@tangle-network/agent-eval 0.203.2 → 0.205.0

package/dist/{search-summary-C1nAHgkB.js → search-summary-BZkoLwbB.js}

@@ -1,7 +1,7 @@
-import "./paired-tests-Ccr16gcW.js";
-import { r as minimumPairsForPairedDeltaTest } from "./paired-promotion-decision-DhrmjdV8.js";
+import "./paired-tests-jbwPdSz5.js";
+import { r as minimumPairsForPairedDeltaTest } from "./paired-promotion-decision-CO0kA0bq.js";
 import { a as SearchLedgerError } from "./search-ledger-file-DY0AXX3W.js";
-import { b as estimateNode } from "./search-kernel-Dpniloks.js";
+import { b as estimateNode } from "./search-kernel-CsISc5t8.js";
 //#region src/campaign/search-summary.ts
 /**
 * The agent's view of a search: the same projection a human reads.
@@ -130,4 +130,4 @@ function formatEstimate(estimate) {
 //#endregion
 export { searchProposerView as n, renderSearchSummary as t };
 
-//# sourceMappingURL=search-summary-C1nAHgkB.js.map
+//# sourceMappingURL=search-summary-BZkoLwbB.js.map

package/dist/{search-summary-C1nAHgkB.js.map → search-summary-BZkoLwbB.js.map}

@@ -1 +1 @@
-{"version":3,"file":"search-summary-C1nAHgkB.js","names":[],"sources":["../src/campaign/search-summary.ts"],"sourcesContent":["/**\n * The agent's view of a search: the same projection a human reads.\n *\n * `renderSearchSummary` turns a `SearchStateView` into compact text — the\n * leading nodes with their estimates, the most recently discarded ideas with\n * why they were discarded, and a log of recent proposals — as AIDE's journal\n * summary does for its own search. `agent-eval search show <ledger>` prints\n * the same text for a terminal; `runOptimization` puts it on every\n * `ProposeContext` a proposer receives.\n *\n * `searchProposerView` is the train-only read `ProposeContext.train` carries.\n * It has no parameter that can select another split, so a proposer built on\n * it cannot reach the selection or test split even by mistake: the sealed\n * splits (§6.6 of the search-tree design) are unreachable by construction,\n * not by convention.\n */\n\nimport { minimumPairsForPairedDeltaTest } from '../paired-delta-test'\nimport { BOOTSTRAP_GATE_MIN_N } from '../statistics'\nimport { estimateNode } from './estimate-node'\nimport { SearchLedgerError } from './search-ledger-errors'\nimport type { NodeEstimate, SearchSplit } from './search-ledger-types'\nimport type { SearchNode, SearchScoredCell, SearchStateView, SearchUnitScore } from './search-state'\n\nconst DESCRIPTIVE_FROM = minimumPairsForPairedDeltaTest(0.95)\n\n/** The proposer's own read of a search: its node's cells on the train split,\n * the only split a proposer may ever see. */\nexport interface SearchProposerView {\n  readonly searchId: string\n  readonly direction: 'maximize' | 'minimize'\n  /** `nodeId`'s scored train cells. Unscored cells (errored, cancelled or in\n   * flight) are absent, never zero. */\n  scoredCells(nodeId: string): readonly SearchScoredCell[]\n  /** Per-unit train-split means of `nodeId`, in unitId order. */\n  unitScores(nodeId: string): readonly SearchUnitScore[]\n}\n\n/** The train-only view `ProposeContext.train` carries. */\nexport function searchProposerView(state: SearchStateView): SearchProposerView {\n  const header = state.header\n  if (!header) throw new SearchLedgerError(`search ${state.searchId} has not been opened`)\n  return {\n    searchId: state.searchId,\n    direction: header.objective.direction,\n    scoredCells: (nodeId) => state.scoredCells(nodeId, 'train'),\n    unitScores: (nodeId) => state.unitScores(nodeId, 'train'),\n  }\n}\n\nexport interface SearchSummaryOptions {\n  /**\n   * The split every estimate in the summary is computed on. A proposer's\n   * context must always pass `'train'`: the selection and test splits stay\n   * unread by anything a proposer's text can reach. The CLI passes the\n   * search's own ranking split, `'selection'` when the search declares one\n   * else `'train'`, matching `searchPolicyView`.\n   */\n  split: SearchSplit\n  /** Nodes shown under \"Leading\" and under \"Recently discarded\". Default 5. */\n  limit?: number\n}\n\n/**\n * `SearchState` as compact text: a header line with the search's shape and\n * spend, the leading nodes against the root with their estimates, the most\n * recently discarded nodes with the measurement that discarded them, and a\n * log of recent proposals (operator and label — the ledger's inline,\n * redacted summary of each rationale; the full rationale is a blob this\n * function never reads, so it stays a pure, I/O-free projection of the\n * state already in hand).\n */\nexport function renderSearchSummary(state: SearchStateView, options: SearchSummaryOptions): string {\n  const header = state.header\n  if (!header) return `search ${state.searchId}: not yet opened`\n  const limit = options.limit ?? 5\n  if (limit < 1 || !Number.isInteger(limit)) {\n    throw new SearchLedgerError(\n      `renderSearchSummary: limit must be a positive integer, got ${limit}`,\n    )\n  }\n  const split = options.split\n  const direction = header.objective.direction\n  const root = state.rootNodeId\n  const lines: string[] = [headerLine(state), spendLine(state), statusLine(state)]\n\n  if (root === null) {\n    lines.push('', 'No nodes yet.')\n    return lines.join('\\n')\n  }\n\n  const others = state.nodes().filter((node) => node.nodeId !== root)\n  const estimated = others\n    .filter((node) => node.status !== 'invalid')\n    .map((node) => ({ node, estimate: estimateNode(state, node.nodeId, { against: root, split }) }))\n\n  const leading = estimated\n    .filter(({ estimate }) => estimate.delta !== null)\n    .sort((a, b) => signed(b.estimate, direction) - signed(a.estimate, direction))\n    .slice(0, limit)\n  lines.push('', `Leading nodes (vs root, ${split} split):`)\n  if (leading.length === 0) lines.push('  none measured yet')\n  for (const { node, estimate } of leading)\n    lines.push(`  ${nodeLine(node)}: ${formatEstimate(estimate)}`)\n\n  const discarded = others\n    .filter(\n      (node) => node.status === 'pruned' || node.status === 'rejected' || node.status === 'invalid',\n    )\n    .sort((a, b) => b.updatedSequence - a.updatedSequence)\n    .slice(0, limit)\n  lines.push('', 'Recently discarded:')\n  if (discarded.length === 0) lines.push('  none yet')\n  for (const node of discarded) lines.push(`  ${nodeLine(node)}: ${discardLine(node)}`)\n\n  const recent = [...state.edges()].reverse().slice(0, limit)\n  lines.push('', 'Recent proposals:')\n  if (recent.length === 0) lines.push('  none yet')\n  for (const edge of recent) {\n    const childId = edge.childNodeId\n    const label = edge.label.length > 0 ? edge.label : '(no label)'\n    lines.push(`  ${edge.operator} → ${childId}: ${label}`)\n  }\n\n  return lines.join('\\n')\n}\n\nfunction headerLine(state: SearchStateView): string {\n  const header = state.header!\n  return `search ${state.searchId} — ${header.subject} — ${header.objective.direction} ${header.objective.metric}`\n}\n\nfunction spendLine(state: SearchStateView): string {\n  const { audit } = state\n  const known = audit.spend.knownUsd.toFixed(2)\n  const floor = audit.spend.floorUsd.toFixed(2)\n  const unknown = audit.spend.unknownCostCells\n  const floorPart =\n    unknown > 0\n      ? ` + floor $${floor} across ${unknown} unknown-cost cell${unknown === 1 ? '' : 's'}`\n      : ''\n  return `  ${audit.nodes} nodes · ${audit.cells.settled} settled cells (${audit.cells.open} open) · $${known} known${floorPart}`\n}\n\nfunction statusLine(state: SearchStateView): string {\n  if (!state.closed) return '  status: open'\n  const claim = state.closed.claim\n  const claimText =\n    claim === null\n      ? ''\n      : claim.decision === 'ship'\n        ? ` · claim: ship ${claim.selected}`\n        : ` · claim: ${claim.decision}`\n  return `  status: closed (${state.closed.reason})${claimText}`\n}\n\nfunction nodeLine(node: SearchNode): string {\n  const rung = node.status === 'advanced' ? ` rung ${node.rung}` : ''\n  return `${node.nodeId} (${node.status ?? 'undecided'}${rung})`\n}\n\nfunction discardLine(node: SearchNode): string {\n  const decision = node.decisions.at(-1)\n  if (!decision) return 'no decision recorded'\n  const basis = decision.basis === null ? 'no measured basis' : formatEstimate(decision.basis)\n  return `${basis} — ${decision.rule}: ${decision.reason}`\n}\n\n/** `estimate.delta` in the objective's own units; positive is always better. */\nfunction signed(estimate: NodeEstimate, direction: 'maximize' | 'minimize'): number {\n  if (estimate.delta === null) return Number.NEGATIVE_INFINITY\n  return direction === 'maximize' ? estimate.delta : -estimate.delta\n}\n\n/** The display rules of search-tree-design §6.4: no whisker below 6 shared\n * units, a dashed descriptive interval with its exact sign p from 6 to 19,\n * a decision-grade whisker from `BOOTSTRAP_GATE_MIN_N`. */\nfunction formatEstimate(estimate: NodeEstimate): string {\n  const unit = (n: number) => `${n} unit${n === 1 ? '' : 's'}`\n  if (estimate.method === 'none') return `unknown (${unit(estimate.pairs)})`\n  if (estimate.indeterminate)\n    return `no measurable difference (indeterminate, ${unit(estimate.pairs)})`\n  const delta =\n    estimate.delta === null ? 'null' : (estimate.delta >= 0 ? '+' : '') + estimate.delta.toFixed(4)\n  if (estimate.method === 'insufficient') {\n    return `Δ=${delta} (${estimate.pairs} of ${DESCRIPTIVE_FROM} units, insufficient)`\n  }\n  const interval = estimate.interval\n    ? ` [${estimate.interval[0].toFixed(4)}, ${estimate.interval[1].toFixed(4)}]`\n    : ''\n  if (estimate.method === 'descriptive') {\n    const signP = estimate.exactSignP === null ? 'null' : estimate.exactSignP.toPrecision(3)\n    return `Δ=${delta}${interval} dashed, sign p=${signP} (${unit(estimate.pairs)})`\n  }\n  return `Δ=${delta}${interval} (${unit(estimate.pairs)}, ≥${BOOTSTRAP_GATE_MIN_N})`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAwBA,MAAM,mBAAmB,+BAA+B,GAAI;;AAe5D,SAAgB,mBAAmB,OAA4C;CAC7E,MAAM,SAAS,MAAM;CACrB,IAAI,CAAC,QAAQ,MAAM,IAAI,kBAAkB,UAAU,MAAM,SAAS,qBAAqB;CACvF,OAAO;EACL,UAAU,MAAM;EAChB,WAAW,OAAO,UAAU;EAC5B,cAAc,WAAW,MAAM,YAAY,QAAQ,OAAO;EAC1D,aAAa,WAAW,MAAM,WAAW,QAAQ,OAAO;CAC1D;AACF;;;;;;;;;;AAwBA,SAAgB,oBAAoB,OAAwB,SAAuC;CACjG,MAAM,SAAS,MAAM;CACrB,IAAI,CAAC,QAAQ,OAAO,UAAU,MAAM,SAAS;CAC7C,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,QAAQ,KAAK,CAAC,OAAO,UAAU,KAAK,GACtC,MAAM,IAAI,kBACR,8DAA8D,OAChE;CAEF,MAAM,QAAQ,QAAQ;CACtB,MAAM,YAAY,OAAO,UAAU;CACnC,MAAM,OAAO,MAAM;CACnB,MAAM,QAAkB;EAAC,WAAW,KAAK;EAAG,UAAU,KAAK;EAAG,WAAW,KAAK;CAAC;CAE/E,IAAI,SAAS,MAAM;EACjB,MAAM,KAAK,IAAI,eAAe;EAC9B,OAAO,MAAM,KAAK,IAAI;CACxB;CAEA,MAAM,SAAS,MAAM,MAAM,CAAC,CAAC,QAAQ,SAAS,KAAK,WAAW,IAAI;CAKlE,MAAM,UAJY,OACf,QAAQ,SAAS,KAAK,WAAW,SAAS,CAAC,CAC3C,KAAK,UAAU;EAAE;EAAM,UAAU,aAAa,OAAO,KAAK,QAAQ;GAAE,SAAS;GAAM;EAAM,CAAC;CAAE,EAEvE,CAAC,CACtB,QAAQ,EAAE,eAAe,SAAS,UAAU,IAAI,CAAC,CACjD,MAAM,GAAG,MAAM,OAAO,EAAE,UAAU,SAAS,IAAI,OAAO,EAAE,UAAU,SAAS,CAAC,CAAC,CAC7E,MAAM,GAAG,KAAK;CACjB,MAAM,KAAK,IAAI,2BAA2B,MAAM,SAAS;CACzD,IAAI,QAAQ,WAAW,GAAG,MAAM,KAAK,qBAAqB;CAC1D,KAAK,MAAM,EAAE,MAAM,cAAc,SAC/B,MAAM,KAAK,KAAK,SAAS,IAAI,EAAE,IAAI,eAAe,QAAQ,GAAG;CAE/D,MAAM,YAAY,OACf,QACE,SAAS,KAAK,WAAW,YAAY,KAAK,WAAW,cAAc,KAAK,WAAW,SACtF,CAAC,CACA,MAAM,GAAG,MAAM,EAAE,kBAAkB,EAAE,eAAe,CAAC,CACrD,MAAM,GAAG,KAAK;CACjB,MAAM,KAAK,IAAI,qBAAqB;CACpC,IAAI,UAAU,WAAW,GAAG,MAAM,KAAK,YAAY;CACnD,KAAK,MAAM,QAAQ,WAAW,MAAM,KAAK,KAAK,SAAS,IAAI,EAAE,IAAI,YAAY,IAAI,GAAG;CAEpF,MAAM,SAAS,CAAC,GAAG,MAAM,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,MAAM,GAAG,KAAK;CAC1D,MAAM,KAAK,IAAI,mBAAmB;CAClC,IAAI,OAAO,WAAW,GAAG,MAAM,KAAK,YAAY;CAChD,KAAK,MAAM,QAAQ,QAAQ;EACzB,MAAM,UAAU,KAAK;EACrB,MAAM,QAAQ,KAAK,MAAM,SAAS,IAAI,KAAK,QAAQ;EACnD,MAAM,KAAK,KAAK,KAAK,SAAS,KAAK,QAAQ,IAAI,OAAO;CACxD;CAEA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,WAAW,OAAgC;CAClD,MAAM,SAAS,MAAM;CACrB,OAAO,UAAU,MAAM,SAAS,KAAK,OAAO,QAAQ,KAAK,OAAO,UAAU,UAAU,GAAG,OAAO,UAAU;AAC1G;AAEA,SAAS,UAAU,OAAgC;CACjD,MAAM,EAAE,UAAU;CAClB,MAAM,QAAQ,MAAM,MAAM,SAAS,QAAQ,CAAC;CAC5C,MAAM,QAAQ,MAAM,MAAM,SAAS,QAAQ,CAAC;CAC5C,MAAM,UAAU,MAAM,MAAM;CAC5B,MAAM,YACJ,UAAU,IACN,aAAa,MAAM,UAAU,QAAQ,oBAAoB,YAAY,IAAI,KAAK,QAC9E;CACN,OAAO,KAAK,MAAM,MAAM,WAAW,MAAM,MAAM,QAAQ,kBAAkB,MAAM,MAAM,KAAK,YAAY,MAAM,QAAQ;AACtH;AAEA,SAAS,WAAW,OAAgC;CAClD,IAAI,CAAC,MAAM,QAAQ,OAAO;CAC1B,MAAM,QAAQ,MAAM,OAAO;CAC3B,MAAM,YACJ,UAAU,OACN,KACA,MAAM,aAAa,SACjB,kBAAkB,MAAM,aACxB,aAAa,MAAM;CAC3B,OAAO,qBAAqB,MAAM,OAAO,OAAO,GAAG;AACrD;AAEA,SAAS,SAAS,MAA0B;CAC1C,MAAM,OAAO,KAAK,WAAW,aAAa,SAAS,KAAK,SAAS;CACjE,OAAO,GAAG,KAAK,OAAO,IAAI,KAAK,UAAU,cAAc,KAAK;AAC9D;AAEA,SAAS,YAAY,MAA0B;CAC7C,MAAM,WAAW,KAAK,UAAU,GAAG,EAAE;CACrC,IAAI,CAAC,UAAU,OAAO;CAEtB,OAAO,GADO,SAAS,UAAU,OAAO,sBAAsB,eAAe,SAAS,KAAK,EAC3E,KAAK,SAAS,KAAK,IAAI,SAAS;AAClD;;AAGA,SAAS,OAAO,UAAwB,WAA4C;CAClF,IAAI,SAAS,UAAU,MAAM,OAAO,OAAO;CAC3C,OAAO,cAAc,aAAa,SAAS,QAAQ,CAAC,SAAS;AAC/D;;;;AAKA,SAAS,eAAe,UAAgC;CACtD,MAAM,QAAQ,MAAc,GAAG,EAAE,OAAO,MAAM,IAAI,KAAK;CACvD,IAAI,SAAS,WAAW,QAAQ,OAAO,YAAY,KAAK,SAAS,KAAK,EAAE;CACxE,IAAI,SAAS,eACX,OAAO,4CAA4C,KAAK,SAAS,KAAK,EAAE;CAC1E,MAAM,QACJ,SAAS,UAAU,OAAO,UAAU,SAAS,SAAS,IAAI,MAAM,MAAM,SAAS,MAAM,QAAQ,CAAC;CAChG,IAAI,SAAS,WAAW,gBACtB,OAAO,KAAK,MAAM,IAAI,SAAS,MAAM,MAAM,iBAAiB;CAE9D,MAAM,WAAW,SAAS,WACtB,KAAK,SAAS,SAAS,EAAE,CAAC,QAAQ,CAAC,EAAE,IAAI,SAAS,SAAS,EAAE,CAAC,QAAQ,CAAC,EAAE,KACzE;CACJ,IAAI,SAAS,WAAW,eAEtB,OAAO,KAAK,QAAQ,SAAS,kBADf,SAAS,eAAe,OAAO,SAAS,SAAS,WAAW,YAAY,CAAC,EAClC,IAAI,KAAK,SAAS,KAAK,EAAE;CAEhF,OAAO,KAAK,QAAQ,SAAS,IAAI,KAAK,SAAS,KAAK,EAAE;AACxD"}
+{"version":3,"file":"search-summary-BZkoLwbB.js","names":[],"sources":["../src/campaign/search-summary.ts"],"sourcesContent":["/**\n * The agent's view of a search: the same projection a human reads.\n *\n * `renderSearchSummary` turns a `SearchStateView` into compact text — the\n * leading nodes with their estimates, the most recently discarded ideas with\n * why they were discarded, and a log of recent proposals — as AIDE's journal\n * summary does for its own search. `agent-eval search show <ledger>` prints\n * the same text for a terminal; `runOptimization` puts it on every\n * `ProposeContext` a proposer receives.\n *\n * `searchProposerView` is the train-only read `ProposeContext.train` carries.\n * It has no parameter that can select another split, so a proposer built on\n * it cannot reach the selection or test split even by mistake: the sealed\n * splits (§6.6 of the search-tree design) are unreachable by construction,\n * not by convention.\n */\n\nimport { minimumPairsForPairedDeltaTest } from '../paired-delta-test'\nimport { BOOTSTRAP_GATE_MIN_N } from '../statistics'\nimport { estimateNode } from './estimate-node'\nimport { SearchLedgerError } from './search-ledger-errors'\nimport type { NodeEstimate, SearchSplit } from './search-ledger-types'\nimport type { SearchNode, SearchScoredCell, SearchStateView, SearchUnitScore } from './search-state'\n\nconst DESCRIPTIVE_FROM = minimumPairsForPairedDeltaTest(0.95)\n\n/** The proposer's own read of a search: its node's cells on the train split,\n * the only split a proposer may ever see. */\nexport interface SearchProposerView {\n  readonly searchId: string\n  readonly direction: 'maximize' | 'minimize'\n  /** `nodeId`'s scored train cells. Unscored cells (errored, cancelled or in\n   * flight) are absent, never zero. */\n  scoredCells(nodeId: string): readonly SearchScoredCell[]\n  /** Per-unit train-split means of `nodeId`, in unitId order. */\n  unitScores(nodeId: string): readonly SearchUnitScore[]\n}\n\n/** The train-only view `ProposeContext.train` carries. */\nexport function searchProposerView(state: SearchStateView): SearchProposerView {\n  const header = state.header\n  if (!header) throw new SearchLedgerError(`search ${state.searchId} has not been opened`)\n  return {\n    searchId: state.searchId,\n    direction: header.objective.direction,\n    scoredCells: (nodeId) => state.scoredCells(nodeId, 'train'),\n    unitScores: (nodeId) => state.unitScores(nodeId, 'train'),\n  }\n}\n\nexport interface SearchSummaryOptions {\n  /**\n   * The split every estimate in the summary is computed on. A proposer's\n   * context must always pass `'train'`: the selection and test splits stay\n   * unread by anything a proposer's text can reach. The CLI passes the\n   * search's own ranking split, `'selection'` when the search declares one\n   * else `'train'`, matching `searchPolicyView`.\n   */\n  split: SearchSplit\n  /** Nodes shown under \"Leading\" and under \"Recently discarded\". Default 5. */\n  limit?: number\n}\n\n/**\n * `SearchState` as compact text: a header line with the search's shape and\n * spend, the leading nodes against the root with their estimates, the most\n * recently discarded nodes with the measurement that discarded them, and a\n * log of recent proposals (operator and label — the ledger's inline,\n * redacted summary of each rationale; the full rationale is a blob this\n * function never reads, so it stays a pure, I/O-free projection of the\n * state already in hand).\n */\nexport function renderSearchSummary(state: SearchStateView, options: SearchSummaryOptions): string {\n  const header = state.header\n  if (!header) return `search ${state.searchId}: not yet opened`\n  const limit = options.limit ?? 5\n  if (limit < 1 || !Number.isInteger(limit)) {\n    throw new SearchLedgerError(\n      `renderSearchSummary: limit must be a positive integer, got ${limit}`,\n    )\n  }\n  const split = options.split\n  const direction = header.objective.direction\n  const root = state.rootNodeId\n  const lines: string[] = [headerLine(state), spendLine(state), statusLine(state)]\n\n  if (root === null) {\n    lines.push('', 'No nodes yet.')\n    return lines.join('\\n')\n  }\n\n  const others = state.nodes().filter((node) => node.nodeId !== root)\n  const estimated = others\n    .filter((node) => node.status !== 'invalid')\n    .map((node) => ({ node, estimate: estimateNode(state, node.nodeId, { against: root, split }) }))\n\n  const leading = estimated\n    .filter(({ estimate }) => estimate.delta !== null)\n    .sort((a, b) => signed(b.estimate, direction) - signed(a.estimate, direction))\n    .slice(0, limit)\n  lines.push('', `Leading nodes (vs root, ${split} split):`)\n  if (leading.length === 0) lines.push('  none measured yet')\n  for (const { node, estimate } of leading)\n    lines.push(`  ${nodeLine(node)}: ${formatEstimate(estimate)}`)\n\n  const discarded = others\n    .filter(\n      (node) => node.status === 'pruned' || node.status === 'rejected' || node.status === 'invalid',\n    )\n    .sort((a, b) => b.updatedSequence - a.updatedSequence)\n    .slice(0, limit)\n  lines.push('', 'Recently discarded:')\n  if (discarded.length === 0) lines.push('  none yet')\n  for (const node of discarded) lines.push(`  ${nodeLine(node)}: ${discardLine(node)}`)\n\n  const recent = [...state.edges()].reverse().slice(0, limit)\n  lines.push('', 'Recent proposals:')\n  if (recent.length === 0) lines.push('  none yet')\n  for (const edge of recent) {\n    const childId = edge.childNodeId\n    const label = edge.label.length > 0 ? edge.label : '(no label)'\n    lines.push(`  ${edge.operator} → ${childId}: ${label}`)\n  }\n\n  return lines.join('\\n')\n}\n\nfunction headerLine(state: SearchStateView): string {\n  const header = state.header!\n  return `search ${state.searchId} — ${header.subject} — ${header.objective.direction} ${header.objective.metric}`\n}\n\nfunction spendLine(state: SearchStateView): string {\n  const { audit } = state\n  const known = audit.spend.knownUsd.toFixed(2)\n  const floor = audit.spend.floorUsd.toFixed(2)\n  const unknown = audit.spend.unknownCostCells\n  const floorPart =\n    unknown > 0\n      ? ` + floor $${floor} across ${unknown} unknown-cost cell${unknown === 1 ? '' : 's'}`\n      : ''\n  return `  ${audit.nodes} nodes · ${audit.cells.settled} settled cells (${audit.cells.open} open) · $${known} known${floorPart}`\n}\n\nfunction statusLine(state: SearchStateView): string {\n  if (!state.closed) return '  status: open'\n  const claim = state.closed.claim\n  const claimText =\n    claim === null\n      ? ''\n      : claim.decision === 'ship'\n        ? ` · claim: ship ${claim.selected}`\n        : ` · claim: ${claim.decision}`\n  return `  status: closed (${state.closed.reason})${claimText}`\n}\n\nfunction nodeLine(node: SearchNode): string {\n  const rung = node.status === 'advanced' ? ` rung ${node.rung}` : ''\n  return `${node.nodeId} (${node.status ?? 'undecided'}${rung})`\n}\n\nfunction discardLine(node: SearchNode): string {\n  const decision = node.decisions.at(-1)\n  if (!decision) return 'no decision recorded'\n  const basis = decision.basis === null ? 'no measured basis' : formatEstimate(decision.basis)\n  return `${basis} — ${decision.rule}: ${decision.reason}`\n}\n\n/** `estimate.delta` in the objective's own units; positive is always better. */\nfunction signed(estimate: NodeEstimate, direction: 'maximize' | 'minimize'): number {\n  if (estimate.delta === null) return Number.NEGATIVE_INFINITY\n  return direction === 'maximize' ? estimate.delta : -estimate.delta\n}\n\n/** The display rules of search-tree-design §6.4: no whisker below 6 shared\n * units, a dashed descriptive interval with its exact sign p from 6 to 19,\n * a decision-grade whisker from `BOOTSTRAP_GATE_MIN_N`. */\nfunction formatEstimate(estimate: NodeEstimate): string {\n  const unit = (n: number) => `${n} unit${n === 1 ? '' : 's'}`\n  if (estimate.method === 'none') return `unknown (${unit(estimate.pairs)})`\n  if (estimate.indeterminate)\n    return `no measurable difference (indeterminate, ${unit(estimate.pairs)})`\n  const delta =\n    estimate.delta === null ? 'null' : (estimate.delta >= 0 ? '+' : '') + estimate.delta.toFixed(4)\n  if (estimate.method === 'insufficient') {\n    return `Δ=${delta} (${estimate.pairs} of ${DESCRIPTIVE_FROM} units, insufficient)`\n  }\n  const interval = estimate.interval\n    ? ` [${estimate.interval[0].toFixed(4)}, ${estimate.interval[1].toFixed(4)}]`\n    : ''\n  if (estimate.method === 'descriptive') {\n    const signP = estimate.exactSignP === null ? 'null' : estimate.exactSignP.toPrecision(3)\n    return `Δ=${delta}${interval} dashed, sign p=${signP} (${unit(estimate.pairs)})`\n  }\n  return `Δ=${delta}${interval} (${unit(estimate.pairs)}, ≥${BOOTSTRAP_GATE_MIN_N})`\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AAwBA,MAAM,mBAAmB,+BAA+B,GAAI;;AAe5D,SAAgB,mBAAmB,OAA4C;CAC7E,MAAM,SAAS,MAAM;CACrB,IAAI,CAAC,QAAQ,MAAM,IAAI,kBAAkB,UAAU,MAAM,SAAS,qBAAqB;CACvF,OAAO;EACL,UAAU,MAAM;EAChB,WAAW,OAAO,UAAU;EAC5B,cAAc,WAAW,MAAM,YAAY,QAAQ,OAAO;EAC1D,aAAa,WAAW,MAAM,WAAW,QAAQ,OAAO;CAC1D;AACF;;;;;;;;;;AAwBA,SAAgB,oBAAoB,OAAwB,SAAuC;CACjG,MAAM,SAAS,MAAM;CACrB,IAAI,CAAC,QAAQ,OAAO,UAAU,MAAM,SAAS;CAC7C,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,QAAQ,KAAK,CAAC,OAAO,UAAU,KAAK,GACtC,MAAM,IAAI,kBACR,8DAA8D,OAChE;CAEF,MAAM,QAAQ,QAAQ;CACtB,MAAM,YAAY,OAAO,UAAU;CACnC,MAAM,OAAO,MAAM;CACnB,MAAM,QAAkB;EAAC,WAAW,KAAK;EAAG,UAAU,KAAK;EAAG,WAAW,KAAK;CAAC;CAE/E,IAAI,SAAS,MAAM;EACjB,MAAM,KAAK,IAAI,eAAe;EAC9B,OAAO,MAAM,KAAK,IAAI;CACxB;CAEA,MAAM,SAAS,MAAM,MAAM,CAAC,CAAC,QAAQ,SAAS,KAAK,WAAW,IAAI;CAKlE,MAAM,UAJY,OACf,QAAQ,SAAS,KAAK,WAAW,SAAS,CAAC,CAC3C,KAAK,UAAU;EAAE;EAAM,UAAU,aAAa,OAAO,KAAK,QAAQ;GAAE,SAAS;GAAM;EAAM,CAAC;CAAE,EAEvE,CAAC,CACtB,QAAQ,EAAE,eAAe,SAAS,UAAU,IAAI,CAAC,CACjD,MAAM,GAAG,MAAM,OAAO,EAAE,UAAU,SAAS,IAAI,OAAO,EAAE,UAAU,SAAS,CAAC,CAAC,CAC7E,MAAM,GAAG,KAAK;CACjB,MAAM,KAAK,IAAI,2BAA2B,MAAM,SAAS;CACzD,IAAI,QAAQ,WAAW,GAAG,MAAM,KAAK,qBAAqB;CAC1D,KAAK,MAAM,EAAE,MAAM,cAAc,SAC/B,MAAM,KAAK,KAAK,SAAS,IAAI,EAAE,IAAI,eAAe,QAAQ,GAAG;CAE/D,MAAM,YAAY,OACf,QACE,SAAS,KAAK,WAAW,YAAY,KAAK,WAAW,cAAc,KAAK,WAAW,SACtF,CAAC,CACA,MAAM,GAAG,MAAM,EAAE,kBAAkB,EAAE,eAAe,CAAC,CACrD,MAAM,GAAG,KAAK;CACjB,MAAM,KAAK,IAAI,qBAAqB;CACpC,IAAI,UAAU,WAAW,GAAG,MAAM,KAAK,YAAY;CACnD,KAAK,MAAM,QAAQ,WAAW,MAAM,KAAK,KAAK,SAAS,IAAI,EAAE,IAAI,YAAY,IAAI,GAAG;CAEpF,MAAM,SAAS,CAAC,GAAG,MAAM,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,MAAM,GAAG,KAAK;CAC1D,MAAM,KAAK,IAAI,mBAAmB;CAClC,IAAI,OAAO,WAAW,GAAG,MAAM,KAAK,YAAY;CAChD,KAAK,MAAM,QAAQ,QAAQ;EACzB,MAAM,UAAU,KAAK;EACrB,MAAM,QAAQ,KAAK,MAAM,SAAS,IAAI,KAAK,QAAQ;EACnD,MAAM,KAAK,KAAK,KAAK,SAAS,KAAK,QAAQ,IAAI,OAAO;CACxD;CAEA,OAAO,MAAM,KAAK,IAAI;AACxB;AAEA,SAAS,WAAW,OAAgC;CAClD,MAAM,SAAS,MAAM;CACrB,OAAO,UAAU,MAAM,SAAS,KAAK,OAAO,QAAQ,KAAK,OAAO,UAAU,UAAU,GAAG,OAAO,UAAU;AAC1G;AAEA,SAAS,UAAU,OAAgC;CACjD,MAAM,EAAE,UAAU;CAClB,MAAM,QAAQ,MAAM,MAAM,SAAS,QAAQ,CAAC;CAC5C,MAAM,QAAQ,MAAM,MAAM,SAAS,QAAQ,CAAC;CAC5C,MAAM,UAAU,MAAM,MAAM;CAC5B,MAAM,YACJ,UAAU,IACN,aAAa,MAAM,UAAU,QAAQ,oBAAoB,YAAY,IAAI,KAAK,QAC9E;CACN,OAAO,KAAK,MAAM,MAAM,WAAW,MAAM,MAAM,QAAQ,kBAAkB,MAAM,MAAM,KAAK,YAAY,MAAM,QAAQ;AACtH;AAEA,SAAS,WAAW,OAAgC;CAClD,IAAI,CAAC,MAAM,QAAQ,OAAO;CAC1B,MAAM,QAAQ,MAAM,OAAO;CAC3B,MAAM,YACJ,UAAU,OACN,KACA,MAAM,aAAa,SACjB,kBAAkB,MAAM,aACxB,aAAa,MAAM;CAC3B,OAAO,qBAAqB,MAAM,OAAO,OAAO,GAAG;AACrD;AAEA,SAAS,SAAS,MAA0B;CAC1C,MAAM,OAAO,KAAK,WAAW,aAAa,SAAS,KAAK,SAAS;CACjE,OAAO,GAAG,KAAK,OAAO,IAAI,KAAK,UAAU,cAAc,KAAK;AAC9D;AAEA,SAAS,YAAY,MAA0B;CAC7C,MAAM,WAAW,KAAK,UAAU,GAAG,EAAE;CACrC,IAAI,CAAC,UAAU,OAAO;CAEtB,OAAO,GADO,SAAS,UAAU,OAAO,sBAAsB,eAAe,SAAS,KAAK,EAC3E,KAAK,SAAS,KAAK,IAAI,SAAS;AAClD;;AAGA,SAAS,OAAO,UAAwB,WAA4C;CAClF,IAAI,SAAS,UAAU,MAAM,OAAO,OAAO;CAC3C,OAAO,cAAc,aAAa,SAAS,QAAQ,CAAC,SAAS;AAC/D;;;;AAKA,SAAS,eAAe,UAAgC;CACtD,MAAM,QAAQ,MAAc,GAAG,EAAE,OAAO,MAAM,IAAI,KAAK;CACvD,IAAI,SAAS,WAAW,QAAQ,OAAO,YAAY,KAAK,SAAS,KAAK,EAAE;CACxE,IAAI,SAAS,eACX,OAAO,4CAA4C,KAAK,SAAS,KAAK,EAAE;CAC1E,MAAM,QACJ,SAAS,UAAU,OAAO,UAAU,SAAS,SAAS,IAAI,MAAM,MAAM,SAAS,MAAM,QAAQ,CAAC;CAChG,IAAI,SAAS,WAAW,gBACtB,OAAO,KAAK,MAAM,IAAI,SAAS,MAAM,MAAM,iBAAiB;CAE9D,MAAM,WAAW,SAAS,WACtB,KAAK,SAAS,SAAS,EAAE,CAAC,QAAQ,CAAC,EAAE,IAAI,SAAS,SAAS,EAAE,CAAC,QAAQ,CAAC,EAAE,KACzE;CACJ,IAAI,SAAS,WAAW,eAEtB,OAAO,KAAK,QAAQ,SAAS,kBADf,SAAS,eAAe,OAAO,SAAS,SAAS,WAAW,YAAY,CAAC,EAClC,IAAI,KAAK,SAAS,KAAK,EAAE;CAEhF,OAAO,KAAK,QAAQ,SAAS,IAAI,KAAK,SAAS,KAAK,EAAE;AACxD"}

package/dist/{semantic-concept-judge-BgWxa4L_.js → semantic-concept-judge-BbiYuv7U.js}

@@ -1,9 +1,9 @@
 import { i as CostLedger } from "./cost-ledger-_IbkejCF.js";
 import { p as Mutex } from "./journal-DvKCvjty.js";
 import { c as extractJsonPayload } from "./llm-client-DmCNzrc9.js";
-import { N as RawAnalystFindingSchema } from "./kind-factory-CdPPHtGZ.js";
+import { P as RawAnalystFindingSchema } from "./kind-factory-Cn1lbsAd.js";
 import { n as paidJsonChat, t as paidChat } from "./chat-json-call-Digk-pfF.js";
-import { n as decodeRawFindingArray } from "./dspy-rlm-engine-D0h-P2Mz.js";
+import { n as decodeRawFindingArray } from "./dspy-rlm-engine-DCQKNQMn.js";
 import { appendFileSync, existsSync, mkdirSync, readFileSync } from "node:fs";
 import { z } from "zod";
 import { dirname } from "node:path";
@@ -777,4 +777,4 @@ async function runSemanticConceptJudge(input, options) {
 //#endregion
 export { diffFindings as a, defaultIsMaterial as i, runSemanticConceptJudge as n, createChatTraceEngine as o, FindingsStore as r, SEMANTIC_CONCEPT_JUDGE_VERSION as t };
 
-//# sourceMappingURL=semantic-concept-judge-BgWxa4L_.js.map
+//# sourceMappingURL=semantic-concept-judge-BbiYuv7U.js.map

package/dist/{semantic-concept-judge-BgWxa4L_.js.map → semantic-concept-judge-BbiYuv7U.js.map}

@@ -1 +1 @@
-{"version":3,"file":"semantic-concept-judge-BgWxa4L_.js","names":[],"sources":["../src/analyst/chat-trace-engine.ts","../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * A `TraceAnalysisEngine` that runs entirely inside Node against a\n * caller-owned `ChatClient`.\n *\n * The other engine in this package, `createDspyRlmTraceEngine`, reaches the\n * DSPy RLM through a Python subprocess. That made every model-backed analyst\n * unreachable for a consumer that already owns a model seam and no Python:\n * `createTraceAnalyst` requires an engine, the only exported constructor\n * required a Python runner, so `buildDefaultAnalystRegistry()` registered the\n * deterministic analyst alone and `analystsFromRegistry` refused the result.\n *\n * This engine closes that path. It drives the same investigation contract —\n * bounded trace tools, a prose answer, a strict findings array — with native\n * function calls over the transport the caller already bound with\n * `createChatClient`. A caller holding a bare\n * `(request: LlmCallRequest) => Promise<LlmCallResult>` adapts it in one line:\n *\n *   createChatClient({ transport: 'custom', chat: call, defaultModel, maximumAttempts })\n *\n * agent-eval still executes no paid model and holds no provider credential.\n * Every call goes through `paidChat`, so the shared cost ledger reserves the\n * priced maximum before the call and settles the receipt after it.\n *\n * Shape of one investigation:\n *   1. Investigation turns. The model reads the trace store through the tools\n *      its analyst kind was given. A turn that requests no tool ends the\n *      phase; an exhausted iteration or tool budget also ends it.\n *   2. One report turn. The model returns `{ answer, findings }` as JSON,\n *      decoded by the same `decodeRawFindingArray` the Python bridge uses, so\n *      a row this engine accepts is a row that engine could report.\n */\n\nimport { z } from 'zod'\nimport { paidChat } from '../chat-json-call'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport {\n  extractJsonPayload,\n  type LlmCallRequest,\n  type LlmMessage,\n  type LlmThinkingMode,\n  type LlmToolCall,\n  type LlmToolDefinition,\n} from '../llm-client'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\nimport type { ChatClient } from './chat-client'\nimport {\n  DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS,\n  type TraceAnalysisEngine,\n  type TraceAnalysisEngineRequest,\n  type TraceAnalysisEngineResult,\n} from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { RawAnalystFindingSchema } from './finding-signature'\n\n/** Bumped whenever this engine's execution behavior changes. */\nconst CHAT_TRACE_ENGINE_VERSION = '1.0.0'\nconst ENGINE_ID = 'chat-trace'\n\n/** Marker appended to a tool result cut down to the retained-output budget. */\nconst TRUNCATION_MARKER = '\\n…[truncated to the analyst maxOutputChars budget]'\n\nexport interface ChatTraceEngineOptions {\n  /**\n   * Caller-owned transport. Build it with `createChatClient`; agent-eval never\n   * receives the provider credential.\n   */\n  chat: ChatClient\n  /** Model for every call. Defaults to the client's `defaultModel`. */\n  model?: string\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n  /** Completion cap per turn. Default: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. */\n  maxOutputTokens?: number\n  /** Sampling temperature. Omitted from the request when absent. */\n  temperature?: number\n  /** Provider reasoning mode. Omitted when the provider default should apply. */\n  thinking?: LlmThinkingMode\n  /** Per-call deadline handed to the transport. */\n  requestTimeoutMs?: number\n  /**\n   * How the report turn asks for structured output.\n   *\n   * `json-object` sends JSON mode, which every OpenAI-compatible endpoint\n   * accepts, and relies on the finding grammar the analyst kind already puts\n   * in its instructions. `json-schema` sends the generated response schema as\n   * well, which stricter endpoints honour and older ones reject. Default:\n   * `json-object`.\n   */\n  reportFormat?: 'json-object' | 'json-schema'\n}\n\n/**\n * The report envelope. Its `findings` rows are generated from\n * `RawAnalystFindingSchema`, so the schema offered to a provider and the\n * decoder that accepts the answer cannot drift apart.\n */\nfunction buildReportJsonSchema(): { name: string; schema: Record<string, unknown> } {\n  const row = z.toJSONSchema(RawAnalystFindingSchema, { target: 'draft-7' }) as Record<\n    string,\n    unknown\n  >\n  delete row.$schema\n  return {\n    name: 'trace_analysis_report',\n    schema: {\n      type: 'object',\n      additionalProperties: false,\n      required: ['answer', 'findings'],\n      properties: {\n        answer: { type: 'string', description: 'Direct prose answer to the question.' },\n        findings: { type: 'array', items: row },\n      },\n    },\n  }\n}\n\nconst REPORT_JSON_SCHEMA = buildReportJsonSchema()\n\n/**\n * Run bounded recursive trace analysis in-process over a caller-owned chat\n * transport. No Python, no subprocess, no loopback proxy.\n */\nexport function createChatTraceEngine(options: ChatTraceEngineOptions): TraceAnalysisEngine {\n  const model = resolveModel(options)\n  const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS\n  if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n    throw new TypeError('chat trace engine maxOutputTokens must be a positive safe integer')\n  }\n  if (\n    options.temperature !== undefined &&\n    (!Number.isFinite(options.temperature) || options.temperature < 0)\n  ) {\n    throw new TypeError('chat trace engine temperature must be a non-negative finite number')\n  }\n  if (\n    options.requestTimeoutMs !== undefined &&\n    (!Number.isSafeInteger(options.requestTimeoutMs) || options.requestTimeoutMs <= 0)\n  ) {\n    throw new TypeError('chat trace engine requestTimeoutMs must be a positive safe integer')\n  }\n  const reportFormat = options.reportFormat ?? 'json-object'\n  if (reportFormat !== 'json-object' && reportFormat !== 'json-schema') {\n    throw new TypeError(`chat trace engine reportFormat must be json-object or json-schema`)\n  }\n\n  return {\n    id: ENGINE_ID,\n    description:\n      'In-process recursive trace analysis over a caller-owned ChatClient with native tool calls.',\n    model,\n    version: CHAT_TRACE_ENGINE_VERSION,\n    executionConfig: {\n      kind: ENGINE_ID,\n      model,\n      transport: options.chat.transport,\n      maximum_attempts: options.chat.maximumAttempts ?? null,\n      pricing: options.pricing ? { ...options.pricing } : null,\n      max_output_tokens: maxOutputTokens,\n      temperature: options.temperature ?? null,\n      thinking: options.thinking ?? null,\n      request_timeout_ms: options.requestTimeoutMs ?? null,\n      report_format: reportFormat,\n      report_schema_name: REPORT_JSON_SCHEMA.name,\n    },\n    analyze: (request) => analyze(request, { ...options, model, maxOutputTokens, reportFormat }),\n  }\n}\n\ninterface ResolvedOptions extends ChatTraceEngineOptions {\n  model: string\n  maxOutputTokens: number\n  reportFormat: 'json-object' | 'json-schema'\n}\n\nasync function analyze(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n): Promise<TraceAnalysisEngineResult> {\n  // One model call is spent on the report, so a budget of one call would buy\n  // an answer with no investigation behind it. Refuse instead of pretending.\n  if (request.limits.maxLlmCalls < 2) {\n    throw new Error(\n      `chat trace engine reserves one model call for the report, so maxLlmCalls must be at least 2 (analyst '${request.analystId}' declared ${request.limits.maxLlmCalls})`,\n    )\n  }\n  const toolsByName = new Map(request.tools.map((tool) => [tool.name, tool]))\n  if (toolsByName.size !== request.tools.length) {\n    throw new Error(`chat trace engine received duplicate tool names for '${request.analystId}'`)\n  }\n  const toolDefinitions: LlmToolDefinition[] = request.tools.map((tool) => ({\n    type: 'function',\n    function: { name: tool.name, description: tool.description, parameters: tool.parameters },\n  }))\n  const investigationTurns = Math.max(\n    1,\n    Math.min(request.limits.maxIterations, request.limits.maxLlmCalls - 1),\n  )\n\n  const messages: LlmMessage[] = [\n    { role: 'system', content: systemPrompt(request, investigationTurns) },\n  ]\n  if (request.taskInputs) {\n    messages.push({ role: 'user', content: renderTaskInputs(request.taskInputs) })\n  }\n  messages.push({ role: 'user', content: request.question })\n\n  request.log?.('trace analyst engine started', {\n    engine: ENGINE_ID,\n    model: options.model,\n    transport: options.chat.transport,\n    tools: request.tools.map((tool) => tool.name),\n    limits: request.limits,\n  })\n\n  const trajectory: unknown[] = []\n  const servedModels = new Set<string>()\n  let modelCalls = 0\n  let toolCalls = 0\n  let truncatedToolResults = 0\n  let toolBudgetExhausted = false\n  let turnsUsed = 0\n  let stoppedOnAnswer = false\n\n  for (let turn = 0; turn < investigationTurns; turn++) {\n    turnsUsed = turn + 1\n    const response = await callModel(request, options, {\n      messages,\n      tools: toolDefinitions,\n      toolChoice: 'auto',\n      purpose: 'investigation',\n    })\n    modelCalls += 1\n    servedModels.add(response.servedModel ?? 'unreported')\n    const requested = response.toolCalls ?? []\n    messages.push({\n      role: 'assistant',\n      content: response.content,\n      ...(requested.length > 0 ? { toolCalls: requested } : {}),\n    })\n    trajectory.push({\n      turn: turnsUsed,\n      phase: 'investigation',\n      content: response.content,\n      tool_calls: requested.map((call) => ({ id: call.id, name: call.name })),\n      finish_reason: response.finishReason ?? null,\n    })\n    if (requested.length === 0) {\n      stoppedOnAnswer = true\n      break\n    }\n    for (const call of requested) {\n      if (toolCalls >= request.limits.maxToolCalls) {\n        toolBudgetExhausted = true\n        // Every requested id still needs an answer: a provider rejects the\n        // next turn when one tool call is left unanswered.\n        messages.push(toolMessage(call, exhaustedToolBudget(request.limits.maxToolCalls)))\n        continue\n      }\n      toolCalls += 1\n      const executed = await executeTool(call, toolsByName, request)\n      if (executed.truncated) truncatedToolResults += 1\n      messages.push(toolMessage(call, executed.payload))\n      trajectory.push({\n        turn: turnsUsed,\n        phase: 'tool',\n        name: call.name,\n        ok: executed.ok,\n        truncated: executed.truncated,\n        result_chars: executed.payload.length,\n      })\n    }\n    if (toolBudgetExhausted) break\n  }\n\n  messages.push({ role: 'user', content: reportPrompt(options.reportFormat) })\n  const report = await callModel(request, options, {\n    messages,\n    purpose: 'report',\n    json: options.reportFormat,\n  })\n  modelCalls += 1\n  servedModels.add(report.servedModel ?? 'unreported')\n  trajectory.push({\n    turn: turnsUsed + 1,\n    phase: 'report',\n    content: report.content,\n    finish_reason: report.finishReason ?? null,\n  })\n\n  const parsed = parseReport(report.content, request.analystId, (index, reason) => {\n    request.log?.('finding rejected: report row failed schema validation', {\n      engine: ENGINE_ID,\n      index,\n      reason,\n    })\n  })\n\n  const result: TraceAnalysisEngineResult = {\n    answer: parsed.answer,\n    findings: parsed.findings,\n    trajectory,\n    modelCalls,\n    toolCalls,\n    runtime: {\n      engine: ENGINE_ID,\n      model: options.model,\n      transport: options.chat.transport,\n      report_format: options.reportFormat,\n      investigation_turns: turnsUsed,\n      investigation_stopped_on_answer: stoppedOnAnswer,\n      tool_budget_exhausted: toolBudgetExhausted,\n      truncated_tool_results: truncatedToolResults,\n      rejected_findings: parsed.rejectedFindings,\n      // A gateway can answer `model: X` from another model on HTTP 200, so the\n      // ids the provider echoed travel with the result as evidence.\n      served_models: [...servedModels].sort(),\n      task_inputs: request.taskInputs ? 'prompt-delivered' : 'none',\n    },\n  }\n  request.log?.('trace analyst engine completed', {\n    engine: ENGINE_ID,\n    model_calls: result.modelCalls,\n    tool_calls: result.toolCalls,\n    findings: result.findings.length,\n  })\n  return result\n}\n\ninterface ModelTurn {\n  content: string\n  toolCalls?: LlmToolCall[]\n  servedModel?: string | null\n  finishReason?: string | null\n}\n\nasync function callModel(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n  turn: {\n    messages: LlmMessage[]\n    tools?: LlmToolDefinition[]\n    toolChoice?: LlmCallRequest['toolChoice']\n    purpose: 'investigation' | 'report'\n    json?: 'json-object' | 'json-schema'\n  },\n): Promise<ModelTurn> {\n  const chatRequest: LlmCallRequest = {\n    model: options.model,\n    // The transport may mutate what it is handed; the conversation is ours.\n    messages: turn.messages.map((message) => ({ ...message })),\n    maxTokens: options.maxOutputTokens,\n    ...(turn.tools && turn.tools.length > 0\n      ? { tools: turn.tools, ...(turn.toolChoice ? { toolChoice: turn.toolChoice } : {}) }\n      : {}),\n    ...(turn.json ? { jsonMode: true } : {}),\n    ...(turn.json === 'json-schema' ? { jsonSchema: REPORT_JSON_SCHEMA } : {}),\n    ...(options.temperature === undefined ? {} : { temperature: options.temperature }),\n    ...(options.thinking === undefined ? {} : { thinking: options.thinking }),\n    ...(options.requestTimeoutMs === undefined ? {} : { timeoutMs: options.requestTimeoutMs }),\n  }\n  const paid = await paidChat({\n    chat: options.chat,\n    request: chatRequest,\n    ledger: request.costLedger,\n    channel: 'analyst',\n    phase: request.costPhase,\n    actor: request.analystId,\n    ...(request.costTags ? { tags: request.costTags } : {}),\n    ...(options.pricing ? { pricing: options.pricing } : {}),\n    ...(request.signal ? { signal: request.signal } : {}),\n  })\n  if (!paid.succeeded) {\n    // The transport already spent its own attempts. A failure here ends the\n    // investigation: continuing would report an answer the model never gave.\n    throw new Error(\n      `chat trace engine ${turn.purpose} call failed for '${request.analystId}': ${paid.error.message}`,\n      { cause: paid.error },\n    )\n  }\n  return {\n    content: paid.response.content,\n    ...(paid.response.toolCalls ? { toolCalls: paid.response.toolCalls } : {}),\n    ...(paid.response.servedModel === undefined ? {} : { servedModel: paid.response.servedModel }),\n    ...(paid.response.finishReason === undefined\n      ? {}\n      : { finishReason: paid.response.finishReason }),\n  }\n}\n\ninterface ExecutedTool {\n  ok: boolean\n  truncated: boolean\n  payload: string\n}\n\nasync function executeTool(\n  call: LlmToolCall,\n  toolsByName: ReadonlyMap<string, TraceAnalysisToolDescriptor>,\n  request: TraceAnalysisEngineRequest,\n): Promise<ExecutedTool> {\n  const descriptor = toolsByName.get(call.name)\n  if (!descriptor) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `unknown tool '${call.name}'; available tools are ${[...toolsByName.keys()].sort().join(', ')}`,\n      ),\n    }\n  }\n  let args: unknown\n  try {\n    args = call.argumentsJson.trim() === '' ? {} : JSON.parse(call.argumentsJson)\n  } catch (error) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `arguments for '${call.name}' were not JSON: ${messageOf(error)}. Send valid JSON arguments.`,\n      ),\n    }\n  }\n  try {\n    const value = await descriptor.handler(\n      args,\n      request.signal ? { signal: request.signal } : undefined,\n    )\n    return truncateToolPayload(JSON.stringify(value ?? null), request.limits.maxOutputChars)\n  } catch (error) {\n    // A bounded store refuses an oversized or malformed read by design. That\n    // refusal is the model's next instruction, not the run's failure.\n    if (request.signal?.aborted) throw error\n    request.log?.('trace tool failed', {\n      engine: ENGINE_ID,\n      analyst_id: request.analystId,\n      tool: call.name,\n      reason: messageOf(error),\n    })\n    return { ok: false, truncated: false, payload: toolError(messageOf(error)) }\n  }\n}\n\nfunction truncateToolPayload(payload: string, maxOutputChars: number): ExecutedTool {\n  if (payload.length <= maxOutputChars) return { ok: true, truncated: false, payload }\n  const keep = Math.max(0, maxOutputChars - TRUNCATION_MARKER.length)\n  return { ok: true, truncated: true, payload: `${payload.slice(0, keep)}${TRUNCATION_MARKER}` }\n}\n\nfunction toolMessage(call: LlmToolCall, content: string): LlmMessage {\n  return { role: 'tool', toolCallId: call.id, content }\n}\n\nfunction toolError(reason: string): string {\n  return JSON.stringify({ error: reason })\n}\n\nfunction exhaustedToolBudget(maxToolCalls: number): string {\n  return toolError(\n    `the trace-tool budget of ${maxToolCalls} calls is spent; answer from the evidence already read`,\n  )\n}\n\nfunction systemPrompt(request: TraceAnalysisEngineRequest, investigationTurns: number): string {\n  return [\n    request.instructions.trim(),\n    [\n      'HOW THIS INVESTIGATION RUNS:',\n      `- You have ${investigationTurns} investigation turns and at most ${request.limits.maxToolCalls} trace-tool calls.`,\n      '- Call the trace tools to read the store. Never state a trace fact you did not read.',\n      `- A tool result longer than ${request.limits.maxOutputChars} characters is truncated; narrow the query instead of asking again.`,\n      '- A tool result carrying an \"error\" field is feedback: fix the call or take another route.',\n      '- Answer with no tool call once you have the evidence. You are then asked for the final report.',\n    ].join('\\n'),\n  ].join('\\n\\n')\n}\n\nfunction reportPrompt(reportFormat: 'json-object' | 'json-schema'): string {\n  return [\n    'Report now. Do not call any more tools.',\n    'Return one JSON object with exactly two fields:',\n    '  \"answer\": a direct prose answer to the question.',\n    '  \"findings\": an array of findings in the schema above. Emit [] when there is nothing to report.',\n    reportFormat === 'json-object'\n      ? 'Return the object alone, with no surrounding prose and no code fence.'\n      : 'Return the object in the response schema you were given.',\n  ].join('\\n')\n}\n\nfunction renderTaskInputs(taskInputs: Readonly<Record<string, unknown>>): string {\n  // The engine contract forbids dropping structured inputs. This engine has no\n  // code environment to bind them as variables, so they are delivered whole as\n  // conversation material and the runtime record says which way they arrived.\n  return [\n    'TASK INPUTS — structured material delivered with the question, not fetched through tools:',\n    JSON.stringify(taskInputs, null, 2),\n  ].join('\\n')\n}\n\ninterface ParsedReport {\n  answer: string\n  findings: TraceAnalysisEngineResult['findings']\n  rejectedFindings: number\n}\n\nfunction parseReport(\n  content: string,\n  analystId: string,\n  onRejectedFinding: (index: number, reason: string) => void,\n): ParsedReport {\n  const envelope = coerceReportEnvelope(content, analystId)\n  if (typeof envelope.answer !== 'string' || !envelope.answer.trim()) {\n    throw new Error(`chat trace engine report for '${analystId}' carried no answer`)\n  }\n  // Findings are model output: one malformed row is model noise, not an engine\n  // fault, and the rest of the paid investigation must survive it.\n  const decoded = decodeRawFindingArray(envelope.findings)\n  if (decoded.topLevelError !== undefined) {\n    throw new Error(\n      `chat trace engine report for '${analystId}' had a malformed findings array: ${decoded.topLevelError}`,\n    )\n  }\n  for (const rejection of decoded.rejected) {\n    onRejectedFinding(\n      rejection.index,\n      `${rejection.code}${rejection.path ? ` at ${rejection.path}` : ''}: ${rejection.message}`,\n    )\n  }\n  return {\n    answer: envelope.answer,\n    findings: decoded.accepted,\n    rejectedFindings: decoded.rejected.length,\n  }\n}\n\nfunction coerceReportEnvelope(content: string, analystId: string): Record<string, unknown> {\n  let value: unknown\n  try {\n    value = JSON.parse(extractJsonPayload(content))\n  } catch (error) {\n    throw new Error(`chat trace engine report for '${analystId}' was not JSON: ${messageOf(error)}`)\n  }\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) {\n    throw new Error(`chat trace engine report for '${analystId}' was not a JSON object`)\n  }\n  return value as Record<string, unknown>\n}\n\nfunction resolveModel(options: ChatTraceEngineOptions): string {\n  const model = options.model ?? options.chat.defaultModel\n  if (typeof model !== 'string' || !model.trim() || model !== model.trim()) {\n    throw new TypeError(\n      'chat trace engine needs a model: pass ChatTraceEngineOptions.model or bind defaultModel on the ChatClient',\n    )\n  }\n  return model\n}\n\nfunction messageOf(error: unknown): string {\n  return error instanceof Error ? error.message : String(error)\n}\n","/**\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 type { ChatClient } from './analyst/chat-client'\nimport { paidJsonChat } from './chat-json-call'\nimport {\n  CostLedger,\n  type CostLedgerHandle,\n  type CostReceipt,\n  type CustomTokenPricing,\n} from './cost-ledger'\nimport type { LlmCallRequest } 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  /** Caller-owned transport. Required: agent-eval executes no paid model. */\n  chat: ChatClient\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\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: { maxPerFileChars: number; maxSourceChars: number; maxHtmlChars: number },\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 = {\n    chat: options.chat,\n    model: options.model ?? options.chat.defaultModel ?? 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    ...(options.pricing ? { pricing: options.pricing } : {}),\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 paidJsonChat<{ summary: string; concepts: ConceptFinding[] }>({\n      chat: opts.chat,\n      request,\n      ledger: opts.costLedger,\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      tags: opts.costTags,\n      signal: opts.signal,\n      ...(opts.pricing ? { pricing: opts.pricing } : {}),\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid\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":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,MAAM,4BAA4B;AAClC,MAAM,YAAY;;AAGlB,MAAM,oBAAoB;;;;;;AAqC1B,SAAS,wBAA2E;CAClF,MAAM,MAAM,EAAE,aAAa,yBAAyB,EAAE,QAAQ,UAAU,CAAC;CAIzE,OAAO,IAAI;CACX,OAAO;EACL,MAAM;EACN,QAAQ;GACN,MAAM;GACN,sBAAsB;GACtB,UAAU,CAAC,UAAU,UAAU;GAC/B,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;IAAuC;IAC9E,UAAU;KAAE,MAAM;KAAS,OAAO;IAAI;GACxC;EACF;CACF;AACF;AAEA,MAAM,qBAAqB,sBAAsB;;;;;AAMjD,SAAgB,sBAAsB,SAAsD;CAC1F,MAAM,QAAQ,aAAa,OAAO;CAClC,MAAM,kBAAkB,QAAQ,mBAAA;CAChC,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,mEAAmE;CAEzF,IACE,QAAQ,gBAAgB,KAAA,MACvB,CAAC,OAAO,SAAS,QAAQ,WAAW,KAAK,QAAQ,cAAc,IAEhE,MAAM,IAAI,UAAU,oEAAoE;CAE1F,IACE,QAAQ,qBAAqB,KAAA,MAC5B,CAAC,OAAO,cAAc,QAAQ,gBAAgB,KAAK,QAAQ,oBAAoB,IAEhF,MAAM,IAAI,UAAU,oEAAoE;CAE1F,MAAM,eAAe,QAAQ,gBAAgB;CAC7C,IAAI,iBAAiB,iBAAiB,iBAAiB,eACrD,MAAM,IAAI,UAAU,mEAAmE;CAGzF,OAAO;EACL,IAAI;EACJ,aACE;EACF;EACA,SAAS;EACT,iBAAiB;GACf,MAAM;GACN;GACA,WAAW,QAAQ,KAAK;GACxB,kBAAkB,QAAQ,KAAK,mBAAmB;GAClD,SAAS,QAAQ,UAAU,EAAE,GAAG,QAAQ,QAAQ,IAAI;GACpD,mBAAmB;GACnB,aAAa,QAAQ,eAAe;GACpC,UAAU,QAAQ,YAAY;GAC9B,oBAAoB,QAAQ,oBAAoB;GAChD,eAAe;GACf,oBAAoB,mBAAmB;EACzC;EACA,UAAU,YAAY,QAAQ,SAAS;GAAE,GAAG;GAAS;GAAO;GAAiB;EAAa,CAAC;CAC7F;AACF;AAQA,eAAe,QACb,SACA,SACoC;CAGpC,IAAI,QAAQ,OAAO,cAAc,GAC/B,MAAM,IAAI,MACR,yGAAyG,QAAQ,UAAU,aAAa,QAAQ,OAAO,YAAY,EACrK;CAEF,MAAM,cAAc,IAAI,IAAI,QAAQ,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAC1E,IAAI,YAAY,SAAS,QAAQ,MAAM,QACrC,MAAM,IAAI,MAAM,wDAAwD,QAAQ,UAAU,EAAE;CAE9F,MAAM,kBAAuC,QAAQ,MAAM,KAAK,UAAU;EACxE,MAAM;EACN,UAAU;GAAE,MAAM,KAAK;GAAM,aAAa,KAAK;GAAa,YAAY,KAAK;EAAW;CAC1F,EAAE;CACF,MAAM,qBAAqB,KAAK,IAC9B,GACA,KAAK,IAAI,QAAQ,OAAO,eAAe,QAAQ,OAAO,cAAc,CAAC,CACvE;CAEA,MAAM,WAAyB,CAC7B;EAAE,MAAM;EAAU,SAAS,aAAa,SAAS,kBAAkB;CAAE,CACvE;CACA,IAAI,QAAQ,YACV,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,iBAAiB,QAAQ,UAAU;CAAE,CAAC;CAE/E,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,QAAQ;CAAS,CAAC;CAEzD,QAAQ,MAAM,gCAAgC;EAC5C,QAAQ;EACR,OAAO,QAAQ;EACf,WAAW,QAAQ,KAAK;EACxB,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;EAC5C,QAAQ,QAAQ;CAClB,CAAC;CAED,MAAM,aAAwB,CAAC;CAC/B,MAAM,+BAAe,IAAI,IAAY;CACrC,IAAI,aAAa;CACjB,IAAI,YAAY;CAChB,IAAI,uBAAuB;CAC3B,IAAI,sBAAsB;CAC1B,IAAI,YAAY;CAChB,IAAI,kBAAkB;CAEtB,KAAK,IAAI,OAAO,GAAG,OAAO,oBAAoB,QAAQ;EACpD,YAAY,OAAO;EACnB,MAAM,WAAW,MAAM,UAAU,SAAS,SAAS;GACjD;GACA,OAAO;GACP,YAAY;GACZ,SAAS;EACX,CAAC;EACD,cAAc;EACd,aAAa,IAAI,SAAS,eAAe,YAAY;EACrD,MAAM,YAAY,SAAS,aAAa,CAAC;EACzC,SAAS,KAAK;GACZ,MAAM;GACN,SAAS,SAAS;GAClB,GAAI,UAAU,SAAS,IAAI,EAAE,WAAW,UAAU,IAAI,CAAC;EACzD,CAAC;EACD,WAAW,KAAK;GACd,MAAM;GACN,OAAO;GACP,SAAS,SAAS;GAClB,YAAY,UAAU,KAAK,UAAU;IAAE,IAAI,KAAK;IAAI,MAAM,KAAK;GAAK,EAAE;GACtE,eAAe,SAAS,gBAAgB;EAC1C,CAAC;EACD,IAAI,UAAU,WAAW,GAAG;GAC1B,kBAAkB;GAClB;EACF;EACA,KAAK,MAAM,QAAQ,WAAW;GAC5B,IAAI,aAAa,QAAQ,OAAO,cAAc;IAC5C,sBAAsB;IAGtB,SAAS,KAAK,YAAY,MAAM,oBAAoB,QAAQ,OAAO,YAAY,CAAC,CAAC;IACjF;GACF;GACA,aAAa;GACb,MAAM,WAAW,MAAM,YAAY,MAAM,aAAa,OAAO;GAC7D,IAAI,SAAS,WAAW,wBAAwB;GAChD,SAAS,KAAK,YAAY,MAAM,SAAS,OAAO,CAAC;GACjD,WAAW,KAAK;IACd,MAAM;IACN,OAAO;IACP,MAAM,KAAK;IACX,IAAI,SAAS;IACb,WAAW,SAAS;IACpB,cAAc,SAAS,QAAQ;GACjC,CAAC;EACH;EACA,IAAI,qBAAqB;CAC3B;CAEA,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,aAAa,QAAQ,YAAY;CAAE,CAAC;CAC3E,MAAM,SAAS,MAAM,UAAU,SAAS,SAAS;EAC/C;EACA,SAAS;EACT,MAAM,QAAQ;CAChB,CAAC;CACD,cAAc;CACd,aAAa,IAAI,OAAO,eAAe,YAAY;CACnD,WAAW,KAAK;EACd,MAAM,YAAY;EAClB,OAAO;EACP,SAAS,OAAO;EAChB,eAAe,OAAO,gBAAgB;CACxC,CAAC;CAED,MAAM,SAAS,YAAY,OAAO,SAAS,QAAQ,YAAY,OAAO,WAAW;EAC/E,QAAQ,MAAM,yDAAyD;GACrE,QAAQ;GACR;GACA;EACF,CAAC;CACH,CAAC;CAED,MAAM,SAAoC;EACxC,QAAQ,OAAO;EACf,UAAU,OAAO;EACjB;EACA;EACA;EACA,SAAS;GACP,QAAQ;GACR,OAAO,QAAQ;GACf,WAAW,QAAQ,KAAK;GACxB,eAAe,QAAQ;GACvB,qBAAqB;GACrB,iCAAiC;GACjC,uBAAuB;GACvB,wBAAwB;GACxB,mBAAmB,OAAO;GAG1B,eAAe,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;GACtC,aAAa,QAAQ,aAAa,qBAAqB;EACzD;CACF;CACA,QAAQ,MAAM,kCAAkC;EAC9C,QAAQ;EACR,aAAa,OAAO;EACpB,YAAY,OAAO;EACnB,UAAU,OAAO,SAAS;CAC5B,CAAC;CACD,OAAO;AACT;AASA,eAAe,UACb,SACA,SACA,MAOoB;CACpB,MAAM,cAA8B;EAClC,OAAO,QAAQ;EAEf,UAAU,KAAK,SAAS,KAAK,aAAa,EAAE,GAAG,QAAQ,EAAE;EACzD,WAAW,QAAQ;EACnB,GAAI,KAAK,SAAS,KAAK,MAAM,SAAS,IAClC;GAAE,OAAO,KAAK;GAAO,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EAAG,IACjF,CAAC;EACL,GAAI,KAAK,OAAO,EAAE,UAAU,KAAK,IAAI,CAAC;EACtC,GAAI,KAAK,SAAS,gBAAgB,EAAE,YAAY,mBAAmB,IAAI,CAAC;EACxE,GAAI,QAAQ,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,QAAQ,YAAY;EAChF,GAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;EACvE,GAAI,QAAQ,qBAAqB,KAAA,IAAY,CAAC,IAAI,EAAE,WAAW,QAAQ,iBAAiB;CAC1F;CACA,MAAM,OAAO,MAAM,SAAS;EAC1B,MAAM,QAAQ;EACd,SAAS;EACT,QAAQ,QAAQ;EAChB,SAAS;EACT,OAAO,QAAQ;EACf,OAAO,QAAQ;EACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;EACrD,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;CACrD,CAAC;CACD,IAAI,CAAC,KAAK,WAGR,MAAM,IAAI,MACR,qBAAqB,KAAK,QAAQ,oBAAoB,QAAQ,UAAU,KAAK,KAAK,MAAM,WACxF,EAAE,OAAO,KAAK,MAAM,CACtB;CAEF,OAAO;EACL,SAAS,KAAK,SAAS;EACvB,GAAI,KAAK,SAAS,YAAY,EAAE,WAAW,KAAK,SAAS,UAAU,IAAI,CAAC;EACxE,GAAI,KAAK,SAAS,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,KAAK,SAAS,YAAY;EAC5F,GAAI,KAAK,SAAS,iBAAiB,KAAA,IAC/B,CAAC,IACD,EAAE,cAAc,KAAK,SAAS,aAAa;CACjD;AACF;AAQA,eAAe,YACb,MACA,aACA,SACuB;CACvB,MAAM,aAAa,YAAY,IAAI,KAAK,IAAI;CAC5C,IAAI,CAAC,YACH,OAAO;EACL,IAAI;EACJ,WAAW;EACX,SAAS,UACP,iBAAiB,KAAK,KAAK,yBAAyB,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC9F;CACF;CAEF,IAAI;CACJ,IAAI;EACF,OAAO,KAAK,cAAc,KAAK,MAAM,KAAK,CAAC,IAAI,KAAK,MAAM,KAAK,aAAa;CAC9E,SAAS,OAAO;EACd,OAAO;GACL,IAAI;GACJ,WAAW;GACX,SAAS,UACP,kBAAkB,KAAK,KAAK,mBAAmB,UAAU,KAAK,EAAE,6BAClE;EACF;CACF;CACA,IAAI;EACF,MAAM,QAAQ,MAAM,WAAW,QAC7B,MACA,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,KAAA,CAChD;EACA,OAAO,oBAAoB,KAAK,UAAU,SAAS,IAAI,GAAG,QAAQ,OAAO,cAAc;CACzF,SAAS,OAAO;EAGd,IAAI,QAAQ,QAAQ,SAAS,MAAM;EACnC,QAAQ,MAAM,qBAAqB;GACjC,QAAQ;GACR,YAAY,QAAQ;GACpB,MAAM,KAAK;GACX,QAAQ,UAAU,KAAK;EACzB,CAAC;EACD,OAAO;GAAE,IAAI;GAAO,WAAW;GAAO,SAAS,UAAU,UAAU,KAAK,CAAC;EAAE;CAC7E;AACF;AAEA,SAAS,oBAAoB,SAAiB,gBAAsC;CAClF,IAAI,QAAQ,UAAU,gBAAgB,OAAO;EAAE,IAAI;EAAM,WAAW;EAAO;CAAQ;CACnF,MAAM,OAAO,KAAK,IAAI,GAAG,iBAAiB,EAAwB;CAClE,OAAO;EAAE,IAAI;EAAM,WAAW;EAAM,SAAS,GAAG,QAAQ,MAAM,GAAG,IAAI,IAAI;CAAoB;AAC/F;AAEA,SAAS,YAAY,MAAmB,SAA6B;CACnE,OAAO;EAAE,MAAM;EAAQ,YAAY,KAAK;EAAI;CAAQ;AACtD;AAEA,SAAS,UAAU,QAAwB;CACzC,OAAO,KAAK,UAAU,EAAE,OAAO,OAAO,CAAC;AACzC;AAEA,SAAS,oBAAoB,cAA8B;CACzD,OAAO,UACL,4BAA4B,aAAa,uDAC3C;AACF;AAEA,SAAS,aAAa,SAAqC,oBAAoC;CAC7F,OAAO,CACL,QAAQ,aAAa,KAAK,GAC1B;EACE;EACA,cAAc,mBAAmB,mCAAmC,QAAQ,OAAO,aAAa;EAChG;EACA,+BAA+B,QAAQ,OAAO,eAAe;EAC7D;EACA;CACF,CAAC,CAAC,KAAK,IAAI,CACb,CAAC,CAAC,KAAK,MAAM;AACf;AAEA,SAAS,aAAa,cAAqD;CACzE,OAAO;EACL;EACA;EACA;EACA;EACA,iBAAiB,gBACb,0EACA;CACN,CAAC,CAAC,KAAK,IAAI;AACb;AAEA,SAAS,iBAAiB,YAAuD;CAI/E,OAAO,CACL,6FACA,KAAK,UAAU,YAAY,MAAM,CAAC,CACpC,CAAC,CAAC,KAAK,IAAI;AACb;AAQA,SAAS,YACP,SACA,WACA,mBACc;CACd,MAAM,WAAW,qBAAqB,SAAS,SAAS;CACxD,IAAI,OAAO,SAAS,WAAW,YAAY,CAAC,SAAS,OAAO,KAAK,GAC/D,MAAM,IAAI,MAAM,iCAAiC,UAAU,oBAAoB;CAIjF,MAAM,UAAU,sBAAsB,SAAS,QAAQ;CACvD,IAAI,QAAQ,kBAAkB,KAAA,GAC5B,MAAM,IAAI,MACR,iCAAiC,UAAU,oCAAoC,QAAQ,eACzF;CAEF,KAAK,MAAM,aAAa,QAAQ,UAC9B,kBACE,UAAU,OACV,GAAG,UAAU,OAAO,UAAU,OAAO,OAAO,UAAU,SAAS,GAAG,IAAI,UAAU,SAClF;CAEF,OAAO;EACL,QAAQ,SAAS;EACjB,UAAU,QAAQ;EAClB,kBAAkB,QAAQ,SAAS;CACrC;AACF;AAEA,SAAS,qBAAqB,SAAiB,WAA4C;CACzF,IAAI;CACJ,IAAI;EACF,QAAQ,KAAK,MAAM,mBAAmB,OAAO,CAAC;CAChD,SAAS,OAAO;EACd,MAAM,IAAI,MAAM,iCAAiC,UAAU,kBAAkB,UAAU,KAAK,GAAG;CACjG;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,MAAM,iCAAiC,UAAU,wBAAwB;CAErF,OAAO;AACT;AAEA,SAAS,aAAa,SAAyC;CAC7D,MAAM,QAAQ,QAAQ,SAAS,QAAQ,KAAK;CAC5C,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,UAAU,MAAM,KAAK,GACrE,MAAM,IAAI,UACR,2GACF;CAEF,OAAO;AACT;AAEA,SAAS,UAAU,OAAwB;CACzC,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC9D;;;;;;;;;;;;;AChiBA,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;AAmCA,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,SACqC;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,OAAO;EACX,MAAM,QAAQ;EACd,OAAO,QAAQ,SAAS,QAAQ,KAAK,gBAAgB;EACrD,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,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,aAA8D;GAC/E,MAAM,KAAK;GACX;GACA,QAAQ,KAAK;GACb,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,MAAM,KAAK;GACX,QAAQ,KAAK;GACb,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;EAClD,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU;EAElB,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-BbiYuv7U.js","names":[],"sources":["../src/analyst/chat-trace-engine.ts","../src/locked-jsonl-appender.ts","../src/analyst/findings-store.ts","../src/semantic-concept-judge.ts"],"sourcesContent":["/**\n * A `TraceAnalysisEngine` that runs entirely inside Node against a\n * caller-owned `ChatClient`.\n *\n * The other engine in this package, `createDspyRlmTraceEngine`, reaches the\n * DSPy RLM through a Python subprocess. That made every model-backed analyst\n * unreachable for a consumer that already owns a model seam and no Python:\n * `createTraceAnalyst` requires an engine, the only exported constructor\n * required a Python runner, so `buildDefaultAnalystRegistry()` registered the\n * deterministic analyst alone and `analystsFromRegistry` refused the result.\n *\n * This engine closes that path. It drives the same investigation contract —\n * bounded trace tools, a prose answer, a strict findings array — with native\n * function calls over the transport the caller already bound with\n * `createChatClient`. A caller holding a bare\n * `(request: LlmCallRequest) => Promise<LlmCallResult>` adapts it in one line:\n *\n *   createChatClient({ transport: 'custom', chat: call, defaultModel, maximumAttempts })\n *\n * agent-eval still executes no paid model and holds no provider credential.\n * Every call goes through `paidChat`, so the shared cost ledger reserves the\n * priced maximum before the call and settles the receipt after it.\n *\n * Shape of one investigation:\n *   1. Investigation turns. The model reads the trace store through the tools\n *      its analyst kind was given. A turn that requests no tool ends the\n *      phase; an exhausted iteration or tool budget also ends it.\n *   2. One report turn. The model returns `{ answer, findings }` as JSON,\n *      decoded by the same `decodeRawFindingArray` the Python bridge uses, so\n *      a row this engine accepts is a row that engine could report.\n */\n\nimport { z } from 'zod'\nimport { paidChat } from '../chat-json-call'\nimport type { CustomTokenPricing } from '../cost-ledger'\nimport {\n  extractJsonPayload,\n  type LlmCallRequest,\n  type LlmMessage,\n  type LlmThinkingMode,\n  type LlmToolCall,\n  type LlmToolDefinition,\n} from '../llm-client'\nimport type { TraceAnalysisToolDescriptor } from '../trace-analyst/tools'\nimport type { ChatClient } from './chat-client'\nimport {\n  DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS,\n  type TraceAnalysisEngine,\n  type TraceAnalysisEngineRequest,\n  type TraceAnalysisEngineResult,\n} from './engine'\nimport { decodeRawFindingArray } from './finding-codec'\nimport { RawAnalystFindingSchema } from './finding-signature'\n\n/** Bumped whenever this engine's execution behavior changes. */\nconst CHAT_TRACE_ENGINE_VERSION = '1.0.0'\nconst ENGINE_ID = 'chat-trace'\n\n/** Marker appended to a tool result cut down to the retained-output budget. */\nconst TRUNCATION_MARKER = '\\n…[truncated to the analyst maxOutputChars budget]'\n\nexport interface ChatTraceEngineOptions {\n  /**\n   * Caller-owned transport. Build it with `createChatClient`; agent-eval never\n   * receives the provider credential.\n   */\n  chat: ChatClient\n  /** Model for every call. Defaults to the client's `defaultModel`. */\n  model?: string\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\n  /** Completion cap per turn. Default: {@link DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS}. */\n  maxOutputTokens?: number\n  /** Sampling temperature. Omitted from the request when absent. */\n  temperature?: number\n  /** Provider reasoning mode. Omitted when the provider default should apply. */\n  thinking?: LlmThinkingMode\n  /** Per-call deadline handed to the transport. */\n  requestTimeoutMs?: number\n  /**\n   * How the report turn asks for structured output.\n   *\n   * `json-object` sends JSON mode, which every OpenAI-compatible endpoint\n   * accepts, and relies on the finding grammar the analyst kind already puts\n   * in its instructions. `json-schema` sends the generated response schema as\n   * well, which stricter endpoints honour and older ones reject. Default:\n   * `json-object`.\n   */\n  reportFormat?: 'json-object' | 'json-schema'\n}\n\n/**\n * The report envelope. Its `findings` rows are generated from\n * `RawAnalystFindingSchema`, so the schema offered to a provider and the\n * decoder that accepts the answer cannot drift apart.\n */\nfunction buildReportJsonSchema(): { name: string; schema: Record<string, unknown> } {\n  const row = z.toJSONSchema(RawAnalystFindingSchema, { target: 'draft-7' }) as Record<\n    string,\n    unknown\n  >\n  delete row.$schema\n  return {\n    name: 'trace_analysis_report',\n    schema: {\n      type: 'object',\n      additionalProperties: false,\n      required: ['answer', 'findings'],\n      properties: {\n        answer: { type: 'string', description: 'Direct prose answer to the question.' },\n        findings: { type: 'array', items: row },\n      },\n    },\n  }\n}\n\nconst REPORT_JSON_SCHEMA = buildReportJsonSchema()\n\n/**\n * Run bounded recursive trace analysis in-process over a caller-owned chat\n * transport. No Python, no subprocess, no loopback proxy.\n */\nexport function createChatTraceEngine(options: ChatTraceEngineOptions): TraceAnalysisEngine {\n  const model = resolveModel(options)\n  const maxOutputTokens = options.maxOutputTokens ?? DEFAULT_TRACE_ANALYST_OUTPUT_TOKENS\n  if (!Number.isSafeInteger(maxOutputTokens) || maxOutputTokens <= 0) {\n    throw new TypeError('chat trace engine maxOutputTokens must be a positive safe integer')\n  }\n  if (\n    options.temperature !== undefined &&\n    (!Number.isFinite(options.temperature) || options.temperature < 0)\n  ) {\n    throw new TypeError('chat trace engine temperature must be a non-negative finite number')\n  }\n  if (\n    options.requestTimeoutMs !== undefined &&\n    (!Number.isSafeInteger(options.requestTimeoutMs) || options.requestTimeoutMs <= 0)\n  ) {\n    throw new TypeError('chat trace engine requestTimeoutMs must be a positive safe integer')\n  }\n  const reportFormat = options.reportFormat ?? 'json-object'\n  if (reportFormat !== 'json-object' && reportFormat !== 'json-schema') {\n    throw new TypeError(`chat trace engine reportFormat must be json-object or json-schema`)\n  }\n\n  return {\n    id: ENGINE_ID,\n    description:\n      'In-process recursive trace analysis over a caller-owned ChatClient with native tool calls.',\n    model,\n    version: CHAT_TRACE_ENGINE_VERSION,\n    executionConfig: {\n      kind: ENGINE_ID,\n      model,\n      transport: options.chat.transport,\n      maximum_attempts: options.chat.maximumAttempts ?? null,\n      pricing: options.pricing ? { ...options.pricing } : null,\n      max_output_tokens: maxOutputTokens,\n      temperature: options.temperature ?? null,\n      thinking: options.thinking ?? null,\n      request_timeout_ms: options.requestTimeoutMs ?? null,\n      report_format: reportFormat,\n      report_schema_name: REPORT_JSON_SCHEMA.name,\n    },\n    analyze: (request) => analyze(request, { ...options, model, maxOutputTokens, reportFormat }),\n  }\n}\n\ninterface ResolvedOptions extends ChatTraceEngineOptions {\n  model: string\n  maxOutputTokens: number\n  reportFormat: 'json-object' | 'json-schema'\n}\n\nasync function analyze(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n): Promise<TraceAnalysisEngineResult> {\n  // One model call is spent on the report, so a budget of one call would buy\n  // an answer with no investigation behind it. Refuse instead of pretending.\n  if (request.limits.maxLlmCalls < 2) {\n    throw new Error(\n      `chat trace engine reserves one model call for the report, so maxLlmCalls must be at least 2 (analyst '${request.analystId}' declared ${request.limits.maxLlmCalls})`,\n    )\n  }\n  const toolsByName = new Map(request.tools.map((tool) => [tool.name, tool]))\n  if (toolsByName.size !== request.tools.length) {\n    throw new Error(`chat trace engine received duplicate tool names for '${request.analystId}'`)\n  }\n  const toolDefinitions: LlmToolDefinition[] = request.tools.map((tool) => ({\n    type: 'function',\n    function: { name: tool.name, description: tool.description, parameters: tool.parameters },\n  }))\n  const investigationTurns = Math.max(\n    1,\n    Math.min(request.limits.maxIterations, request.limits.maxLlmCalls - 1),\n  )\n\n  const messages: LlmMessage[] = [\n    { role: 'system', content: systemPrompt(request, investigationTurns) },\n  ]\n  if (request.taskInputs) {\n    messages.push({ role: 'user', content: renderTaskInputs(request.taskInputs) })\n  }\n  messages.push({ role: 'user', content: request.question })\n\n  request.log?.('trace analyst engine started', {\n    engine: ENGINE_ID,\n    model: options.model,\n    transport: options.chat.transport,\n    tools: request.tools.map((tool) => tool.name),\n    limits: request.limits,\n  })\n\n  const trajectory: unknown[] = []\n  const servedModels = new Set<string>()\n  let modelCalls = 0\n  let toolCalls = 0\n  let truncatedToolResults = 0\n  let toolBudgetExhausted = false\n  let turnsUsed = 0\n  let stoppedOnAnswer = false\n\n  for (let turn = 0; turn < investigationTurns; turn++) {\n    turnsUsed = turn + 1\n    const response = await callModel(request, options, {\n      messages,\n      tools: toolDefinitions,\n      toolChoice: 'auto',\n      purpose: 'investigation',\n    })\n    modelCalls += 1\n    servedModels.add(response.servedModel ?? 'unreported')\n    const requested = response.toolCalls ?? []\n    messages.push({\n      role: 'assistant',\n      content: response.content,\n      ...(requested.length > 0 ? { toolCalls: requested } : {}),\n    })\n    trajectory.push({\n      turn: turnsUsed,\n      phase: 'investigation',\n      content: response.content,\n      tool_calls: requested.map((call) => ({ id: call.id, name: call.name })),\n      finish_reason: response.finishReason ?? null,\n    })\n    if (requested.length === 0) {\n      stoppedOnAnswer = true\n      break\n    }\n    for (const call of requested) {\n      if (toolCalls >= request.limits.maxToolCalls) {\n        toolBudgetExhausted = true\n        // Every requested id still needs an answer: a provider rejects the\n        // next turn when one tool call is left unanswered.\n        messages.push(toolMessage(call, exhaustedToolBudget(request.limits.maxToolCalls)))\n        continue\n      }\n      toolCalls += 1\n      const executed = await executeTool(call, toolsByName, request)\n      if (executed.truncated) truncatedToolResults += 1\n      messages.push(toolMessage(call, executed.payload))\n      trajectory.push({\n        turn: turnsUsed,\n        phase: 'tool',\n        name: call.name,\n        ok: executed.ok,\n        truncated: executed.truncated,\n        result_chars: executed.payload.length,\n      })\n    }\n    if (toolBudgetExhausted) break\n  }\n\n  messages.push({ role: 'user', content: reportPrompt(options.reportFormat) })\n  const report = await callModel(request, options, {\n    messages,\n    purpose: 'report',\n    json: options.reportFormat,\n  })\n  modelCalls += 1\n  servedModels.add(report.servedModel ?? 'unreported')\n  trajectory.push({\n    turn: turnsUsed + 1,\n    phase: 'report',\n    content: report.content,\n    finish_reason: report.finishReason ?? null,\n  })\n\n  const parsed = parseReport(report.content, request.analystId, (index, reason) => {\n    request.log?.('finding rejected: report row failed schema validation', {\n      engine: ENGINE_ID,\n      index,\n      reason,\n    })\n  })\n\n  const result: TraceAnalysisEngineResult = {\n    answer: parsed.answer,\n    findings: parsed.findings,\n    trajectory,\n    modelCalls,\n    toolCalls,\n    runtime: {\n      engine: ENGINE_ID,\n      model: options.model,\n      transport: options.chat.transport,\n      report_format: options.reportFormat,\n      investigation_turns: turnsUsed,\n      investigation_stopped_on_answer: stoppedOnAnswer,\n      tool_budget_exhausted: toolBudgetExhausted,\n      truncated_tool_results: truncatedToolResults,\n      rejected_findings: parsed.rejectedFindings,\n      // A gateway can answer `model: X` from another model on HTTP 200, so the\n      // ids the provider echoed travel with the result as evidence.\n      served_models: [...servedModels].sort(),\n      task_inputs: request.taskInputs ? 'prompt-delivered' : 'none',\n    },\n  }\n  request.log?.('trace analyst engine completed', {\n    engine: ENGINE_ID,\n    model_calls: result.modelCalls,\n    tool_calls: result.toolCalls,\n    findings: result.findings.length,\n  })\n  return result\n}\n\ninterface ModelTurn {\n  content: string\n  toolCalls?: LlmToolCall[]\n  servedModel?: string | null\n  finishReason?: string | null\n}\n\nasync function callModel(\n  request: TraceAnalysisEngineRequest,\n  options: ResolvedOptions,\n  turn: {\n    messages: LlmMessage[]\n    tools?: LlmToolDefinition[]\n    toolChoice?: LlmCallRequest['toolChoice']\n    purpose: 'investigation' | 'report'\n    json?: 'json-object' | 'json-schema'\n  },\n): Promise<ModelTurn> {\n  const chatRequest: LlmCallRequest = {\n    model: options.model,\n    // The transport may mutate what it is handed; the conversation is ours.\n    messages: turn.messages.map((message) => ({ ...message })),\n    maxTokens: options.maxOutputTokens,\n    ...(turn.tools && turn.tools.length > 0\n      ? { tools: turn.tools, ...(turn.toolChoice ? { toolChoice: turn.toolChoice } : {}) }\n      : {}),\n    ...(turn.json ? { jsonMode: true } : {}),\n    ...(turn.json === 'json-schema' ? { jsonSchema: REPORT_JSON_SCHEMA } : {}),\n    ...(options.temperature === undefined ? {} : { temperature: options.temperature }),\n    ...(options.thinking === undefined ? {} : { thinking: options.thinking }),\n    ...(options.requestTimeoutMs === undefined ? {} : { timeoutMs: options.requestTimeoutMs }),\n  }\n  const paid = await paidChat({\n    chat: options.chat,\n    request: chatRequest,\n    ledger: request.costLedger,\n    channel: 'analyst',\n    phase: request.costPhase,\n    actor: request.analystId,\n    ...(request.costTags ? { tags: request.costTags } : {}),\n    ...(options.pricing ? { pricing: options.pricing } : {}),\n    ...(request.signal ? { signal: request.signal } : {}),\n  })\n  if (!paid.succeeded) {\n    // The transport already spent its own attempts. A failure here ends the\n    // investigation: continuing would report an answer the model never gave.\n    throw new Error(\n      `chat trace engine ${turn.purpose} call failed for '${request.analystId}': ${paid.error.message}`,\n      { cause: paid.error },\n    )\n  }\n  return {\n    content: paid.response.content,\n    ...(paid.response.toolCalls ? { toolCalls: paid.response.toolCalls } : {}),\n    ...(paid.response.servedModel === undefined ? {} : { servedModel: paid.response.servedModel }),\n    ...(paid.response.finishReason === undefined\n      ? {}\n      : { finishReason: paid.response.finishReason }),\n  }\n}\n\ninterface ExecutedTool {\n  ok: boolean\n  truncated: boolean\n  payload: string\n}\n\nasync function executeTool(\n  call: LlmToolCall,\n  toolsByName: ReadonlyMap<string, TraceAnalysisToolDescriptor>,\n  request: TraceAnalysisEngineRequest,\n): Promise<ExecutedTool> {\n  const descriptor = toolsByName.get(call.name)\n  if (!descriptor) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `unknown tool '${call.name}'; available tools are ${[...toolsByName.keys()].sort().join(', ')}`,\n      ),\n    }\n  }\n  let args: unknown\n  try {\n    args = call.argumentsJson.trim() === '' ? {} : JSON.parse(call.argumentsJson)\n  } catch (error) {\n    return {\n      ok: false,\n      truncated: false,\n      payload: toolError(\n        `arguments for '${call.name}' were not JSON: ${messageOf(error)}. Send valid JSON arguments.`,\n      ),\n    }\n  }\n  try {\n    const value = await descriptor.handler(\n      args,\n      request.signal ? { signal: request.signal } : undefined,\n    )\n    return truncateToolPayload(JSON.stringify(value ?? null), request.limits.maxOutputChars)\n  } catch (error) {\n    // A bounded store refuses an oversized or malformed read by design. That\n    // refusal is the model's next instruction, not the run's failure.\n    if (request.signal?.aborted) throw error\n    request.log?.('trace tool failed', {\n      engine: ENGINE_ID,\n      analyst_id: request.analystId,\n      tool: call.name,\n      reason: messageOf(error),\n    })\n    return { ok: false, truncated: false, payload: toolError(messageOf(error)) }\n  }\n}\n\nfunction truncateToolPayload(payload: string, maxOutputChars: number): ExecutedTool {\n  if (payload.length <= maxOutputChars) return { ok: true, truncated: false, payload }\n  const keep = Math.max(0, maxOutputChars - TRUNCATION_MARKER.length)\n  return { ok: true, truncated: true, payload: `${payload.slice(0, keep)}${TRUNCATION_MARKER}` }\n}\n\nfunction toolMessage(call: LlmToolCall, content: string): LlmMessage {\n  return { role: 'tool', toolCallId: call.id, content }\n}\n\nfunction toolError(reason: string): string {\n  return JSON.stringify({ error: reason })\n}\n\nfunction exhaustedToolBudget(maxToolCalls: number): string {\n  return toolError(\n    `the trace-tool budget of ${maxToolCalls} calls is spent; answer from the evidence already read`,\n  )\n}\n\nfunction systemPrompt(request: TraceAnalysisEngineRequest, investigationTurns: number): string {\n  return [\n    request.instructions.trim(),\n    [\n      'HOW THIS INVESTIGATION RUNS:',\n      `- You have ${investigationTurns} investigation turns and at most ${request.limits.maxToolCalls} trace-tool calls.`,\n      '- Call the trace tools to read the store. Never state a trace fact you did not read.',\n      `- A tool result longer than ${request.limits.maxOutputChars} characters is truncated; narrow the query instead of asking again.`,\n      '- A tool result carrying an \"error\" field is feedback: fix the call or take another route.',\n      '- Answer with no tool call once you have the evidence. You are then asked for the final report.',\n    ].join('\\n'),\n  ].join('\\n\\n')\n}\n\nfunction reportPrompt(reportFormat: 'json-object' | 'json-schema'): string {\n  return [\n    'Report now. Do not call any more tools.',\n    'Return one JSON object with exactly two fields:',\n    '  \"answer\": a direct prose answer to the question.',\n    '  \"findings\": an array of findings in the schema above. Emit [] when there is nothing to report.',\n    reportFormat === 'json-object'\n      ? 'Return the object alone, with no surrounding prose and no code fence.'\n      : 'Return the object in the response schema you were given.',\n  ].join('\\n')\n}\n\nfunction renderTaskInputs(taskInputs: Readonly<Record<string, unknown>>): string {\n  // The engine contract forbids dropping structured inputs. This engine has no\n  // code environment to bind them as variables, so they are delivered whole as\n  // conversation material and the runtime record says which way they arrived.\n  return [\n    'TASK INPUTS — structured material delivered with the question, not fetched through tools:',\n    JSON.stringify(taskInputs, null, 2),\n  ].join('\\n')\n}\n\ninterface ParsedReport {\n  answer: string\n  findings: TraceAnalysisEngineResult['findings']\n  rejectedFindings: number\n}\n\nfunction parseReport(\n  content: string,\n  analystId: string,\n  onRejectedFinding: (index: number, reason: string) => void,\n): ParsedReport {\n  const envelope = coerceReportEnvelope(content, analystId)\n  if (typeof envelope.answer !== 'string' || !envelope.answer.trim()) {\n    throw new Error(`chat trace engine report for '${analystId}' carried no answer`)\n  }\n  // Findings are model output: one malformed row is model noise, not an engine\n  // fault, and the rest of the paid investigation must survive it.\n  const decoded = decodeRawFindingArray(envelope.findings)\n  if (decoded.topLevelError !== undefined) {\n    throw new Error(\n      `chat trace engine report for '${analystId}' had a malformed findings array: ${decoded.topLevelError}`,\n    )\n  }\n  for (const rejection of decoded.rejected) {\n    onRejectedFinding(\n      rejection.index,\n      `${rejection.code}${rejection.path ? ` at ${rejection.path}` : ''}: ${rejection.message}`,\n    )\n  }\n  return {\n    answer: envelope.answer,\n    findings: decoded.accepted,\n    rejectedFindings: decoded.rejected.length,\n  }\n}\n\nfunction coerceReportEnvelope(content: string, analystId: string): Record<string, unknown> {\n  let value: unknown\n  try {\n    value = JSON.parse(extractJsonPayload(content))\n  } catch (error) {\n    throw new Error(`chat trace engine report for '${analystId}' was not JSON: ${messageOf(error)}`)\n  }\n  if (typeof value !== 'object' || value === null || Array.isArray(value)) {\n    throw new Error(`chat trace engine report for '${analystId}' was not a JSON object`)\n  }\n  return value as Record<string, unknown>\n}\n\nfunction resolveModel(options: ChatTraceEngineOptions): string {\n  const model = options.model ?? options.chat.defaultModel\n  if (typeof model !== 'string' || !model.trim() || model !== model.trim()) {\n    throw new TypeError(\n      'chat trace engine needs a model: pass ChatTraceEngineOptions.model or bind defaultModel on the ChatClient',\n    )\n  }\n  return model\n}\n\nfunction messageOf(error: unknown): string {\n  return error instanceof Error ? error.message : String(error)\n}\n","/**\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 type { ChatClient } from './analyst/chat-client'\nimport { paidJsonChat } from './chat-json-call'\nimport {\n  CostLedger,\n  type CostLedgerHandle,\n  type CostReceipt,\n  type CustomTokenPricing,\n} from './cost-ledger'\nimport type { LlmCallRequest } 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  /** Caller-owned transport. Required: agent-eval executes no paid model. */\n  chat: ChatClient\n  /** Endpoint rates used when the transport reports no billed amount. */\n  pricing?: CustomTokenPricing\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: { maxPerFileChars: number; maxSourceChars: number; maxHtmlChars: number },\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 = {\n    chat: options.chat,\n    model: options.model ?? options.chat.defaultModel ?? 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    ...(options.pricing ? { pricing: options.pricing } : {}),\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 paidJsonChat<{ summary: string; concepts: ConceptFinding[] }>({\n      chat: opts.chat,\n      request,\n      ledger: opts.costLedger,\n      channel: 'judge',\n      phase: opts.costPhase,\n      actor: 'semantic-concept',\n      tags: opts.costTags,\n      signal: opts.signal,\n      ...(opts.pricing ? { pricing: opts.pricing } : {}),\n    })\n    receipt = paid.receipt\n    if (!paid.succeeded) throw paid.error\n    const { value } = paid\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":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuDA,MAAM,4BAA4B;AAClC,MAAM,YAAY;;AAGlB,MAAM,oBAAoB;;;;;;AAqC1B,SAAS,wBAA2E;CAClF,MAAM,MAAM,EAAE,aAAa,yBAAyB,EAAE,QAAQ,UAAU,CAAC;CAIzE,OAAO,IAAI;CACX,OAAO;EACL,MAAM;EACN,QAAQ;GACN,MAAM;GACN,sBAAsB;GACtB,UAAU,CAAC,UAAU,UAAU;GAC/B,YAAY;IACV,QAAQ;KAAE,MAAM;KAAU,aAAa;IAAuC;IAC9E,UAAU;KAAE,MAAM;KAAS,OAAO;IAAI;GACxC;EACF;CACF;AACF;AAEA,MAAM,qBAAqB,sBAAsB;;;;;AAMjD,SAAgB,sBAAsB,SAAsD;CAC1F,MAAM,QAAQ,aAAa,OAAO;CAClC,MAAM,kBAAkB,QAAQ,mBAAA;CAChC,IAAI,CAAC,OAAO,cAAc,eAAe,KAAK,mBAAmB,GAC/D,MAAM,IAAI,UAAU,mEAAmE;CAEzF,IACE,QAAQ,gBAAgB,KAAA,MACvB,CAAC,OAAO,SAAS,QAAQ,WAAW,KAAK,QAAQ,cAAc,IAEhE,MAAM,IAAI,UAAU,oEAAoE;CAE1F,IACE,QAAQ,qBAAqB,KAAA,MAC5B,CAAC,OAAO,cAAc,QAAQ,gBAAgB,KAAK,QAAQ,oBAAoB,IAEhF,MAAM,IAAI,UAAU,oEAAoE;CAE1F,MAAM,eAAe,QAAQ,gBAAgB;CAC7C,IAAI,iBAAiB,iBAAiB,iBAAiB,eACrD,MAAM,IAAI,UAAU,mEAAmE;CAGzF,OAAO;EACL,IAAI;EACJ,aACE;EACF;EACA,SAAS;EACT,iBAAiB;GACf,MAAM;GACN;GACA,WAAW,QAAQ,KAAK;GACxB,kBAAkB,QAAQ,KAAK,mBAAmB;GAClD,SAAS,QAAQ,UAAU,EAAE,GAAG,QAAQ,QAAQ,IAAI;GACpD,mBAAmB;GACnB,aAAa,QAAQ,eAAe;GACpC,UAAU,QAAQ,YAAY;GAC9B,oBAAoB,QAAQ,oBAAoB;GAChD,eAAe;GACf,oBAAoB,mBAAmB;EACzC;EACA,UAAU,YAAY,QAAQ,SAAS;GAAE,GAAG;GAAS;GAAO;GAAiB;EAAa,CAAC;CAC7F;AACF;AAQA,eAAe,QACb,SACA,SACoC;CAGpC,IAAI,QAAQ,OAAO,cAAc,GAC/B,MAAM,IAAI,MACR,yGAAyG,QAAQ,UAAU,aAAa,QAAQ,OAAO,YAAY,EACrK;CAEF,MAAM,cAAc,IAAI,IAAI,QAAQ,MAAM,KAAK,SAAS,CAAC,KAAK,MAAM,IAAI,CAAC,CAAC;CAC1E,IAAI,YAAY,SAAS,QAAQ,MAAM,QACrC,MAAM,IAAI,MAAM,wDAAwD,QAAQ,UAAU,EAAE;CAE9F,MAAM,kBAAuC,QAAQ,MAAM,KAAK,UAAU;EACxE,MAAM;EACN,UAAU;GAAE,MAAM,KAAK;GAAM,aAAa,KAAK;GAAa,YAAY,KAAK;EAAW;CAC1F,EAAE;CACF,MAAM,qBAAqB,KAAK,IAC9B,GACA,KAAK,IAAI,QAAQ,OAAO,eAAe,QAAQ,OAAO,cAAc,CAAC,CACvE;CAEA,MAAM,WAAyB,CAC7B;EAAE,MAAM;EAAU,SAAS,aAAa,SAAS,kBAAkB;CAAE,CACvE;CACA,IAAI,QAAQ,YACV,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,iBAAiB,QAAQ,UAAU;CAAE,CAAC;CAE/E,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,QAAQ;CAAS,CAAC;CAEzD,QAAQ,MAAM,gCAAgC;EAC5C,QAAQ;EACR,OAAO,QAAQ;EACf,WAAW,QAAQ,KAAK;EACxB,OAAO,QAAQ,MAAM,KAAK,SAAS,KAAK,IAAI;EAC5C,QAAQ,QAAQ;CAClB,CAAC;CAED,MAAM,aAAwB,CAAC;CAC/B,MAAM,+BAAe,IAAI,IAAY;CACrC,IAAI,aAAa;CACjB,IAAI,YAAY;CAChB,IAAI,uBAAuB;CAC3B,IAAI,sBAAsB;CAC1B,IAAI,YAAY;CAChB,IAAI,kBAAkB;CAEtB,KAAK,IAAI,OAAO,GAAG,OAAO,oBAAoB,QAAQ;EACpD,YAAY,OAAO;EACnB,MAAM,WAAW,MAAM,UAAU,SAAS,SAAS;GACjD;GACA,OAAO;GACP,YAAY;GACZ,SAAS;EACX,CAAC;EACD,cAAc;EACd,aAAa,IAAI,SAAS,eAAe,YAAY;EACrD,MAAM,YAAY,SAAS,aAAa,CAAC;EACzC,SAAS,KAAK;GACZ,MAAM;GACN,SAAS,SAAS;GAClB,GAAI,UAAU,SAAS,IAAI,EAAE,WAAW,UAAU,IAAI,CAAC;EACzD,CAAC;EACD,WAAW,KAAK;GACd,MAAM;GACN,OAAO;GACP,SAAS,SAAS;GAClB,YAAY,UAAU,KAAK,UAAU;IAAE,IAAI,KAAK;IAAI,MAAM,KAAK;GAAK,EAAE;GACtE,eAAe,SAAS,gBAAgB;EAC1C,CAAC;EACD,IAAI,UAAU,WAAW,GAAG;GAC1B,kBAAkB;GAClB;EACF;EACA,KAAK,MAAM,QAAQ,WAAW;GAC5B,IAAI,aAAa,QAAQ,OAAO,cAAc;IAC5C,sBAAsB;IAGtB,SAAS,KAAK,YAAY,MAAM,oBAAoB,QAAQ,OAAO,YAAY,CAAC,CAAC;IACjF;GACF;GACA,aAAa;GACb,MAAM,WAAW,MAAM,YAAY,MAAM,aAAa,OAAO;GAC7D,IAAI,SAAS,WAAW,wBAAwB;GAChD,SAAS,KAAK,YAAY,MAAM,SAAS,OAAO,CAAC;GACjD,WAAW,KAAK;IACd,MAAM;IACN,OAAO;IACP,MAAM,KAAK;IACX,IAAI,SAAS;IACb,WAAW,SAAS;IACpB,cAAc,SAAS,QAAQ;GACjC,CAAC;EACH;EACA,IAAI,qBAAqB;CAC3B;CAEA,SAAS,KAAK;EAAE,MAAM;EAAQ,SAAS,aAAa,QAAQ,YAAY;CAAE,CAAC;CAC3E,MAAM,SAAS,MAAM,UAAU,SAAS,SAAS;EAC/C;EACA,SAAS;EACT,MAAM,QAAQ;CAChB,CAAC;CACD,cAAc;CACd,aAAa,IAAI,OAAO,eAAe,YAAY;CACnD,WAAW,KAAK;EACd,MAAM,YAAY;EAClB,OAAO;EACP,SAAS,OAAO;EAChB,eAAe,OAAO,gBAAgB;CACxC,CAAC;CAED,MAAM,SAAS,YAAY,OAAO,SAAS,QAAQ,YAAY,OAAO,WAAW;EAC/E,QAAQ,MAAM,yDAAyD;GACrE,QAAQ;GACR;GACA;EACF,CAAC;CACH,CAAC;CAED,MAAM,SAAoC;EACxC,QAAQ,OAAO;EACf,UAAU,OAAO;EACjB;EACA;EACA;EACA,SAAS;GACP,QAAQ;GACR,OAAO,QAAQ;GACf,WAAW,QAAQ,KAAK;GACxB,eAAe,QAAQ;GACvB,qBAAqB;GACrB,iCAAiC;GACjC,uBAAuB;GACvB,wBAAwB;GACxB,mBAAmB,OAAO;GAG1B,eAAe,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;GACtC,aAAa,QAAQ,aAAa,qBAAqB;EACzD;CACF;CACA,QAAQ,MAAM,kCAAkC;EAC9C,QAAQ;EACR,aAAa,OAAO;EACpB,YAAY,OAAO;EACnB,UAAU,OAAO,SAAS;CAC5B,CAAC;CACD,OAAO;AACT;AASA,eAAe,UACb,SACA,SACA,MAOoB;CACpB,MAAM,cAA8B;EAClC,OAAO,QAAQ;EAEf,UAAU,KAAK,SAAS,KAAK,aAAa,EAAE,GAAG,QAAQ,EAAE;EACzD,WAAW,QAAQ;EACnB,GAAI,KAAK,SAAS,KAAK,MAAM,SAAS,IAClC;GAAE,OAAO,KAAK;GAAO,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EAAG,IACjF,CAAC;EACL,GAAI,KAAK,OAAO,EAAE,UAAU,KAAK,IAAI,CAAC;EACtC,GAAI,KAAK,SAAS,gBAAgB,EAAE,YAAY,mBAAmB,IAAI,CAAC;EACxE,GAAI,QAAQ,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,QAAQ,YAAY;EAChF,GAAI,QAAQ,aAAa,KAAA,IAAY,CAAC,IAAI,EAAE,UAAU,QAAQ,SAAS;EACvE,GAAI,QAAQ,qBAAqB,KAAA,IAAY,CAAC,IAAI,EAAE,WAAW,QAAQ,iBAAiB;CAC1F;CACA,MAAM,OAAO,MAAM,SAAS;EAC1B,MAAM,QAAQ;EACd,SAAS;EACT,QAAQ,QAAQ;EAChB,SAAS;EACT,OAAO,QAAQ;EACf,OAAO,QAAQ;EACf,GAAI,QAAQ,WAAW,EAAE,MAAM,QAAQ,SAAS,IAAI,CAAC;EACrD,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;CACrD,CAAC;CACD,IAAI,CAAC,KAAK,WAGR,MAAM,IAAI,MACR,qBAAqB,KAAK,QAAQ,oBAAoB,QAAQ,UAAU,KAAK,KAAK,MAAM,WACxF,EAAE,OAAO,KAAK,MAAM,CACtB;CAEF,OAAO;EACL,SAAS,KAAK,SAAS;EACvB,GAAI,KAAK,SAAS,YAAY,EAAE,WAAW,KAAK,SAAS,UAAU,IAAI,CAAC;EACxE,GAAI,KAAK,SAAS,gBAAgB,KAAA,IAAY,CAAC,IAAI,EAAE,aAAa,KAAK,SAAS,YAAY;EAC5F,GAAI,KAAK,SAAS,iBAAiB,KAAA,IAC/B,CAAC,IACD,EAAE,cAAc,KAAK,SAAS,aAAa;CACjD;AACF;AAQA,eAAe,YACb,MACA,aACA,SACuB;CACvB,MAAM,aAAa,YAAY,IAAI,KAAK,IAAI;CAC5C,IAAI,CAAC,YACH,OAAO;EACL,IAAI;EACJ,WAAW;EACX,SAAS,UACP,iBAAiB,KAAK,KAAK,yBAAyB,CAAC,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,KAAK,IAAI,GAC9F;CACF;CAEF,IAAI;CACJ,IAAI;EACF,OAAO,KAAK,cAAc,KAAK,MAAM,KAAK,CAAC,IAAI,KAAK,MAAM,KAAK,aAAa;CAC9E,SAAS,OAAO;EACd,OAAO;GACL,IAAI;GACJ,WAAW;GACX,SAAS,UACP,kBAAkB,KAAK,KAAK,mBAAmB,UAAU,KAAK,EAAE,6BAClE;EACF;CACF;CACA,IAAI;EACF,MAAM,QAAQ,MAAM,WAAW,QAC7B,MACA,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,KAAA,CAChD;EACA,OAAO,oBAAoB,KAAK,UAAU,SAAS,IAAI,GAAG,QAAQ,OAAO,cAAc;CACzF,SAAS,OAAO;EAGd,IAAI,QAAQ,QAAQ,SAAS,MAAM;EACnC,QAAQ,MAAM,qBAAqB;GACjC,QAAQ;GACR,YAAY,QAAQ;GACpB,MAAM,KAAK;GACX,QAAQ,UAAU,KAAK;EACzB,CAAC;EACD,OAAO;GAAE,IAAI;GAAO,WAAW;GAAO,SAAS,UAAU,UAAU,KAAK,CAAC;EAAE;CAC7E;AACF;AAEA,SAAS,oBAAoB,SAAiB,gBAAsC;CAClF,IAAI,QAAQ,UAAU,gBAAgB,OAAO;EAAE,IAAI;EAAM,WAAW;EAAO;CAAQ;CACnF,MAAM,OAAO,KAAK,IAAI,GAAG,iBAAiB,EAAwB;CAClE,OAAO;EAAE,IAAI;EAAM,WAAW;EAAM,SAAS,GAAG,QAAQ,MAAM,GAAG,IAAI,IAAI;CAAoB;AAC/F;AAEA,SAAS,YAAY,MAAmB,SAA6B;CACnE,OAAO;EAAE,MAAM;EAAQ,YAAY,KAAK;EAAI;CAAQ;AACtD;AAEA,SAAS,UAAU,QAAwB;CACzC,OAAO,KAAK,UAAU,EAAE,OAAO,OAAO,CAAC;AACzC;AAEA,SAAS,oBAAoB,cAA8B;CACzD,OAAO,UACL,4BAA4B,aAAa,uDAC3C;AACF;AAEA,SAAS,aAAa,SAAqC,oBAAoC;CAC7F,OAAO,CACL,QAAQ,aAAa,KAAK,GAC1B;EACE;EACA,cAAc,mBAAmB,mCAAmC,QAAQ,OAAO,aAAa;EAChG;EACA,+BAA+B,QAAQ,OAAO,eAAe;EAC7D;EACA;CACF,CAAC,CAAC,KAAK,IAAI,CACb,CAAC,CAAC,KAAK,MAAM;AACf;AAEA,SAAS,aAAa,cAAqD;CACzE,OAAO;EACL;EACA;EACA;EACA;EACA,iBAAiB,gBACb,0EACA;CACN,CAAC,CAAC,KAAK,IAAI;AACb;AAEA,SAAS,iBAAiB,YAAuD;CAI/E,OAAO,CACL,6FACA,KAAK,UAAU,YAAY,MAAM,CAAC,CACpC,CAAC,CAAC,KAAK,IAAI;AACb;AAQA,SAAS,YACP,SACA,WACA,mBACc;CACd,MAAM,WAAW,qBAAqB,SAAS,SAAS;CACxD,IAAI,OAAO,SAAS,WAAW,YAAY,CAAC,SAAS,OAAO,KAAK,GAC/D,MAAM,IAAI,MAAM,iCAAiC,UAAU,oBAAoB;CAIjF,MAAM,UAAU,sBAAsB,SAAS,QAAQ;CACvD,IAAI,QAAQ,kBAAkB,KAAA,GAC5B,MAAM,IAAI,MACR,iCAAiC,UAAU,oCAAoC,QAAQ,eACzF;CAEF,KAAK,MAAM,aAAa,QAAQ,UAC9B,kBACE,UAAU,OACV,GAAG,UAAU,OAAO,UAAU,OAAO,OAAO,UAAU,SAAS,GAAG,IAAI,UAAU,SAClF;CAEF,OAAO;EACL,QAAQ,SAAS;EACjB,UAAU,QAAQ;EAClB,kBAAkB,QAAQ,SAAS;CACrC;AACF;AAEA,SAAS,qBAAqB,SAAiB,WAA4C;CACzF,IAAI;CACJ,IAAI;EACF,QAAQ,KAAK,MAAM,mBAAmB,OAAO,CAAC;CAChD,SAAS,OAAO;EACd,MAAM,IAAI,MAAM,iCAAiC,UAAU,kBAAkB,UAAU,KAAK,GAAG;CACjG;CACA,IAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,MAAM,QAAQ,KAAK,GACpE,MAAM,IAAI,MAAM,iCAAiC,UAAU,wBAAwB;CAErF,OAAO;AACT;AAEA,SAAS,aAAa,SAAyC;CAC7D,MAAM,QAAQ,QAAQ,SAAS,QAAQ,KAAK;CAC5C,IAAI,OAAO,UAAU,YAAY,CAAC,MAAM,KAAK,KAAK,UAAU,MAAM,KAAK,GACrE,MAAM,IAAI,UACR,2GACF;CAEF,OAAO;AACT;AAEA,SAAS,UAAU,OAAwB;CACzC,OAAO,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;AAC9D;;;;;;;;;;;;;AChiBA,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;AAmCA,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,SACqC;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,OAAO;EACX,MAAM,QAAQ;EACd,OAAO,QAAQ,SAAS,QAAQ,KAAK,gBAAgB;EACrD,WAAW,QAAQ,aAAa;EAChC,WAAW,QAAQ,aAAa;EAChC,gBAAgB,QAAQ,kBAAkB;EAC1C,iBAAiB,QAAQ,mBAAmB;EAC5C,cAAc,QAAQ,gBAAgB;EACtC,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;EACtD,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,aAA8D;GAC/E,MAAM,KAAK;GACX;GACA,QAAQ,KAAK;GACb,SAAS;GACT,OAAO,KAAK;GACZ,OAAO;GACP,MAAM,KAAK;GACX,QAAQ,KAAK;GACb,GAAI,KAAK,UAAU,EAAE,SAAS,KAAK,QAAQ,IAAI,CAAC;EAClD,CAAC;EACD,UAAU,KAAK;EACf,IAAI,CAAC,KAAK,WAAW,MAAM,KAAK;EAChC,MAAM,EAAE,UAAU;EAElB,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/{sequential-DWn_np6P.js → sequential-7uxpO0Vm.js}

@@ -1,7 +1,7 @@
 import { h as manifestContentDigest } from "./agent-profile-cell-DcFW4vOc.js";
 import { t as mulberry32 } from "./random-Dn5fPWkt.js";
 import { t as eProcess } from "./sequential-eprocess-D1jKoihe.js";
-import { n as aggregatePairedHoldout, o as pairHoldout } from "./statistical-heldout-CNLOvRq_.js";
+import { n as aggregatePairedHoldout, o as pairHoldout } from "./statistical-heldout-C26hOFWD.js";
 //#region src/campaign/gates/sequential.ts
 /**
 * Anytime-valid sequential promotion gate — an e-process (betting
@@ -338,4 +338,4 @@ function sequentialDecide(options = {}) {
 //#endregion
 export { sequentialPairedGate as n, sequentialDecide as t };
 
-//# sourceMappingURL=sequential-DWn_np6P.js.map
+//# sourceMappingURL=sequential-7uxpO0Vm.js.map

package/dist/{sequential-DWn_np6P.js.map → sequential-7uxpO0Vm.js.map}

@@ -1 +1 @@
-{"version":3,"file":"sequential-DWn_np6P.js","names":[],"sources":["../src/campaign/gates/sequential.ts"],"sourcesContent":["/**\n * Anytime-valid sequential promotion gate — an e-process (betting\n * test-martingale, see `eProcess` in `statistics.ts`) over paired\n * per-scenario deltas, so a campaign stops the MOMENT evidence decides\n * instead of burning a fixed-n budget. Decisions remain valid at any\n * data-dependent stopping time (Ville's inequality), which is exactly what\n * fixed-n machinery cannot offer: peeking at a bootstrap CI after every\n * observation and stopping on the first significant peek inflates type-I\n * error far beyond alpha.\n *\n * REPLACES, never layers on, a fixed-n gate. Running `heldoutSignificance`\n * or `paretoSignificanceGate` repeatedly on a growing sample and stopping\n * early is optional stopping no matter how it is dressed up; this gate is\n * the valid way to stop early. Use one or the other per evidence stream.\n *\n * Pre-registration binding: anytime validity holds only for the\n * PRE-REGISTERED statistic. When a `SignedManifest` is bound, the gate takes\n * alpha from `manifest.alpha`, the observation budget from\n * `manifest.preRegisteredN`, orients deltas by `manifest.direction`, and\n * shifts the null boundary by `manifest.minEffect` — re-deciding the same\n * stream under different parameters after seeing data would reopen optional\n * stopping under a fancier name. The manifest's content hash is verified at\n * construction with the canonical RFC 8785 digest used by signManifest.\n *\n * The supermartingale guarantee requires E[delta_t | past] ≤ the null\n * boundary under H0. Independent sampling with that mean bound suffices.\n * Exchangeability or shuffling alone does not: one fair sign repeated many\n * times is exchangeable and has marginal mean zero, but its conditional mean\n * becomes the revealed sign. A seeded shuffle cannot remove that dependence.\n * `decide(ctx)` pairs cells, then averages repetitions within each scenario\n * before testing. Pass independentUnitByScenarioId for scenarios sharing a\n * latent source or incident. Units must still satisfy the conditional-mean null.\n * Direct `observe()` callers own this sampling and aggregation contract.\n */\n\nimport { manifestContentDigest, type SignedManifest } from '../../pre-registration'\nimport { type EProcessState, eProcess, mulberry32 } from '../../statistics'\nimport type { Gate, GateContext, GateResult, GenerationRecord, Scenario } from '../types'\nimport { aggregatePairedHoldout, pairHoldout } from './statistical-heldout'\n\nexport type SequentialDecision = 'promote' | 'continue' | 'undecided-at-maxN'\n\nexport interface SequentialObservation {\n  decision: SequentialDecision\n  /** Current e-value (the betting wealth) against H0. */\n  eValue: number\n  /** Paired deltas consumed so far. */\n  n: number\n  /** Names the decision basis. For 'undecided-at-maxN' it states explicitly\n   *  that exhausting the budget is NOT evidence of no effect. */\n  reason: string\n}\n\nexport interface SequentialPairedGateOptions {\n  /** Type-I budget. With `preRegistration` bound this MUST match\n   *  `manifest.alpha` (conflict throws). Default 0.05. */\n  alpha?: number\n  /** Minimum independent paired units before a promote may fire. The stopping rule is\n   *  \"first n ≥ minN with e-value ≥ 1/alpha\" — still a valid stopping time.\n   *  Default 5. */\n  minN?: number\n  /** Pre-registered independent-unit budget (scenarios by default in decide()). Required unless `preRegistration`\n   *  supplies it via `preRegisteredN` (conflict throws). */\n  maxN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |delta| in the judge's native scale; deltas are mapped to\n   *  x = (d/scale + 1)/2 ∈ [0,1]. A delta outside ±scale throws (use\n   *  `detectScale` to pick 1 vs 100 BEFORE streaming). Default 1. */\n  scale?: number\n  /** Seed for reproducible ordering of scenario deltas in decide().\n   *  Shuffling does not establish the conditional-mean null. Default 1337. */\n  shuffleSeed?: number\n  /** Independent unit for every scenario passed to decide(). Full cell pairs\n   *  are averaged within each unit before testing. Defaults to scenario IDs.\n   *  The mapping is copied at construction. Direct observe() is unchanged. */\n  independentUnitByScenarioId?: ReadonlyMap<string, string>\n  /** Bind the pre-registered hypothesis. Verified (content hash) at\n   *  construction; alpha/maxN/direction/minEffect come FROM the manifest. */\n  preRegistration?: SignedManifest\n  /** Override the gate name in reports. */\n  name?: string\n  /** Continue the observe-stream from a `state()` snapshot taken before a\n   *  restart. The gate's parameters still resolve from this options object\n   *  (and the manifest); the snapshot must have been recorded under the same\n   *  alpha, maxBet, and null boundary, and its decision must be one this\n   *  configuration could have reached at its n, or construction throws.\n   *  `decide(ctx)` is unaffected — it always runs on a fresh stream. */\n  resume?: SequentialStreamState\n}\n\n/** Snapshot of one observe-stream: the e-process state plus the gate's own\n *  decision, which applies `minN` and `maxN` on top of the core latch. A\n *  gate rebuilt with `resume` from this value continues exactly where the\n *  snapshot was taken. */\nexport type SequentialStreamState = EProcessState & { decision: SequentialDecision }\n\nexport interface SequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>\n  extends Gate<TArtifact, TScenario> {\n  /** Streaming entry point: feed one paired independent-unit delta\n   *  (candidate − baseline, native scale). Aggregate correlated repetitions\n   *  before calling; the caller must justify E[delta_t | past] under H0.\n   *  Each gate instance carries ONE\n   *  observe-stream; `decide(ctx)` runs on its own fresh stream and never\n   *  consumes or advances this one. 'promote' is sticky; observing past the\n   *  pre-registered maxN throws (extending a finished stream after seeing\n   *  the result reopens optional stopping — start a NEW pre-registered\n   *  test). */\n  observe(delta: number): SequentialObservation\n  /** Read-only snapshot of the observe-stream. Pass it back as\n   *  `resume` to continue the stream after a restart. */\n  state(): SequentialStreamState\n}\n\ninterface ResolvedConfig {\n  alpha: number\n  minN: number\n  maxN: number\n  maxBet: number\n  scale: number\n  shuffleSeed: number\n  /** 'decrease' negates deltas so \"better\" is always positive. */\n  direction: 'increase' | 'decrease'\n  /** H0 boundary in x-space: 1/2 + minEffect/(2·scale). */\n  nullMean: number\n  minEffect: number\n}\n\n/** Sync twin of `verifyManifest`, so gate construction stays synchronous and\n *  fails loud before any observation is consumed. The digest scheme belongs to\n *  `pre-registration`; this gate only compares. */\nfunction verifyManifestSync(m: SignedManifest): boolean {\n  return manifestContentDigest(m) === m.contentHash\n}\n\nfunction resolveConfig(opts: SequentialPairedGateOptions): ResolvedConfig {\n  const m = opts.preRegistration\n  let alpha: number\n  let maxN: number\n  let direction: 'increase' | 'decrease'\n  let minEffect: number\n  if (m) {\n    if (!verifyManifestSync(m)) {\n      throw new Error(\n        `sequentialPairedGate: pre-registration manifest '${m.id}' content hash mismatch (tampered)`,\n      )\n    }\n    if (opts.alpha !== undefined && opts.alpha !== m.alpha) {\n      throw new Error(\n        `sequentialPairedGate: alpha ${opts.alpha} conflicts with pre-registered alpha ${m.alpha} — ` +\n          `the registered statistic is the only one anytime validity covers`,\n      )\n    }\n    if (opts.maxN !== undefined && opts.maxN !== m.preRegisteredN) {\n      throw new Error(\n        `sequentialPairedGate: maxN ${opts.maxN} conflicts with pre-registered N ${m.preRegisteredN}`,\n      )\n    }\n    alpha = m.alpha\n    maxN = m.preRegisteredN\n    direction = m.direction\n    minEffect = m.minEffect\n  } else {\n    alpha = opts.alpha ?? 0.05\n    if (opts.maxN === undefined) {\n      throw new Error(\n        'sequentialPairedGate: maxN is required (or bind a preRegistration manifest whose ' +\n          'preRegisteredN is the budget) — an unbounded stream has no pre-registered budget',\n      )\n    }\n    maxN = opts.maxN\n    direction = 'increase'\n    minEffect = 0\n  }\n  const scale = opts.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialPairedGate: scale must be > 0, got ${scale}`)\n  }\n  if (!Number.isInteger(maxN) || maxN < 1) {\n    throw new Error(`sequentialPairedGate: maxN must be a positive integer, got ${maxN}`)\n  }\n  const minN = opts.minN ?? 5\n  if (!Number.isInteger(minN) || minN < 1 || minN > maxN) {\n    throw new Error(\n      `sequentialPairedGate: minN must be an integer in [1, maxN=${maxN}], got ${minN}`,\n    )\n  }\n  if (!Number.isFinite(minEffect) || minEffect < 0 || minEffect >= scale) {\n    throw new Error(\n      `sequentialPairedGate: minEffect must be in [0, scale=${scale}), got ${minEffect}`,\n    )\n  }\n  // H0: mean delta ≤ minEffect. In x = (d/scale + 1)/2 space that is\n  // E[x] ≤ 1/2 + minEffect/(2·scale) — a shifted null boundary, NOT a clamp\n  // of the observations (clamping is mean-distorting for skewed deltas).\n  const nullMean = 0.5 + minEffect / (2 * scale)\n  return {\n    alpha,\n    minN,\n    maxN,\n    maxBet: opts.maxBet ?? 0.5,\n    scale,\n    shuffleSeed: opts.shuffleSeed ?? 1337,\n    direction,\n    nullMean,\n    minEffect,\n  }\n}\n\ninterface SequentialStream {\n  observe(delta: number): SequentialObservation\n  state(): SequentialStreamState\n}\n\nconst SEQUENTIAL_DECISIONS: readonly SequentialDecision[] = [\n  'promote',\n  'continue',\n  'undecided-at-maxN',\n]\n\n/** Refuse a snapshot whose gate decision this configuration could not have\n *  reached at the snapshot's n. The e-process parameters are checked by\n *  `eProcess` itself; this covers the gate's own stopping rule. */\nfunction assertResumableStream(resume: SequentialStreamState, cfg: ResolvedConfig): void {\n  const fail = (detail: string): never => {\n    throw new Error(`sequentialPairedGate: cannot resume — ${detail}`)\n  }\n  if (!SEQUENTIAL_DECISIONS.includes(resume.decision)) {\n    fail(`unknown decision '${String(resume.decision)}'`)\n  }\n  if (resume.n > cfg.maxN) {\n    fail(`snapshot n=${resume.n} exceeds the pre-registered maxN=${cfg.maxN}`)\n  }\n  const threshold = 1 / cfg.alpha\n  switch (resume.decision) {\n    case 'promote':\n      if (resume.n < cfg.minN) {\n        fail(`decision 'promote' at n=${resume.n} is below minN=${cfg.minN}`)\n      }\n      break\n    case 'undecided-at-maxN':\n      if (resume.n !== cfg.maxN) {\n        fail(`decision 'undecided-at-maxN' at n=${resume.n} does not match maxN=${cfg.maxN}`)\n      }\n      break\n    case 'continue':\n      if (resume.n >= cfg.maxN) {\n        fail(`decision 'continue' at n=${resume.n} with maxN=${cfg.maxN}; the stream is finished`)\n      }\n      if (resume.n >= cfg.minN && resume.wealth >= threshold) {\n        fail(\n          `decision 'continue' at n=${resume.n} ≥ minN=${cfg.minN} with e-value ` +\n            `${resume.wealth} ≥ 1/α=${threshold}; this stream would have promoted`,\n        )\n      }\n      break\n  }\n}\n\nfunction makeStream(cfg: ResolvedConfig, resume?: SequentialStreamState): SequentialStream {\n  if (resume !== undefined) assertResumableStream(resume, cfg)\n  const proc = eProcess({\n    alpha: cfg.alpha,\n    maxBet: cfg.maxBet,\n    nullMean: cfg.nullMean,\n    ...(resume === undefined ? {} : { resume }),\n  })\n  const threshold = 1 / cfg.alpha\n  let terminal: SequentialDecision | undefined =\n    resume === undefined || resume.decision === 'continue' ? undefined : resume.decision\n  return {\n    observe(delta: number): SequentialObservation {\n      if (terminal === 'undecided-at-maxN') {\n        throw new Error(\n          `sequentialPairedGate: pre-registered maxN=${cfg.maxN} exhausted — extending the ` +\n            'stream after seeing the result reopens optional stopping; start a NEW ' +\n            'pre-registered test',\n        )\n      }\n      if (!Number.isFinite(delta) || Math.abs(delta) > cfg.scale) {\n        throw new Error(\n          `sequentialPairedGate: delta ${delta} outside ±scale=${cfg.scale} — pass the judge's ` +\n            `native scale explicitly (detectScale helps pick 1 vs 100)`,\n        )\n      }\n      const d = cfg.direction === 'decrease' ? -delta : delta\n      const step = proc.update((d / cfg.scale + 1) / 2)\n      // Stopping rule: FIRST n ≥ minN with current wealth ≥ 1/alpha — a valid\n      // stopping time (measurable w.r.t. the past), sticky once latched. The\n      // core's own `decided` latch ignores minN, so the gate re-derives the\n      // decision from current wealth here.\n      if (terminal === undefined && step.n >= cfg.minN && step.wealth >= threshold) {\n        terminal = 'promote'\n      }\n      if (terminal === 'promote') {\n        return {\n          decision: 'promote',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} at n=${step.n} ` +\n            `(minN=${cfg.minN}): the paired improvement exceeds ${cfg.minEffect} at anytime-valid ` +\n            `level α=${cfg.alpha}`,\n        }\n      }\n      if (step.n >= cfg.maxN) {\n        terminal = 'undecided-at-maxN'\n        return {\n          decision: 'undecided-at-maxN',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `undecided at pre-registered maxN=${cfg.maxN} (e-value ${step.wealth.toFixed(2)} < ` +\n            `1/α=${threshold.toFixed(2)}). This is NOT evidence of no effect — the effect may be ` +\n            'real but smaller than this budget can detect; re-register with a larger N to test that',\n        }\n      }\n      return {\n        decision: 'continue',\n        eValue: step.wealth,\n        n: step.n,\n        reason: `e-value ${step.wealth.toFixed(2)} < 1/α=${threshold.toFixed(2)} at n=${step.n}/${cfg.maxN} — keep observing`,\n      }\n    },\n    state() {\n      return { ...proc.state(), decision: terminal ?? 'continue' }\n    },\n  }\n}\n\n/** Data-independent in-place Fisher–Yates with a seeded PRNG — the permutation\n *  depends only on the seed, never the values, so bet predictability survives. */\nfunction seededShuffle<T>(items: T[], seed: number): T[] {\n  const rng = mulberry32(seed)\n  for (let i = items.length - 1; i > 0; i--) {\n    const j = Math.floor(rng() * (i + 1))\n    const tmp = items[i]!\n    items[i] = items[j]!\n    items[j] = tmp\n  }\n  return items\n}\n\n/**\n * Anytime-valid sequential paired gate. Conforms to the existing `Gate`\n * contract: decide(ctx) pairs candidate and baseline by full cellId, then\n * averages deltas within each configured unit (scenario by default). The\n * e-process consumes one observation per unit, with equal unit weights.\n * The streaming observe(delta) entry consumes caller-aggregated independent\n * units. Repetitions improve a unit's precision and do not increase n.\n *\n * Decision mapping onto the substrate's five-valued `GateDecision`:\n *   - 'promote'            → 'ship'\n *   - 'continue'           → 'need_more_work' (stream ended before maxN with\n *                            the e-value undecided — more independent units needed)\n *   - 'undecided-at-maxN'  → 'hold', with the reason stating it is NOT\n *                            evidence of no effect (never a silent default)\n */\nexport function sequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: SequentialPairedGateOptions,\n): SequentialPairedGate<TArtifact, TScenario> {\n  const cfg = resolveConfig(options)\n  const name = options.name ?? 'sequentialPairedGate'\n  const manifest = options.preRegistration\n  const observeStream = makeStream(cfg, options.resume)\n  const independentUnitByScenarioId =\n    options.independentUnitByScenarioId === undefined\n      ? undefined\n      : new Map(options.independentUnitByScenarioId)\n\n  return {\n    name,\n    observe(delta: number): SequentialObservation {\n      return observeStream.observe(delta)\n    },\n    state() {\n      return observeStream.state()\n    },\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      if (!ctx.baselineJudgeScores) {\n        throw new Error(\n          `${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own ` +\n            'scores would compare the candidate against itself (delta 0, silent no-op)',\n        )\n      }\n      const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n      const unitMap = independentUnitByScenarioId ?? new Map([...scenarioIds].map((id) => [id, id]))\n      for (const id of scenarioIds) {\n        const unit = unitMap.get(id)\n        if (typeof unit !== 'string' || unit.length === 0 || unit.trim() !== unit) {\n          throw new Error(`${name}: missing independent unit for scenario '${id}'`)\n        }\n      }\n      const paired = pairHoldout(\n        ctx.judgeScores,\n        ctx.baselineJudgeScores,\n        scenarioIds,\n        (s) => s.composite,\n      )\n      for (let i = 0; i < paired.cellIds.length; i++) {\n        const delta = paired.after[i]! - paired.before[i]!\n        if (Math.abs(delta) > cfg.scale) {\n          throw new Error(\n            `${name}: cell '${paired.cellIds[i]}' delta ${delta} outside ±scale=${cfg.scale}`,\n          )\n        }\n      }\n      const observations = aggregatePairedHoldout(paired, unitMap)\n      const deltas = observations.after.map((after, i) => after - observations.before[i]!)\n      seededShuffle(deltas, cfg.shuffleSeed)\n\n      const stream = makeStream(cfg)\n      let last: SequentialObservation | undefined\n      for (const d of deltas) {\n        last = stream.observe(d)\n        if (last.decision !== 'continue') break\n      }\n\n      const detail = {\n        ...stream.state(),\n        minN: cfg.minN,\n        maxN: cfg.maxN,\n        scale: cfg.scale,\n        shuffleSeed: cfg.shuffleSeed,\n        direction: cfg.direction,\n        minEffect: cfg.minEffect,\n        pairedN: deltas.length,\n        pairedCellN: paired.cellIds.length,\n        observationUnit: independentUnitByScenarioId === undefined ? 'scenario' : 'registered',\n        unitIds: observations.unitIds,\n        ...(manifest ? { preRegisteredId: manifest.id, metric: manifest.metric } : {}),\n      }\n      const meanDelta =\n        deltas.length === 0 ? undefined : deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n      if (last === undefined) {\n        return {\n          decision: 'need_more_work',\n          reasons: [`${name}: no paired holdout observations — nothing to test`],\n          contributingGates: [{ name, status: 'not_evaluated', detail }],\n        }\n      }\n      const decision =\n        last.decision === 'promote'\n          ? 'ship'\n          : last.decision === 'continue'\n            ? 'need_more_work'\n            : 'hold'\n      return {\n        decision,\n        reasons: [`${name}: ${last.reason}`],\n        contributingGates: [\n          {\n            name,\n            status:\n              decision === 'ship'\n                ? 'pass'\n                : decision === 'need_more_work'\n                  ? 'not_evaluated'\n                  : 'fail',\n            detail,\n          },\n        ],\n        delta: meanDelta,\n      }\n    },\n  }\n}\n\nexport interface SequentialDecideOptions {\n  /** Type-I budget for the early-stop evidence. Default 0.05. */\n  alpha?: number\n  /** Minimum paired deltas before a stop may fire. Default 5. */\n  minN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |per-scenario composite delta|. Default 1. */\n  scale?: number\n}\n\nexport interface SequentialDecideFn {\n  (args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n  /** Read-only snapshot of the accumulated e-process (observability + tests). */\n  state(): EProcessState\n}\n\n/**\n * SurfaceProposer.decide adapter that stops exploration at an e-value threshold.\n * The selected candidate still requires an independent held-out decision.\n *\n * Stream: for each generation g ≥ 1, the per-scenario composite deltas of\n * generation g's top candidate vs the generation-0 top candidate (the\n * incumbent the loop set out to beat), paired by scenarioId. H0: no proposed\n * surface improves any scenario's expected composite over the incumbent —\n * under it every delta has conditional mean ≤ 0 and the e-process is valid.\n * Once wealth ≥ 1/alpha the loop stops and hands the winner to the promotion\n * gate (which re-scores on HELD-OUT data — this adapter only spends the\n * exploration budget, it never promotes).\n *\n * This is an exploration stopping heuristic. Selection of the best measured\n * candidate and reuse of measured incumbent scores generally violate the\n * conditional-mean null, even when the candidates have no true improvement.\n * More repetitions reduce noise but do not establish exact type-I control.\n * An UNDECIDED process never stops the loop — absence of a\n * crossing is NOT evidence of no effect, so the loop simply runs its normal\n * course. Calling the adapter repeatedly with a growing history consumes each\n * generation exactly once (re-feeding an already-seen record would double-count\n * evidence).\n */\nexport function sequentialDecide(options: SequentialDecideOptions = {}): SequentialDecideFn {\n  const alpha = options.alpha ?? 0.05\n  const minN = options.minN ?? 5\n  const scale = options.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialDecide: scale must be > 0, got ${scale}`)\n  }\n  const proc = eProcess({ alpha, maxBet: options.maxBet ?? 0.5, nullMean: 0.5 })\n  const threshold = 1 / alpha\n  let processedGenerations = 0\n  let reference: Map<string, number> | undefined\n  let stopped: { stop: true; reason: string } | undefined\n\n  const topCandidate = (record: GenerationRecord) => {\n    if (record.candidates.length === 0) {\n      throw new Error(\n        `sequentialDecide: generation ${record.generationIndex} has no candidates — ` +\n          'cannot extract a top candidate',\n      )\n    }\n    return record.candidates.reduce((best, candidate) => {\n      if (best.composite === null) return candidate\n      if (candidate.composite === null) return best\n      return candidate.composite > best.composite ? candidate : best\n    })\n  }\n\n  const decide: SequentialDecideFn = ({ history }) => {\n    if (stopped) return stopped\n    if (history.length === 0) return { stop: false }\n    if (reference === undefined) {\n      reference = new Map(\n        topCandidate(history[0]!).scenarios.map((s) => [s.scenarioId, s.composite]),\n      )\n    }\n    // Consume only generations not yet seen; start at 1 — generation 0 IS the\n    // reference and would contribute all-zero deltas (pure wealth dilution).\n    for (let g = Math.max(1, processedGenerations); g < history.length; g++) {\n      const top = topCandidate(history[g]!)\n      const byScenario = new Map(top.scenarios.map((s) => [s.scenarioId, s.composite]))\n      for (const [scenarioId, refComposite] of reference) {\n        const candComposite = byScenario.get(scenarioId)\n        if (candComposite === undefined) {\n          throw new Error(\n            `sequentialDecide: generation ${history[g]!.generationIndex} top candidate is missing ` +\n              `scenario '${scenarioId}' — generations must score the same scenario set to pair`,\n          )\n        }\n        const delta = candComposite - refComposite\n        if (!Number.isFinite(delta) || Math.abs(delta) > scale) {\n          throw new Error(\n            `sequentialDecide: paired delta ${delta} outside ±scale=${scale} on scenario ` +\n              `'${scenarioId}' — pass the composite scale explicitly`,\n          )\n        }\n        const step = proc.update((delta / scale + 1) / 2)\n        if (step.n >= minN && step.wealth >= threshold) {\n          stopped = {\n            stop: true,\n            reason:\n              `sequential e-process decided at generation ${history[g]!.generationIndex}: ` +\n              `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} ` +\n              'paired deltas vs the generation-0 incumbent; stop exploring and require ' +\n              'an independent held-out promotion decision',\n          }\n          processedGenerations = history.length\n          return stopped\n        }\n      }\n    }\n    processedGenerations = history.length\n    return { stop: false }\n  }\n  decide.state = () => proc.state()\n  return decide\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmIA,SAAS,mBAAmB,GAA4B;CACtD,OAAO,sBAAsB,CAAC,MAAM,EAAE;AACxC;AAEA,SAAS,cAAc,MAAmD;CACxE,MAAM,IAAI,KAAK;CACf,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI,GAAG;EACL,IAAI,CAAC,mBAAmB,CAAC,GACvB,MAAM,IAAI,MACR,oDAAoD,EAAE,GAAG,mCAC3D;EAEF,IAAI,KAAK,UAAU,KAAA,KAAa,KAAK,UAAU,EAAE,OAC/C,MAAM,IAAI,MACR,+BAA+B,KAAK,MAAM,uCAAuC,EAAE,MAAM,oEAE3F;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,KAAK,SAAS,EAAE,gBAC7C,MAAM,IAAI,MACR,8BAA8B,KAAK,KAAK,mCAAmC,EAAE,gBAC/E;EAEF,QAAQ,EAAE;EACV,OAAO,EAAE;EACT,YAAY,EAAE;EACd,YAAY,EAAE;CAChB,OAAO;EACL,QAAQ,KAAK,SAAS;EACtB,IAAI,KAAK,SAAS,KAAA,GAChB,MAAM,IAAI,MACR,mKAEF;EAEF,OAAO,KAAK;EACZ,YAAY;EACZ,YAAY;CACd;CACA,MAAM,QAAQ,KAAK,SAAS;CAC5B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,gDAAgD,OAAO;CAEzE,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,GACpC,MAAM,IAAI,MAAM,8DAA8D,MAAM;CAEtF,MAAM,OAAO,KAAK,QAAQ;CAC1B,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,MAChD,MAAM,IAAI,MACR,6DAA6D,KAAK,SAAS,MAC7E;CAEF,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,KAAK,aAAa,OAC/D,MAAM,IAAI,MACR,wDAAwD,MAAM,SAAS,WACzE;CAKF,MAAM,WAAW,KAAM,aAAa,IAAI;CACxC,OAAO;EACL;EACA;EACA;EACA,QAAQ,KAAK,UAAU;EACvB;EACA,aAAa,KAAK,eAAe;EACjC;EACA;EACA;CACF;AACF;AAOA,MAAM,uBAAsD;CAC1D;CACA;CACA;AACF;;;;AAKA,SAAS,sBAAsB,QAA+B,KAA2B;CACvF,MAAM,QAAQ,WAA0B;EACtC,MAAM,IAAI,MAAM,yCAAyC,QAAQ;CACnE;CACA,IAAI,CAAC,qBAAqB,SAAS,OAAO,QAAQ,GAChD,KAAK,qBAAqB,OAAO,OAAO,QAAQ,EAAE,EAAE;CAEtD,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,cAAc,OAAO,EAAE,mCAAmC,IAAI,MAAM;CAE3E,MAAM,YAAY,IAAI,IAAI;CAC1B,QAAQ,OAAO,UAAf;EACE,KAAK;GACH,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,2BAA2B,OAAO,EAAE,iBAAiB,IAAI,MAAM;GAEtE;EACF,KAAK;GACH,IAAI,OAAO,MAAM,IAAI,MACnB,KAAK,qCAAqC,OAAO,EAAE,uBAAuB,IAAI,MAAM;GAEtF;EACF,KAAK;GACH,IAAI,OAAO,KAAK,IAAI,MAClB,KAAK,4BAA4B,OAAO,EAAE,aAAa,IAAI,KAAK,yBAAyB;GAE3F,IAAI,OAAO,KAAK,IAAI,QAAQ,OAAO,UAAU,WAC3C,KACE,4BAA4B,OAAO,EAAE,UAAU,IAAI,KAAK,gBACnD,OAAO,OAAO,SAAS,UAAU,kCACxC;CAGN;AACF;AAEA,SAAS,WAAW,KAAqB,QAAkD;CACzF,IAAI,WAAW,KAAA,GAAW,sBAAsB,QAAQ,GAAG;CAC3D,MAAM,OAAO,SAAS;EACpB,OAAO,IAAI;EACX,QAAQ,IAAI;EACZ,UAAU,IAAI;EACd,GAAI,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO;CAC3C,CAAC;CACD,MAAM,YAAY,IAAI,IAAI;CAC1B,IAAI,WACF,WAAW,KAAA,KAAa,OAAO,aAAa,aAAa,KAAA,IAAY,OAAO;CAC9E,OAAO;EACL,QAAQ,OAAsC;GAC5C,IAAI,aAAa,qBACf,MAAM,IAAI,MACR,6CAA6C,IAAI,KAAK,qHAGxD;GAEF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,OACnD,MAAM,IAAI,MACR,+BAA+B,MAAM,kBAAkB,IAAI,MAAM,8EAEnE;GAEF,MAAM,IAAI,IAAI,cAAc,aAAa,CAAC,QAAQ;GAClD,MAAM,OAAO,KAAK,QAAQ,IAAI,IAAI,QAAQ,KAAK,CAAC;GAKhD,IAAI,aAAa,KAAA,KAAa,KAAK,KAAK,IAAI,QAAQ,KAAK,UAAU,WACjE,WAAW;GAEb,IAAI,aAAa,WACf,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QACE,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,SACtE,IAAI,KAAK,oCAAoC,IAAI,UAAU,4BACzD,IAAI;GACnB;GAEF,IAAI,KAAK,KAAK,IAAI,MAAM;IACtB,WAAW;IACX,OAAO;KACL,UAAU;KACV,QAAQ,KAAK;KACb,GAAG,KAAK;KACR,QACE,oCAAoC,IAAI,KAAK,YAAY,KAAK,OAAO,QAAQ,CAAC,EAAE,SACzE,UAAU,QAAQ,CAAC,EAAE;IAEhC;GACF;GACA,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QAAQ,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,GAAG,IAAI,KAAK;GACrG;EACF;EACA,QAAQ;GACN,OAAO;IAAE,GAAG,KAAK,MAAM;IAAG,UAAU,YAAY;GAAW;EAC7D;CACF;AACF;;;AAIA,SAAS,cAAiB,OAAY,MAAmB;CACvD,MAAM,MAAM,WAAW,IAAI;CAC3B,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,IAAI,GAAG,KAAK;EACzC,MAAM,IAAI,KAAK,MAAM,IAAI,KAAK,IAAI,EAAE;EACpC,MAAM,MAAM,MAAM;EAClB,MAAM,KAAK,MAAM;EACjB,MAAM,KAAK;CACb;CACA,OAAO;AACT;;;;;;;;;;;;;;;;AAiBA,SAAgB,qBACd,SAC4C;CAC5C,MAAM,MAAM,cAAc,OAAO;CACjC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,WAAW,QAAQ;CACzB,MAAM,gBAAgB,WAAW,KAAK,QAAQ,MAAM;CACpD,MAAM,8BACJ,QAAQ,gCAAgC,KAAA,IACpC,KAAA,IACA,IAAI,IAAI,QAAQ,2BAA2B;CAEjD,OAAO;EACL;EACA,QAAQ,OAAsC;GAC5C,OAAO,cAAc,QAAQ,KAAK;EACpC;EACA,QAAQ;GACN,OAAO,cAAc,MAAM;EAC7B;EACA,MAAM,OAAO,KAA6D;GACxE,IAAI,CAAC,IAAI,qBACP,MAAM,IAAI,MACR,GAAG,KAAK,sJAEV;GAEF,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;GAC1D,MAAM,UAAU,+BAA+B,IAAI,IAAI,CAAC,GAAG,WAAW,CAAC,CAAC,KAAK,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC;GAC7F,KAAK,MAAM,MAAM,aAAa;IAC5B,MAAM,OAAO,QAAQ,IAAI,EAAE;IAC3B,IAAI,OAAO,SAAS,YAAY,KAAK,WAAW,KAAK,KAAK,KAAK,MAAM,MACnE,MAAM,IAAI,MAAM,GAAG,KAAK,2CAA2C,GAAG,EAAE;GAE5E;GACA,MAAM,SAAS,YACb,IAAI,aACJ,IAAI,qBACJ,cACC,MAAM,EAAE,SACX;GACA,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,QAAQ,KAAK;IAC9C,MAAM,QAAQ,OAAO,MAAM,KAAM,OAAO,OAAO;IAC/C,IAAI,KAAK,IAAI,KAAK,IAAI,IAAI,OACxB,MAAM,IAAI,MACR,GAAG,KAAK,UAAU,OAAO,QAAQ,GAAG,UAAU,MAAM,kBAAkB,IAAI,OAC5E;GAEJ;GACA,MAAM,eAAe,uBAAuB,QAAQ,OAAO;GAC3D,MAAM,SAAS,aAAa,MAAM,KAAK,OAAO,MAAM,QAAQ,aAAa,OAAO,EAAG;GACnF,cAAc,QAAQ,IAAI,WAAW;GAErC,MAAM,SAAS,WAAW,GAAG;GAC7B,IAAI;GACJ,KAAK,MAAM,KAAK,QAAQ;IACtB,OAAO,OAAO,QAAQ,CAAC;IACvB,IAAI,KAAK,aAAa,YAAY;GACpC;GAEA,MAAM,SAAS;IACb,GAAG,OAAO,MAAM;IAChB,MAAM,IAAI;IACV,MAAM,IAAI;IACV,OAAO,IAAI;IACX,aAAa,IAAI;IACjB,WAAW,IAAI;IACf,WAAW,IAAI;IACf,SAAS,OAAO;IAChB,aAAa,OAAO,QAAQ;IAC5B,iBAAiB,gCAAgC,KAAA,IAAY,aAAa;IAC1E,SAAS,aAAa;IACtB,GAAI,WAAW;KAAE,iBAAiB,SAAS;KAAI,QAAQ,SAAS;IAAO,IAAI,CAAC;GAC9E;GACA,MAAM,YACJ,OAAO,WAAW,IAAI,KAAA,IAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;GAE/E,IAAI,SAAS,KAAA,GACX,OAAO;IACL,UAAU;IACV,SAAS,CAAC,GAAG,KAAK,mDAAmD;IACrE,mBAAmB,CAAC;KAAE;KAAM,QAAQ;KAAiB;IAAO,CAAC;GAC/D;GAEF,MAAM,WACJ,KAAK,aAAa,YACd,SACA,KAAK,aAAa,aAChB,mBACA;GACR,OAAO;IACL;IACA,SAAS,CAAC,GAAG,KAAK,IAAI,KAAK,QAAQ;IACnC,mBAAmB,CACjB;KACE;KACA,QACE,aAAa,SACT,SACA,aAAa,mBACX,kBACA;KACR;IACF,CACF;IACA,OAAO;GACT;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,iBAAiB,UAAmC,CAAC,GAAuB;CAC1F,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,4CAA4C,OAAO;CAErE,MAAM,OAAO,SAAS;EAAE;EAAO,QAAQ,QAAQ,UAAU;EAAK,UAAU;CAAI,CAAC;CAC7E,MAAM,YAAY,IAAI;CACtB,IAAI,uBAAuB;CAC3B,IAAI;CACJ,IAAI;CAEJ,MAAM,gBAAgB,WAA6B;EACjD,IAAI,OAAO,WAAW,WAAW,GAC/B,MAAM,IAAI,MACR,gCAAgC,OAAO,gBAAgB,oDAEzD;EAEF,OAAO,OAAO,WAAW,QAAQ,MAAM,cAAc;GACnD,IAAI,KAAK,cAAc,MAAM,OAAO;GACpC,IAAI,UAAU,cAAc,MAAM,OAAO;GACzC,OAAO,UAAU,YAAY,KAAK,YAAY,YAAY;EAC5D,CAAC;CACH;CAEA,MAAM,UAA8B,EAAE,cAAc;EAClD,IAAI,SAAS,OAAO;EACpB,IAAI,QAAQ,WAAW,GAAG,OAAO,EAAE,MAAM,MAAM;EAC/C,IAAI,cAAc,KAAA,GAChB,YAAY,IAAI,IACd,aAAa,QAAQ,EAAG,CAAC,CAAC,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAC5E;EAIF,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,oBAAoB,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvE,MAAM,MAAM,aAAa,QAAQ,EAAG;GACpC,MAAM,aAAa,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC;GAChF,KAAK,MAAM,CAAC,YAAY,iBAAiB,WAAW;IAClD,MAAM,gBAAgB,WAAW,IAAI,UAAU;IAC/C,IAAI,kBAAkB,KAAA,GACpB,MAAM,IAAI,MACR,gCAAgC,QAAQ,EAAE,CAAE,gBAAgB,sCAC7C,WAAW,yDAC5B;IAEF,MAAM,QAAQ,gBAAgB;IAC9B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,OAC/C,MAAM,IAAI,MACR,kCAAkC,MAAM,kBAAkB,MAAM,gBAC1D,WAAW,wCACnB;IAEF,MAAM,OAAO,KAAK,QAAQ,QAAQ,QAAQ,KAAK,CAAC;IAChD,IAAI,KAAK,KAAK,QAAQ,KAAK,UAAU,WAAW;KAC9C,UAAU;MACR,MAAM;MACN,QACE,8CAA8C,QAAQ,EAAE,CAAE,gBAAgB,YAC/D,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,WAAW,KAAK,EAAE;KAGtF;KACA,uBAAuB,QAAQ;KAC/B,OAAO;IACT;GACF;EACF;EACA,uBAAuB,QAAQ;EAC/B,OAAO,EAAE,MAAM,MAAM;CACvB;CACA,OAAO,cAAc,KAAK,MAAM;CAChC,OAAO;AACT"}
+{"version":3,"file":"sequential-7uxpO0Vm.js","names":[],"sources":["../src/campaign/gates/sequential.ts"],"sourcesContent":["/**\n * Anytime-valid sequential promotion gate — an e-process (betting\n * test-martingale, see `eProcess` in `statistics.ts`) over paired\n * per-scenario deltas, so a campaign stops the MOMENT evidence decides\n * instead of burning a fixed-n budget. Decisions remain valid at any\n * data-dependent stopping time (Ville's inequality), which is exactly what\n * fixed-n machinery cannot offer: peeking at a bootstrap CI after every\n * observation and stopping on the first significant peek inflates type-I\n * error far beyond alpha.\n *\n * REPLACES, never layers on, a fixed-n gate. Running `heldoutSignificance`\n * or `paretoSignificanceGate` repeatedly on a growing sample and stopping\n * early is optional stopping no matter how it is dressed up; this gate is\n * the valid way to stop early. Use one or the other per evidence stream.\n *\n * Pre-registration binding: anytime validity holds only for the\n * PRE-REGISTERED statistic. When a `SignedManifest` is bound, the gate takes\n * alpha from `manifest.alpha`, the observation budget from\n * `manifest.preRegisteredN`, orients deltas by `manifest.direction`, and\n * shifts the null boundary by `manifest.minEffect` — re-deciding the same\n * stream under different parameters after seeing data would reopen optional\n * stopping under a fancier name. The manifest's content hash is verified at\n * construction with the canonical RFC 8785 digest used by signManifest.\n *\n * The supermartingale guarantee requires E[delta_t | past] ≤ the null\n * boundary under H0. Independent sampling with that mean bound suffices.\n * Exchangeability or shuffling alone does not: one fair sign repeated many\n * times is exchangeable and has marginal mean zero, but its conditional mean\n * becomes the revealed sign. A seeded shuffle cannot remove that dependence.\n * `decide(ctx)` pairs cells, then averages repetitions within each scenario\n * before testing. Pass independentUnitByScenarioId for scenarios sharing a\n * latent source or incident. Units must still satisfy the conditional-mean null.\n * Direct `observe()` callers own this sampling and aggregation contract.\n */\n\nimport { manifestContentDigest, type SignedManifest } from '../../pre-registration'\nimport { type EProcessState, eProcess, mulberry32 } from '../../statistics'\nimport type { Gate, GateContext, GateResult, GenerationRecord, Scenario } from '../types'\nimport { aggregatePairedHoldout, pairHoldout } from './statistical-heldout'\n\nexport type SequentialDecision = 'promote' | 'continue' | 'undecided-at-maxN'\n\nexport interface SequentialObservation {\n  decision: SequentialDecision\n  /** Current e-value (the betting wealth) against H0. */\n  eValue: number\n  /** Paired deltas consumed so far. */\n  n: number\n  /** Names the decision basis. For 'undecided-at-maxN' it states explicitly\n   *  that exhausting the budget is NOT evidence of no effect. */\n  reason: string\n}\n\nexport interface SequentialPairedGateOptions {\n  /** Type-I budget. With `preRegistration` bound this MUST match\n   *  `manifest.alpha` (conflict throws). Default 0.05. */\n  alpha?: number\n  /** Minimum independent paired units before a promote may fire. The stopping rule is\n   *  \"first n ≥ minN with e-value ≥ 1/alpha\" — still a valid stopping time.\n   *  Default 5. */\n  minN?: number\n  /** Pre-registered independent-unit budget (scenarios by default in decide()). Required unless `preRegistration`\n   *  supplies it via `preRegisteredN` (conflict throws). */\n  maxN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |delta| in the judge's native scale; deltas are mapped to\n   *  x = (d/scale + 1)/2 ∈ [0,1]. A delta outside ±scale throws (use\n   *  `detectScale` to pick 1 vs 100 BEFORE streaming). Default 1. */\n  scale?: number\n  /** Seed for reproducible ordering of scenario deltas in decide().\n   *  Shuffling does not establish the conditional-mean null. Default 1337. */\n  shuffleSeed?: number\n  /** Independent unit for every scenario passed to decide(). Full cell pairs\n   *  are averaged within each unit before testing. Defaults to scenario IDs.\n   *  The mapping is copied at construction. Direct observe() is unchanged. */\n  independentUnitByScenarioId?: ReadonlyMap<string, string>\n  /** Bind the pre-registered hypothesis. Verified (content hash) at\n   *  construction; alpha/maxN/direction/minEffect come FROM the manifest. */\n  preRegistration?: SignedManifest\n  /** Override the gate name in reports. */\n  name?: string\n  /** Continue the observe-stream from a `state()` snapshot taken before a\n   *  restart. The gate's parameters still resolve from this options object\n   *  (and the manifest); the snapshot must have been recorded under the same\n   *  alpha, maxBet, and null boundary, and its decision must be one this\n   *  configuration could have reached at its n, or construction throws.\n   *  `decide(ctx)` is unaffected — it always runs on a fresh stream. */\n  resume?: SequentialStreamState\n}\n\n/** Snapshot of one observe-stream: the e-process state plus the gate's own\n *  decision, which applies `minN` and `maxN` on top of the core latch. A\n *  gate rebuilt with `resume` from this value continues exactly where the\n *  snapshot was taken. */\nexport type SequentialStreamState = EProcessState & { decision: SequentialDecision }\n\nexport interface SequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>\n  extends Gate<TArtifact, TScenario> {\n  /** Streaming entry point: feed one paired independent-unit delta\n   *  (candidate − baseline, native scale). Aggregate correlated repetitions\n   *  before calling; the caller must justify E[delta_t | past] under H0.\n   *  Each gate instance carries ONE\n   *  observe-stream; `decide(ctx)` runs on its own fresh stream and never\n   *  consumes or advances this one. 'promote' is sticky; observing past the\n   *  pre-registered maxN throws (extending a finished stream after seeing\n   *  the result reopens optional stopping — start a NEW pre-registered\n   *  test). */\n  observe(delta: number): SequentialObservation\n  /** Read-only snapshot of the observe-stream. Pass it back as\n   *  `resume` to continue the stream after a restart. */\n  state(): SequentialStreamState\n}\n\ninterface ResolvedConfig {\n  alpha: number\n  minN: number\n  maxN: number\n  maxBet: number\n  scale: number\n  shuffleSeed: number\n  /** 'decrease' negates deltas so \"better\" is always positive. */\n  direction: 'increase' | 'decrease'\n  /** H0 boundary in x-space: 1/2 + minEffect/(2·scale). */\n  nullMean: number\n  minEffect: number\n}\n\n/** Sync twin of `verifyManifest`, so gate construction stays synchronous and\n *  fails loud before any observation is consumed. The digest scheme belongs to\n *  `pre-registration`; this gate only compares. */\nfunction verifyManifestSync(m: SignedManifest): boolean {\n  return manifestContentDigest(m) === m.contentHash\n}\n\nfunction resolveConfig(opts: SequentialPairedGateOptions): ResolvedConfig {\n  const m = opts.preRegistration\n  let alpha: number\n  let maxN: number\n  let direction: 'increase' | 'decrease'\n  let minEffect: number\n  if (m) {\n    if (!verifyManifestSync(m)) {\n      throw new Error(\n        `sequentialPairedGate: pre-registration manifest '${m.id}' content hash mismatch (tampered)`,\n      )\n    }\n    if (opts.alpha !== undefined && opts.alpha !== m.alpha) {\n      throw new Error(\n        `sequentialPairedGate: alpha ${opts.alpha} conflicts with pre-registered alpha ${m.alpha} — ` +\n          `the registered statistic is the only one anytime validity covers`,\n      )\n    }\n    if (opts.maxN !== undefined && opts.maxN !== m.preRegisteredN) {\n      throw new Error(\n        `sequentialPairedGate: maxN ${opts.maxN} conflicts with pre-registered N ${m.preRegisteredN}`,\n      )\n    }\n    alpha = m.alpha\n    maxN = m.preRegisteredN\n    direction = m.direction\n    minEffect = m.minEffect\n  } else {\n    alpha = opts.alpha ?? 0.05\n    if (opts.maxN === undefined) {\n      throw new Error(\n        'sequentialPairedGate: maxN is required (or bind a preRegistration manifest whose ' +\n          'preRegisteredN is the budget) — an unbounded stream has no pre-registered budget',\n      )\n    }\n    maxN = opts.maxN\n    direction = 'increase'\n    minEffect = 0\n  }\n  const scale = opts.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialPairedGate: scale must be > 0, got ${scale}`)\n  }\n  if (!Number.isInteger(maxN) || maxN < 1) {\n    throw new Error(`sequentialPairedGate: maxN must be a positive integer, got ${maxN}`)\n  }\n  const minN = opts.minN ?? 5\n  if (!Number.isInteger(minN) || minN < 1 || minN > maxN) {\n    throw new Error(\n      `sequentialPairedGate: minN must be an integer in [1, maxN=${maxN}], got ${minN}`,\n    )\n  }\n  if (!Number.isFinite(minEffect) || minEffect < 0 || minEffect >= scale) {\n    throw new Error(\n      `sequentialPairedGate: minEffect must be in [0, scale=${scale}), got ${minEffect}`,\n    )\n  }\n  // H0: mean delta ≤ minEffect. In x = (d/scale + 1)/2 space that is\n  // E[x] ≤ 1/2 + minEffect/(2·scale) — a shifted null boundary, NOT a clamp\n  // of the observations (clamping is mean-distorting for skewed deltas).\n  const nullMean = 0.5 + minEffect / (2 * scale)\n  return {\n    alpha,\n    minN,\n    maxN,\n    maxBet: opts.maxBet ?? 0.5,\n    scale,\n    shuffleSeed: opts.shuffleSeed ?? 1337,\n    direction,\n    nullMean,\n    minEffect,\n  }\n}\n\ninterface SequentialStream {\n  observe(delta: number): SequentialObservation\n  state(): SequentialStreamState\n}\n\nconst SEQUENTIAL_DECISIONS: readonly SequentialDecision[] = [\n  'promote',\n  'continue',\n  'undecided-at-maxN',\n]\n\n/** Refuse a snapshot whose gate decision this configuration could not have\n *  reached at the snapshot's n. The e-process parameters are checked by\n *  `eProcess` itself; this covers the gate's own stopping rule. */\nfunction assertResumableStream(resume: SequentialStreamState, cfg: ResolvedConfig): void {\n  const fail = (detail: string): never => {\n    throw new Error(`sequentialPairedGate: cannot resume — ${detail}`)\n  }\n  if (!SEQUENTIAL_DECISIONS.includes(resume.decision)) {\n    fail(`unknown decision '${String(resume.decision)}'`)\n  }\n  if (resume.n > cfg.maxN) {\n    fail(`snapshot n=${resume.n} exceeds the pre-registered maxN=${cfg.maxN}`)\n  }\n  const threshold = 1 / cfg.alpha\n  switch (resume.decision) {\n    case 'promote':\n      if (resume.n < cfg.minN) {\n        fail(`decision 'promote' at n=${resume.n} is below minN=${cfg.minN}`)\n      }\n      break\n    case 'undecided-at-maxN':\n      if (resume.n !== cfg.maxN) {\n        fail(`decision 'undecided-at-maxN' at n=${resume.n} does not match maxN=${cfg.maxN}`)\n      }\n      break\n    case 'continue':\n      if (resume.n >= cfg.maxN) {\n        fail(`decision 'continue' at n=${resume.n} with maxN=${cfg.maxN}; the stream is finished`)\n      }\n      if (resume.n >= cfg.minN && resume.wealth >= threshold) {\n        fail(\n          `decision 'continue' at n=${resume.n} ≥ minN=${cfg.minN} with e-value ` +\n            `${resume.wealth} ≥ 1/α=${threshold}; this stream would have promoted`,\n        )\n      }\n      break\n  }\n}\n\nfunction makeStream(cfg: ResolvedConfig, resume?: SequentialStreamState): SequentialStream {\n  if (resume !== undefined) assertResumableStream(resume, cfg)\n  const proc = eProcess({\n    alpha: cfg.alpha,\n    maxBet: cfg.maxBet,\n    nullMean: cfg.nullMean,\n    ...(resume === undefined ? {} : { resume }),\n  })\n  const threshold = 1 / cfg.alpha\n  let terminal: SequentialDecision | undefined =\n    resume === undefined || resume.decision === 'continue' ? undefined : resume.decision\n  return {\n    observe(delta: number): SequentialObservation {\n      if (terminal === 'undecided-at-maxN') {\n        throw new Error(\n          `sequentialPairedGate: pre-registered maxN=${cfg.maxN} exhausted — extending the ` +\n            'stream after seeing the result reopens optional stopping; start a NEW ' +\n            'pre-registered test',\n        )\n      }\n      if (!Number.isFinite(delta) || Math.abs(delta) > cfg.scale) {\n        throw new Error(\n          `sequentialPairedGate: delta ${delta} outside ±scale=${cfg.scale} — pass the judge's ` +\n            `native scale explicitly (detectScale helps pick 1 vs 100)`,\n        )\n      }\n      const d = cfg.direction === 'decrease' ? -delta : delta\n      const step = proc.update((d / cfg.scale + 1) / 2)\n      // Stopping rule: FIRST n ≥ minN with current wealth ≥ 1/alpha — a valid\n      // stopping time (measurable w.r.t. the past), sticky once latched. The\n      // core's own `decided` latch ignores minN, so the gate re-derives the\n      // decision from current wealth here.\n      if (terminal === undefined && step.n >= cfg.minN && step.wealth >= threshold) {\n        terminal = 'promote'\n      }\n      if (terminal === 'promote') {\n        return {\n          decision: 'promote',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} at n=${step.n} ` +\n            `(minN=${cfg.minN}): the paired improvement exceeds ${cfg.minEffect} at anytime-valid ` +\n            `level α=${cfg.alpha}`,\n        }\n      }\n      if (step.n >= cfg.maxN) {\n        terminal = 'undecided-at-maxN'\n        return {\n          decision: 'undecided-at-maxN',\n          eValue: step.wealth,\n          n: step.n,\n          reason:\n            `undecided at pre-registered maxN=${cfg.maxN} (e-value ${step.wealth.toFixed(2)} < ` +\n            `1/α=${threshold.toFixed(2)}). This is NOT evidence of no effect — the effect may be ` +\n            'real but smaller than this budget can detect; re-register with a larger N to test that',\n        }\n      }\n      return {\n        decision: 'continue',\n        eValue: step.wealth,\n        n: step.n,\n        reason: `e-value ${step.wealth.toFixed(2)} < 1/α=${threshold.toFixed(2)} at n=${step.n}/${cfg.maxN} — keep observing`,\n      }\n    },\n    state() {\n      return { ...proc.state(), decision: terminal ?? 'continue' }\n    },\n  }\n}\n\n/** Data-independent in-place Fisher–Yates with a seeded PRNG — the permutation\n *  depends only on the seed, never the values, so bet predictability survives. */\nfunction seededShuffle<T>(items: T[], seed: number): T[] {\n  const rng = mulberry32(seed)\n  for (let i = items.length - 1; i > 0; i--) {\n    const j = Math.floor(rng() * (i + 1))\n    const tmp = items[i]!\n    items[i] = items[j]!\n    items[j] = tmp\n  }\n  return items\n}\n\n/**\n * Anytime-valid sequential paired gate. Conforms to the existing `Gate`\n * contract: decide(ctx) pairs candidate and baseline by full cellId, then\n * averages deltas within each configured unit (scenario by default). The\n * e-process consumes one observation per unit, with equal unit weights.\n * The streaming observe(delta) entry consumes caller-aggregated independent\n * units. Repetitions improve a unit's precision and do not increase n.\n *\n * Decision mapping onto the substrate's five-valued `GateDecision`:\n *   - 'promote'            → 'ship'\n *   - 'continue'           → 'need_more_work' (stream ended before maxN with\n *                            the e-value undecided — more independent units needed)\n *   - 'undecided-at-maxN'  → 'hold', with the reason stating it is NOT\n *                            evidence of no effect (never a silent default)\n */\nexport function sequentialPairedGate<TArtifact = unknown, TScenario extends Scenario = Scenario>(\n  options: SequentialPairedGateOptions,\n): SequentialPairedGate<TArtifact, TScenario> {\n  const cfg = resolveConfig(options)\n  const name = options.name ?? 'sequentialPairedGate'\n  const manifest = options.preRegistration\n  const observeStream = makeStream(cfg, options.resume)\n  const independentUnitByScenarioId =\n    options.independentUnitByScenarioId === undefined\n      ? undefined\n      : new Map(options.independentUnitByScenarioId)\n\n  return {\n    name,\n    observe(delta: number): SequentialObservation {\n      return observeStream.observe(delta)\n    },\n    state() {\n      return observeStream.state()\n    },\n    async decide(ctx: GateContext<TArtifact, TScenario>): Promise<GateResult> {\n      if (!ctx.baselineJudgeScores) {\n        throw new Error(\n          `${name}: ctx.baselineJudgeScores is required — falling back to the candidate's own ` +\n            'scores would compare the candidate against itself (delta 0, silent no-op)',\n        )\n      }\n      const scenarioIds = new Set(ctx.scenarios.map((s) => s.id))\n      const unitMap = independentUnitByScenarioId ?? new Map([...scenarioIds].map((id) => [id, id]))\n      for (const id of scenarioIds) {\n        const unit = unitMap.get(id)\n        if (typeof unit !== 'string' || unit.length === 0 || unit.trim() !== unit) {\n          throw new Error(`${name}: missing independent unit for scenario '${id}'`)\n        }\n      }\n      const paired = pairHoldout(\n        ctx.judgeScores,\n        ctx.baselineJudgeScores,\n        scenarioIds,\n        (s) => s.composite,\n      )\n      for (let i = 0; i < paired.cellIds.length; i++) {\n        const delta = paired.after[i]! - paired.before[i]!\n        if (Math.abs(delta) > cfg.scale) {\n          throw new Error(\n            `${name}: cell '${paired.cellIds[i]}' delta ${delta} outside ±scale=${cfg.scale}`,\n          )\n        }\n      }\n      const observations = aggregatePairedHoldout(paired, unitMap)\n      const deltas = observations.after.map((after, i) => after - observations.before[i]!)\n      seededShuffle(deltas, cfg.shuffleSeed)\n\n      const stream = makeStream(cfg)\n      let last: SequentialObservation | undefined\n      for (const d of deltas) {\n        last = stream.observe(d)\n        if (last.decision !== 'continue') break\n      }\n\n      const detail = {\n        ...stream.state(),\n        minN: cfg.minN,\n        maxN: cfg.maxN,\n        scale: cfg.scale,\n        shuffleSeed: cfg.shuffleSeed,\n        direction: cfg.direction,\n        minEffect: cfg.minEffect,\n        pairedN: deltas.length,\n        pairedCellN: paired.cellIds.length,\n        observationUnit: independentUnitByScenarioId === undefined ? 'scenario' : 'registered',\n        unitIds: observations.unitIds,\n        ...(manifest ? { preRegisteredId: manifest.id, metric: manifest.metric } : {}),\n      }\n      const meanDelta =\n        deltas.length === 0 ? undefined : deltas.reduce((s, d) => s + d, 0) / deltas.length\n\n      if (last === undefined) {\n        return {\n          decision: 'need_more_work',\n          reasons: [`${name}: no paired holdout observations — nothing to test`],\n          contributingGates: [{ name, status: 'not_evaluated', detail }],\n        }\n      }\n      const decision =\n        last.decision === 'promote'\n          ? 'ship'\n          : last.decision === 'continue'\n            ? 'need_more_work'\n            : 'hold'\n      return {\n        decision,\n        reasons: [`${name}: ${last.reason}`],\n        contributingGates: [\n          {\n            name,\n            status:\n              decision === 'ship'\n                ? 'pass'\n                : decision === 'need_more_work'\n                  ? 'not_evaluated'\n                  : 'fail',\n            detail,\n          },\n        ],\n        delta: meanDelta,\n      }\n    },\n  }\n}\n\nexport interface SequentialDecideOptions {\n  /** Type-I budget for the early-stop evidence. Default 0.05. */\n  alpha?: number\n  /** Minimum paired deltas before a stop may fire. Default 5. */\n  minN?: number\n  /** Bet truncation forwarded to `eProcess`. Default 0.5. */\n  maxBet?: number\n  /** Bound on |per-scenario composite delta|. Default 1. */\n  scale?: number\n}\n\nexport interface SequentialDecideFn {\n  (args: { history: GenerationRecord[] }): { stop: boolean; reason?: string }\n  /** Read-only snapshot of the accumulated e-process (observability + tests). */\n  state(): EProcessState\n}\n\n/**\n * SurfaceProposer.decide adapter that stops exploration at an e-value threshold.\n * The selected candidate still requires an independent held-out decision.\n *\n * Stream: for each generation g ≥ 1, the per-scenario composite deltas of\n * generation g's top candidate vs the generation-0 top candidate (the\n * incumbent the loop set out to beat), paired by scenarioId. H0: no proposed\n * surface improves any scenario's expected composite over the incumbent —\n * under it every delta has conditional mean ≤ 0 and the e-process is valid.\n * Once wealth ≥ 1/alpha the loop stops and hands the winner to the promotion\n * gate (which re-scores on HELD-OUT data — this adapter only spends the\n * exploration budget, it never promotes).\n *\n * This is an exploration stopping heuristic. Selection of the best measured\n * candidate and reuse of measured incumbent scores generally violate the\n * conditional-mean null, even when the candidates have no true improvement.\n * More repetitions reduce noise but do not establish exact type-I control.\n * An UNDECIDED process never stops the loop — absence of a\n * crossing is NOT evidence of no effect, so the loop simply runs its normal\n * course. Calling the adapter repeatedly with a growing history consumes each\n * generation exactly once (re-feeding an already-seen record would double-count\n * evidence).\n */\nexport function sequentialDecide(options: SequentialDecideOptions = {}): SequentialDecideFn {\n  const alpha = options.alpha ?? 0.05\n  const minN = options.minN ?? 5\n  const scale = options.scale ?? 1\n  if (!Number.isFinite(scale) || scale <= 0) {\n    throw new Error(`sequentialDecide: scale must be > 0, got ${scale}`)\n  }\n  const proc = eProcess({ alpha, maxBet: options.maxBet ?? 0.5, nullMean: 0.5 })\n  const threshold = 1 / alpha\n  let processedGenerations = 0\n  let reference: Map<string, number> | undefined\n  let stopped: { stop: true; reason: string } | undefined\n\n  const topCandidate = (record: GenerationRecord) => {\n    if (record.candidates.length === 0) {\n      throw new Error(\n        `sequentialDecide: generation ${record.generationIndex} has no candidates — ` +\n          'cannot extract a top candidate',\n      )\n    }\n    return record.candidates.reduce((best, candidate) => {\n      if (best.composite === null) return candidate\n      if (candidate.composite === null) return best\n      return candidate.composite > best.composite ? candidate : best\n    })\n  }\n\n  const decide: SequentialDecideFn = ({ history }) => {\n    if (stopped) return stopped\n    if (history.length === 0) return { stop: false }\n    if (reference === undefined) {\n      reference = new Map(\n        topCandidate(history[0]!).scenarios.map((s) => [s.scenarioId, s.composite]),\n      )\n    }\n    // Consume only generations not yet seen; start at 1 — generation 0 IS the\n    // reference and would contribute all-zero deltas (pure wealth dilution).\n    for (let g = Math.max(1, processedGenerations); g < history.length; g++) {\n      const top = topCandidate(history[g]!)\n      const byScenario = new Map(top.scenarios.map((s) => [s.scenarioId, s.composite]))\n      for (const [scenarioId, refComposite] of reference) {\n        const candComposite = byScenario.get(scenarioId)\n        if (candComposite === undefined) {\n          throw new Error(\n            `sequentialDecide: generation ${history[g]!.generationIndex} top candidate is missing ` +\n              `scenario '${scenarioId}' — generations must score the same scenario set to pair`,\n          )\n        }\n        const delta = candComposite - refComposite\n        if (!Number.isFinite(delta) || Math.abs(delta) > scale) {\n          throw new Error(\n            `sequentialDecide: paired delta ${delta} outside ±scale=${scale} on scenario ` +\n              `'${scenarioId}' — pass the composite scale explicitly`,\n          )\n        }\n        const step = proc.update((delta / scale + 1) / 2)\n        if (step.n >= minN && step.wealth >= threshold) {\n          stopped = {\n            stop: true,\n            reason:\n              `sequential e-process decided at generation ${history[g]!.generationIndex}: ` +\n              `e-value ${step.wealth.toFixed(2)} ≥ 1/α=${threshold.toFixed(2)} after n=${step.n} ` +\n              'paired deltas vs the generation-0 incumbent; stop exploring and require ' +\n              'an independent held-out promotion decision',\n          }\n          processedGenerations = history.length\n          return stopped\n        }\n      }\n    }\n    processedGenerations = history.length\n    return { stop: false }\n  }\n  decide.state = () => proc.state()\n  return decide\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmIA,SAAS,mBAAmB,GAA4B;CACtD,OAAO,sBAAsB,CAAC,MAAM,EAAE;AACxC;AAEA,SAAS,cAAc,MAAmD;CACxE,MAAM,IAAI,KAAK;CACf,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI;CACJ,IAAI,GAAG;EACL,IAAI,CAAC,mBAAmB,CAAC,GACvB,MAAM,IAAI,MACR,oDAAoD,EAAE,GAAG,mCAC3D;EAEF,IAAI,KAAK,UAAU,KAAA,KAAa,KAAK,UAAU,EAAE,OAC/C,MAAM,IAAI,MACR,+BAA+B,KAAK,MAAM,uCAAuC,EAAE,MAAM,oEAE3F;EAEF,IAAI,KAAK,SAAS,KAAA,KAAa,KAAK,SAAS,EAAE,gBAC7C,MAAM,IAAI,MACR,8BAA8B,KAAK,KAAK,mCAAmC,EAAE,gBAC/E;EAEF,QAAQ,EAAE;EACV,OAAO,EAAE;EACT,YAAY,EAAE;EACd,YAAY,EAAE;CAChB,OAAO;EACL,QAAQ,KAAK,SAAS;EACtB,IAAI,KAAK,SAAS,KAAA,GAChB,MAAM,IAAI,MACR,mKAEF;EAEF,OAAO,KAAK;EACZ,YAAY;EACZ,YAAY;CACd;CACA,MAAM,QAAQ,KAAK,SAAS;CAC5B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,gDAAgD,OAAO;CAEzE,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,GACpC,MAAM,IAAI,MAAM,8DAA8D,MAAM;CAEtF,MAAM,OAAO,KAAK,QAAQ;CAC1B,IAAI,CAAC,OAAO,UAAU,IAAI,KAAK,OAAO,KAAK,OAAO,MAChD,MAAM,IAAI,MACR,6DAA6D,KAAK,SAAS,MAC7E;CAEF,IAAI,CAAC,OAAO,SAAS,SAAS,KAAK,YAAY,KAAK,aAAa,OAC/D,MAAM,IAAI,MACR,wDAAwD,MAAM,SAAS,WACzE;CAKF,MAAM,WAAW,KAAM,aAAa,IAAI;CACxC,OAAO;EACL;EACA;EACA;EACA,QAAQ,KAAK,UAAU;EACvB;EACA,aAAa,KAAK,eAAe;EACjC;EACA;EACA;CACF;AACF;AAOA,MAAM,uBAAsD;CAC1D;CACA;CACA;AACF;;;;AAKA,SAAS,sBAAsB,QAA+B,KAA2B;CACvF,MAAM,QAAQ,WAA0B;EACtC,MAAM,IAAI,MAAM,yCAAyC,QAAQ;CACnE;CACA,IAAI,CAAC,qBAAqB,SAAS,OAAO,QAAQ,GAChD,KAAK,qBAAqB,OAAO,OAAO,QAAQ,EAAE,EAAE;CAEtD,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,cAAc,OAAO,EAAE,mCAAmC,IAAI,MAAM;CAE3E,MAAM,YAAY,IAAI,IAAI;CAC1B,QAAQ,OAAO,UAAf;EACE,KAAK;GACH,IAAI,OAAO,IAAI,IAAI,MACjB,KAAK,2BAA2B,OAAO,EAAE,iBAAiB,IAAI,MAAM;GAEtE;EACF,KAAK;GACH,IAAI,OAAO,MAAM,IAAI,MACnB,KAAK,qCAAqC,OAAO,EAAE,uBAAuB,IAAI,MAAM;GAEtF;EACF,KAAK;GACH,IAAI,OAAO,KAAK,IAAI,MAClB,KAAK,4BAA4B,OAAO,EAAE,aAAa,IAAI,KAAK,yBAAyB;GAE3F,IAAI,OAAO,KAAK,IAAI,QAAQ,OAAO,UAAU,WAC3C,KACE,4BAA4B,OAAO,EAAE,UAAU,IAAI,KAAK,gBACnD,OAAO,OAAO,SAAS,UAAU,kCACxC;CAGN;AACF;AAEA,SAAS,WAAW,KAAqB,QAAkD;CACzF,IAAI,WAAW,KAAA,GAAW,sBAAsB,QAAQ,GAAG;CAC3D,MAAM,OAAO,SAAS;EACpB,OAAO,IAAI;EACX,QAAQ,IAAI;EACZ,UAAU,IAAI;EACd,GAAI,WAAW,KAAA,IAAY,CAAC,IAAI,EAAE,OAAO;CAC3C,CAAC;CACD,MAAM,YAAY,IAAI,IAAI;CAC1B,IAAI,WACF,WAAW,KAAA,KAAa,OAAO,aAAa,aAAa,KAAA,IAAY,OAAO;CAC9E,OAAO;EACL,QAAQ,OAAsC;GAC5C,IAAI,aAAa,qBACf,MAAM,IAAI,MACR,6CAA6C,IAAI,KAAK,qHAGxD;GAEF,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,OACnD,MAAM,IAAI,MACR,+BAA+B,MAAM,kBAAkB,IAAI,MAAM,8EAEnE;GAEF,MAAM,IAAI,IAAI,cAAc,aAAa,CAAC,QAAQ;GAClD,MAAM,OAAO,KAAK,QAAQ,IAAI,IAAI,QAAQ,KAAK,CAAC;GAKhD,IAAI,aAAa,KAAA,KAAa,KAAK,KAAK,IAAI,QAAQ,KAAK,UAAU,WACjE,WAAW;GAEb,IAAI,aAAa,WACf,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QACE,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,SACtE,IAAI,KAAK,oCAAoC,IAAI,UAAU,4BACzD,IAAI;GACnB;GAEF,IAAI,KAAK,KAAK,IAAI,MAAM;IACtB,WAAW;IACX,OAAO;KACL,UAAU;KACV,QAAQ,KAAK;KACb,GAAG,KAAK;KACR,QACE,oCAAoC,IAAI,KAAK,YAAY,KAAK,OAAO,QAAQ,CAAC,EAAE,SACzE,UAAU,QAAQ,CAAC,EAAE;IAEhC;GACF;GACA,OAAO;IACL,UAAU;IACV,QAAQ,KAAK;IACb,GAAG,KAAK;IACR,QAAQ,WAAW,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,QAAQ,KAAK,EAAE,GAAG,IAAI,KAAK;GACrG;EACF;EACA,QAAQ;GACN,OAAO;IAAE,GAAG,KAAK,MAAM;IAAG,UAAU,YAAY;GAAW;EAC7D;CACF;AACF;;;AAIA,SAAS,cAAiB,OAAY,MAAmB;CACvD,MAAM,MAAM,WAAW,IAAI;CAC3B,KAAK,IAAI,IAAI,MAAM,SAAS,GAAG,IAAI,GAAG,KAAK;EACzC,MAAM,IAAI,KAAK,MAAM,IAAI,KAAK,IAAI,EAAE;EACpC,MAAM,MAAM,MAAM;EAClB,MAAM,KAAK,MAAM;EACjB,MAAM,KAAK;CACb;CACA,OAAO;AACT;;;;;;;;;;;;;;;;AAiBA,SAAgB,qBACd,SAC4C;CAC5C,MAAM,MAAM,cAAc,OAAO;CACjC,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,WAAW,QAAQ;CACzB,MAAM,gBAAgB,WAAW,KAAK,QAAQ,MAAM;CACpD,MAAM,8BACJ,QAAQ,gCAAgC,KAAA,IACpC,KAAA,IACA,IAAI,IAAI,QAAQ,2BAA2B;CAEjD,OAAO;EACL;EACA,QAAQ,OAAsC;GAC5C,OAAO,cAAc,QAAQ,KAAK;EACpC;EACA,QAAQ;GACN,OAAO,cAAc,MAAM;EAC7B;EACA,MAAM,OAAO,KAA6D;GACxE,IAAI,CAAC,IAAI,qBACP,MAAM,IAAI,MACR,GAAG,KAAK,sJAEV;GAEF,MAAM,cAAc,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,EAAE,EAAE,CAAC;GAC1D,MAAM,UAAU,+BAA+B,IAAI,IAAI,CAAC,GAAG,WAAW,CAAC,CAAC,KAAK,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC;GAC7F,KAAK,MAAM,MAAM,aAAa;IAC5B,MAAM,OAAO,QAAQ,IAAI,EAAE;IAC3B,IAAI,OAAO,SAAS,YAAY,KAAK,WAAW,KAAK,KAAK,KAAK,MAAM,MACnE,MAAM,IAAI,MAAM,GAAG,KAAK,2CAA2C,GAAG,EAAE;GAE5E;GACA,MAAM,SAAS,YACb,IAAI,aACJ,IAAI,qBACJ,cACC,MAAM,EAAE,SACX;GACA,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,QAAQ,KAAK;IAC9C,MAAM,QAAQ,OAAO,MAAM,KAAM,OAAO,OAAO;IAC/C,IAAI,KAAK,IAAI,KAAK,IAAI,IAAI,OACxB,MAAM,IAAI,MACR,GAAG,KAAK,UAAU,OAAO,QAAQ,GAAG,UAAU,MAAM,kBAAkB,IAAI,OAC5E;GAEJ;GACA,MAAM,eAAe,uBAAuB,QAAQ,OAAO;GAC3D,MAAM,SAAS,aAAa,MAAM,KAAK,OAAO,MAAM,QAAQ,aAAa,OAAO,EAAG;GACnF,cAAc,QAAQ,IAAI,WAAW;GAErC,MAAM,SAAS,WAAW,GAAG;GAC7B,IAAI;GACJ,KAAK,MAAM,KAAK,QAAQ;IACtB,OAAO,OAAO,QAAQ,CAAC;IACvB,IAAI,KAAK,aAAa,YAAY;GACpC;GAEA,MAAM,SAAS;IACb,GAAG,OAAO,MAAM;IAChB,MAAM,IAAI;IACV,MAAM,IAAI;IACV,OAAO,IAAI;IACX,aAAa,IAAI;IACjB,WAAW,IAAI;IACf,WAAW,IAAI;IACf,SAAS,OAAO;IAChB,aAAa,OAAO,QAAQ;IAC5B,iBAAiB,gCAAgC,KAAA,IAAY,aAAa;IAC1E,SAAS,aAAa;IACtB,GAAI,WAAW;KAAE,iBAAiB,SAAS;KAAI,QAAQ,SAAS;IAAO,IAAI,CAAC;GAC9E;GACA,MAAM,YACJ,OAAO,WAAW,IAAI,KAAA,IAAY,OAAO,QAAQ,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;GAE/E,IAAI,SAAS,KAAA,GACX,OAAO;IACL,UAAU;IACV,SAAS,CAAC,GAAG,KAAK,mDAAmD;IACrE,mBAAmB,CAAC;KAAE;KAAM,QAAQ;KAAiB;IAAO,CAAC;GAC/D;GAEF,MAAM,WACJ,KAAK,aAAa,YACd,SACA,KAAK,aAAa,aAChB,mBACA;GACR,OAAO;IACL;IACA,SAAS,CAAC,GAAG,KAAK,IAAI,KAAK,QAAQ;IACnC,mBAAmB,CACjB;KACE;KACA,QACE,aAAa,SACT,SACA,aAAa,mBACX,kBACA;KACR;IACF,CACF;IACA,OAAO;GACT;EACF;CACF;AACF;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,iBAAiB,UAAmC,CAAC,GAAuB;CAC1F,MAAM,QAAQ,QAAQ,SAAS;CAC/B,MAAM,OAAO,QAAQ,QAAQ;CAC7B,MAAM,QAAQ,QAAQ,SAAS;CAC/B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACtC,MAAM,IAAI,MAAM,4CAA4C,OAAO;CAErE,MAAM,OAAO,SAAS;EAAE;EAAO,QAAQ,QAAQ,UAAU;EAAK,UAAU;CAAI,CAAC;CAC7E,MAAM,YAAY,IAAI;CACtB,IAAI,uBAAuB;CAC3B,IAAI;CACJ,IAAI;CAEJ,MAAM,gBAAgB,WAA6B;EACjD,IAAI,OAAO,WAAW,WAAW,GAC/B,MAAM,IAAI,MACR,gCAAgC,OAAO,gBAAgB,oDAEzD;EAEF,OAAO,OAAO,WAAW,QAAQ,MAAM,cAAc;GACnD,IAAI,KAAK,cAAc,MAAM,OAAO;GACpC,IAAI,UAAU,cAAc,MAAM,OAAO;GACzC,OAAO,UAAU,YAAY,KAAK,YAAY,YAAY;EAC5D,CAAC;CACH;CAEA,MAAM,UAA8B,EAAE,cAAc;EAClD,IAAI,SAAS,OAAO;EACpB,IAAI,QAAQ,WAAW,GAAG,OAAO,EAAE,MAAM,MAAM;EAC/C,IAAI,cAAc,KAAA,GAChB,YAAY,IAAI,IACd,aAAa,QAAQ,EAAG,CAAC,CAAC,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAC5E;EAIF,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,oBAAoB,GAAG,IAAI,QAAQ,QAAQ,KAAK;GACvE,MAAM,MAAM,aAAa,QAAQ,EAAG;GACpC,MAAM,aAAa,IAAI,IAAI,IAAI,UAAU,KAAK,MAAM,CAAC,EAAE,YAAY,EAAE,SAAS,CAAC,CAAC;GAChF,KAAK,MAAM,CAAC,YAAY,iBAAiB,WAAW;IAClD,MAAM,gBAAgB,WAAW,IAAI,UAAU;IAC/C,IAAI,kBAAkB,KAAA,GACpB,MAAM,IAAI,MACR,gCAAgC,QAAQ,EAAE,CAAE,gBAAgB,sCAC7C,WAAW,yDAC5B;IAEF,MAAM,QAAQ,gBAAgB;IAC9B,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,KAAK,IAAI,KAAK,IAAI,OAC/C,MAAM,IAAI,MACR,kCAAkC,MAAM,kBAAkB,MAAM,gBAC1D,WAAW,wCACnB;IAEF,MAAM,OAAO,KAAK,QAAQ,QAAQ,QAAQ,KAAK,CAAC;IAChD,IAAI,KAAK,KAAK,QAAQ,KAAK,UAAU,WAAW;KAC9C,UAAU;MACR,MAAM;MACN,QACE,8CAA8C,QAAQ,EAAE,CAAE,gBAAgB,YAC/D,KAAK,OAAO,QAAQ,CAAC,EAAE,SAAS,UAAU,QAAQ,CAAC,EAAE,WAAW,KAAK,EAAE;KAGtF;KACA,uBAAuB,QAAQ;KAC/B,OAAO;IACT;GACF;EACF;EACA,uBAAuB,QAAQ;EAC/B,OAAO,EAAE,MAAM,MAAM;CACvB;CACA,OAAO,cAAc,KAAK,MAAM;CAChC,OAAO;AACT"}

package/dist/{skillopt-optimization-method-Dz5CGZb8.js → skillopt-optimization-method-bEm4jYV4.js}

@@ -1,8 +1,8 @@
 import { a as hashCanonical } from "./canonical-BsZVgBhW.js";
-import { a as heldoutSignificance, o as pairHoldout } from "./statistical-heldout-CNLOvRq_.js";
+import { a as heldoutSignificance, o as pairHoldout } from "./statistical-heldout-C26hOFWD.js";
 import { n as contentHash, t as canonicalJson } from "./verdict-cache-BQEsVy9n.js";
-import { D as surfaceDiff, O as surfaceNode, T as developmentClaim, d as combineComparisonCosts, f as costFromLedgerSummary, l as optimizationTokenUsageFromSummary, m as readGepaCandidatePopulationArtifact, p as assertGepaCandidatePopulationSummary, rt as runCampaign, t as llmJudge, w as SearchRecorder } from "./llm-judge-DyZdrp6I.js";
-import { F as surfaceContentHash, b as campaignBreakdown } from "./campaign-evidence-Dt_rWPk2.js";
+import { D as surfaceDiff, O as surfaceNode, T as developmentClaim, d as combineComparisonCosts, f as costFromLedgerSummary, l as optimizationTokenUsageFromSummary, m as readGepaCandidatePopulationArtifact, p as assertGepaCandidatePopulationSummary, rt as runCampaign, t as llmJudge, w as SearchRecorder } from "./llm-judge-Bka5q-ds.js";
+import { F as surfaceContentHash, b as campaignBreakdown } from "./campaign-evidence-C__cUpCt.js";
 import { t as deepFreezeCanonicalJson } from "./deep-freeze-CurcOQg3.js";
 import { S as fsCampaignStorage, _ as removeCredentialEnvironment, b as safePathComponent, d as assertExternalOptimizerModelBudget, f as assertJsonValue, g as isRecord, h as isExternalTextCandidate, i as startExternalOptimizerModelProxy, m as isCandidateText, n as closeExternalOptimizerResources, p as assertNoCredentialValues, r as runWithCleanup, t as runExternalOptimizerProcess, v as resolveExternalOptimizerCallbackLimits, x as createRunCostLedger, y as resolveExternalOptimizerProcessLimits } from "./external-optimizer-subprocess-BmKtH3T6.js";
 import { n as openSearchLedger } from "./search-ledger-CXQPwdWz.js";
@@ -2583,4 +2583,4 @@ function skillOptOptimizationMethod(config) {
 //#endregion
 export { importGepaPopulation as a, heldOutGate as c, decodeExternalTextCandidate as d, readExternalOptimizerObservationArtifact as f, runReferenceEquivalenceJudge as g, createReferenceEquivalenceJudge as h, importExternalEvaluations as i, composeGate as l, REFERENCE_EQUIVALENCE_JUDGE_VERSION as m, gepaOptimizationMethod as n, recordGepaSearch as o, REFERENCE_EQUIVALENCE_INPUT_LIMITS as p, externalSurface as r, recordSkillOptSearch as s, skillOptOptimizationMethod as t, externalTextOptimizationMethod as u };
 
-//# sourceMappingURL=skillopt-optimization-method-Dz5CGZb8.js.map
+//# sourceMappingURL=skillopt-optimization-method-bEm4jYV4.js.map