAI;GAAE,CAC7D;GACA,YAAY;IAAE,MAAM;IAA0B,QAAQ;GAAgB;GACtE,aAAa;GACb,WAAW,KAAK;GAChB,WAAW,KAAK;EAClB;EACA,MAAM,OAAO,MAAM,KAAK,WAAW,YAAY;GAC7C,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,OAAO,KAAK;GACZ,GAAI,OAAO,KAAK,KAAK,QAAQ,CAAC,CAAC,SAAS,IAAI,EAAE,MAAM,KAAK,SAAS,IAAI,CAAC;GACvE,eAAe,2BAA2B,SAAS,KAAK,GAAG;GAC3D,QAAQ,KAAK;GACb,UAAU,QAAQ,WAChB,YAA6D,SAAS;IACpE,GAAG,KAAK;IACR;IACA,gBAAgB;GAClB,CAAC;GACH,UAAU,EAAE,aAAa,mBAAmB,MAAM;GAClD,kBAAkB;EACpB,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU,KAAK;EAEvB,IAAI,CAAC,OAAO,YAAY,CAAC,MAAM,QAAQ,MAAM,QAAQ,GACnD,MAAM,IAAI,MAAM,uEAAqE;EAGvF,MAAM,WAA6B,MAAM,SAAS,KAAK,OAAO;GAC5D,SAAS,OAAO,EAAE,OAAO;GACzB,SAAS,QAAQ,EAAE,OAAO;GAC1B,OAAO,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,OAAO,EAAE,SAAS,CAAC,CAAC,CAAC;GACrD,UAAU,OAAO,EAAE,YAAY,EAAE;GACjC,UAAW;IAAC;IAAY;IAAS;IAAS;GAAM,CAAC,CAAW,SAAS,EAAE,QAAQ,IAC3E,EAAE,WACF;EACN,EAAE;EAEF,MAAM,eAAe,SAAS,QAAQ,MAAM,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC,CAAC;EACvE,IAAI,YAAY;EAChB,IAAI,mBAAmB;EACvB,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,IAAI,aAAa,IAAI,EAAE,OAAO,KAAK;GACzC,aAAa;GACb,oBAAoB,IAAI,EAAE;EAC5B;EACA,MAAM,WACJ,YAAY,IACR,mBAAmB,YACnB,SAAS,QAAQ,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,KAAK,IAAI,GAAG,SAAS,MAAM;EAE7E,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO,QAAQ,WAAW,GAAA,CAAI,QAAQ,CAAC,CAAC;GACxC;GACA;GACA;GACA,SAAS,OAAO,MAAM,WAAW,EAAE;GACnC,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,KAAK,QAAQ,cAAc,OAAO,KAAK,QAAQ;GACxD,WAAW;EACb;CACF,SAAS,KAAK;EACZ,OAAO;GACL,MAAM;GACN,SAAS;GACT,OAAO;GACP,cAAc;GACd;GACA,UAAU,CAAC;GACX,SAAS;GACT,YAAY,KAAK,IAAI,IAAI;GACzB,SAAS,WAAW,CAAC,QAAQ,cAAc,QAAQ,UAAU;GAC7D,WAAW;GACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;EACxD;CACF;AACF"}

package/dist/{series-convergence-D2fsoJ2w.d.ts → series-convergence-BxKEgBwA.d.ts}

@@ -1,4 +1,4 @@
-import { s as JudgeScore } from "./types-DOhGq4S1.js";
+import { s as JudgeScore } from "./types-jUBXJ7Iz.js";
 import { i as ContinuousAgreementOptions, r as ContinuousAgreement } from "./judge-calibration-C5CbMYce.js";
 //#region src/statistics/agreement-irr.d.ts
 /**
@@ -123,4 +123,4 @@ interface SeriesConvergenceResult {
 declare function analyzeSeries(values: number[], options?: SeriesConvergenceOptions): SeriesConvergenceResult;
 //#endregion
 export { CorpusAgreementPerDimension as a, corpusInterRaterAgreement as c, CorpusAgreementOptions as i, corpusInterRaterAgreementFromJudgeScores as l, SeriesConvergenceResult as n, CorpusAgreementReport as o, analyzeSeries as r, CorpusScoreRecord as s, SeriesConvergenceOptions as t, interRaterReliability as u };
-//# sourceMappingURL=series-convergence-D2fsoJ2w.d.ts.map
+//# sourceMappingURL=series-convergence-BxKEgBwA.d.ts.map

package/dist/{series-convergence-D2fsoJ2w.d.ts.map → series-convergence-BxKEgBwA.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"series-convergence-D2fsoJ2w.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../src/series-convergence.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;;;;;;;;;;;;UCvRc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR"}
+{"version":3,"file":"series-convergence-BxKEgBwA.d.ts","names":[],"sources":["../src/statistics/agreement-irr.ts","../src/series-convergence.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;iBAyBgB,sBAAsB,aAAa;UA0ElC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe,oCAAoC;EACnD;;EAEA;;EAEA;;UAGe;;EAEf,cAAc;;EAEd;;EAEA;;EAEA;;EAEA;;UAGe,+BAA+B;;;;;;EAM9C;;;;;;EAMA;;;;;;;;;;;;;;;;;;iBAmBc,0BACd,SAAS,qBACT,OAAM,yBACL;;;;;;;;;iBA0Ha,yCACd,aAAa;EAAQ;EAAgB,QAAQ;IAC7C,OAAM,yBACL;;;;;;;;;;;;;UCvRc;;EAEf;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;iBAGc,cACd,kBACA,UAAS,2BACR"}

package/dist/{server-D9wQclzG.js → server-dIWwF3j_.js}

@@ -1,6 +1,6 @@
 import { i as CostLedger } from "./cost-ledger-B1qx30B4.js";
 import { t as packageVersion } from "./package-version-D7lQHt_-.js";
-import { a as assertLlmRoute, c as callLlmJson, f as maximumChargeForLlmRequest, i as LlmRouteAssertionError, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-BMuxYoZy.js";
+import { a as assertLlmRoute, c as callLlmJson, f as maximumChargeForLlmRequest, i as LlmRouteAssertionError, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-Bg32RW0j.js";
 import { z } from "zod";
 import { OpenAPIRegistry, OpenApiGeneratorV31, extendZodWithOpenApi } from "@asteasolutions/zod-to-openapi";
 import { serve } from "@hono/node-server";
@@ -1021,4 +1021,4 @@ function startServerAsync(opts = {}) {
 //#endregion
 export { TraceEventSchema as A, HealthResponseSchema as C, RubricDimensionSchema as D, ListRubricsResponseSchema as E, hashRubric as F, TracesIngestResponseSchema as M, VersionResponseSchema as N, RubricInfoSchema as O, WIRE_VERSION as P, FeedbackTrajectorySchema as S, JudgeResultSchema as T, ErrorResponseSchema as _, runRpcBatch as a, FeedbackIngestResponseSchema as b, WireError as c, handleListRubrics as d, handleTracesIngest as f, listBuiltinRubrics as g, getBuiltinRubric as h, dispatchRpc as i, TracesIngestRequestSchema as j, RubricSchema as k, handleFeedbackIngest as l, BUILTIN_RUBRICS as m, startServer as n, runRpcOnce as o, handleVersion as p, startServerAsync as r, buildOpenApi as s, createApp as t, handleJudge as u, FailureModeSchema as v, JudgeRequestSchema as w, FeedbackLabelSchema as x, FeedbackAttemptSchema as y };
 
-//# sourceMappingURL=server-D9wQclzG.js.map
+//# sourceMappingURL=server-dIWwF3j_.js.map