@tangle-network/agent-eval 0.190.1 → 0.192.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (167) hide show
  1. package/CHANGELOG.md +69 -0
  2. package/dist/adapters/http.d.ts +2 -2
  3. package/dist/{agent-profile-DNV_U6QR.d.ts → agent-profile-DDfIjdyS.d.ts} +3 -3
  4. package/dist/{agent-profile-DNV_U6QR.d.ts.map → agent-profile-DDfIjdyS.d.ts.map} +1 -1
  5. package/dist/analyst/index.d.ts +11 -11
  6. package/dist/analyst/index.js +5 -5
  7. package/dist/{insight-report-CvPkEIcC.d.ts → analyze-runs-CPiKLJFs.d.ts} +73 -3
  8. package/dist/analyze-runs-CPiKLJFs.d.ts.map +1 -0
  9. package/dist/{attribute-vocabulary-DLJ6303h.d.ts → attribute-vocabulary-CcXqHODT.d.ts} +23 -12
  10. package/dist/attribute-vocabulary-CcXqHODT.d.ts.map +1 -0
  11. package/dist/{benchmark-command-BXIr_k8D.js → benchmark-command-C458yEFK.js} +9 -9
  12. package/dist/{benchmark-command-BXIr_k8D.js.map → benchmark-command-C458yEFK.js.map} +1 -1
  13. package/dist/benchmarks/index.d.ts +3 -3
  14. package/dist/benchmarks/index.js +3 -3
  15. package/dist/campaign/index.d.ts +6 -6
  16. package/dist/campaign/index.js +8 -7
  17. package/dist/{campaign-B7q74uIh.js → campaign-BVQVBEfH.js} +13 -325
  18. package/dist/campaign-BVQVBEfH.js.map +1 -0
  19. package/dist/{campaign-evidence-D40-LHIL.js → campaign-evidence-DcQiYhJI.js} +18 -154
  20. package/dist/campaign-evidence-DcQiYhJI.js.map +1 -0
  21. package/dist/claim-V9Zqbq0J.js +54 -0
  22. package/dist/claim-V9Zqbq0J.js.map +1 -0
  23. package/dist/cli.js +99 -2
  24. package/dist/cli.js.map +1 -1
  25. package/dist/client-wFyDisxi.d.ts +236 -0
  26. package/dist/client-wFyDisxi.d.ts.map +1 -0
  27. package/dist/contract/index.d.ts +10 -11
  28. package/dist/contract/index.js +10 -140
  29. package/dist/contract/index.js.map +1 -1
  30. package/dist/{default-registry-B5HrwrPM.d.ts → default-registry-DdqA_-_o.d.ts} +6 -6
  31. package/dist/{default-registry-B5HrwrPM.d.ts.map → default-registry-DdqA_-_o.d.ts.map} +1 -1
  32. package/dist/{default-registry-16I-Bxlf.js → default-registry-xMLX160E.js} +3 -3
  33. package/dist/{default-registry-16I-Bxlf.js.map → default-registry-xMLX160E.js.map} +1 -1
  34. package/dist/{define-agent-eval-DPizhHfy.js → define-agent-eval-DQuos3Aw.js} +30 -113
  35. package/dist/define-agent-eval-DQuos3Aw.js.map +1 -0
  36. package/dist/{define-agent-eval-DblGedQk.d.ts → define-agent-eval-DcludNJS.d.ts} +25 -38
  37. package/dist/define-agent-eval-DcludNJS.d.ts.map +1 -0
  38. package/dist/diagnosis/index.d.ts +2 -2
  39. package/dist/diagnosis/index.js +1 -1
  40. package/dist/{dspy-rlm-engine-Cs5LzJ-f.js → dspy-rlm-engine-BX34vPVY.js} +3 -3
  41. package/dist/{dspy-rlm-engine-Cs5LzJ-f.js.map → dspy-rlm-engine-BX34vPVY.js.map} +1 -1
  42. package/dist/{engine-BkLCnZQX.d.ts → engine-C1DIAEpn.d.ts} +4 -4
  43. package/dist/{engine-BkLCnZQX.d.ts.map → engine-C1DIAEpn.d.ts.map} +1 -1
  44. package/dist/evaluation.d.ts +2 -2
  45. package/dist/{exact-types-CfpP2A4Z.d.ts → exact-types-KGxBWH4b.d.ts} +2 -2
  46. package/dist/{exact-types-CfpP2A4Z.d.ts.map → exact-types-KGxBWH4b.d.ts.map} +1 -1
  47. package/dist/experiment/index.d.ts +1 -1
  48. package/dist/experiment/index.js +2 -1
  49. package/dist/experiment/index.js.map +1 -1
  50. package/dist/{external-optimizer-process-BnSmrkgL.js → external-optimizer-process-D-Sab4v3.js} +3 -62
  51. package/dist/external-optimizer-process-D-Sab4v3.js.map +1 -0
  52. package/dist/{external-optimizer-subprocess-kbcPu5po.js → external-optimizer-subprocess-kbqC-iWf.js} +65 -27
  53. package/dist/external-optimizer-subprocess-kbqC-iWf.js.map +1 -0
  54. package/dist/{extract-usage-BrQ8mCLX.js → extract-usage-DFATEuWW.js} +2 -2
  55. package/dist/{extract-usage-BrQ8mCLX.js.map → extract-usage-DFATEuWW.js.map} +1 -1
  56. package/dist/{feedback-trajectory-BYjh7NEM.d.ts → feedback-trajectory-BXCGTRkz.d.ts} +3 -3
  57. package/dist/{feedback-trajectory-BYjh7NEM.d.ts.map → feedback-trajectory-BXCGTRkz.d.ts.map} +1 -1
  58. package/dist/hosted/index.d.ts +66 -6
  59. package/dist/hosted/index.d.ts.map +1 -1
  60. package/dist/hosted/index.js +2 -2
  61. package/dist/{index-I4UJGRno.d.ts → index-Bj96W0jZ.d.ts} +28 -2
  62. package/dist/{index-I4UJGRno.d.ts.map → index-Bj96W0jZ.d.ts.map} +1 -1
  63. package/dist/{index-2e_MTznj.d.ts → index-DBoXFaz0.d.ts} +6 -6
  64. package/dist/{index-2e_MTznj.d.ts.map → index-DBoXFaz0.d.ts.map} +1 -1
  65. package/dist/{index-CWoIJHmy.d.ts → index-DNVzu9Vv.d.ts} +10 -104
  66. package/dist/index-DNVzu9Vv.d.ts.map +1 -0
  67. package/dist/{index-CJv56rhX.d.ts → index-DwQxdWve.d.ts} +351 -135
  68. package/dist/index-DwQxdWve.d.ts.map +1 -0
  69. package/dist/index.d.ts +16 -17
  70. package/dist/index.d.ts.map +1 -1
  71. package/dist/index.js +11 -11
  72. package/dist/jev.d.ts +2 -2
  73. package/dist/{journal-c3qY0SDP.js → journal-DvKCvjty.js} +94 -2
  74. package/dist/{journal-c3qY0SDP.js.map → journal-DvKCvjty.js.map} +1 -1
  75. package/dist/{kind-factory-ByZeXDqF.js → kind-factory-CdPPHtGZ.js} +10 -26
  76. package/dist/kind-factory-CdPPHtGZ.js.map +1 -0
  77. package/dist/ledger-core/index.d.ts +2 -2
  78. package/dist/ledger-core/index.js +2 -2
  79. package/dist/{llm-judge-BnxAv7ed.js → llm-judge-Z6MkTkd0.js} +1805 -2012
  80. package/dist/llm-judge-Z6MkTkd0.js.map +1 -0
  81. package/dist/{matrix-CNypzP-v.d.ts → matrix-CQMMrEQ5.d.ts} +2 -2
  82. package/dist/{matrix-CNypzP-v.d.ts.map → matrix-CQMMrEQ5.d.ts.map} +1 -1
  83. package/dist/multishot/golden/index.d.ts +1 -1
  84. package/dist/multishot/index.d.ts +2 -2
  85. package/dist/openapi.json +1 -1
  86. package/dist/{otel-spans-CQY1pNiN.js → otel-spans-DS7t1t1J.js} +8 -19
  87. package/dist/otel-spans-DS7t1t1J.js.map +1 -0
  88. package/dist/otlp-attributes-D9ptWaIJ.js +53 -0
  89. package/dist/otlp-attributes-D9ptWaIJ.js.map +1 -0
  90. package/dist/{prime-protocol-Fwvctmhm.d.ts → prime-protocol-T0i6xcAs.d.ts} +2 -2
  91. package/dist/{prime-protocol-Fwvctmhm.d.ts.map → prime-protocol-T0i6xcAs.d.ts.map} +1 -1
  92. package/dist/{produced-state-BXjXJHJT.js → produced-state-CVmCPQDJ.js} +3 -3
  93. package/dist/{produced-state-BXjXJHJT.js.map → produced-state-CVmCPQDJ.js.map} +1 -1
  94. package/dist/{registry-mAQVtuLx.d.ts → registry-Bu3lNole.d.ts} +3 -3
  95. package/dist/{registry-mAQVtuLx.d.ts.map → registry-Bu3lNole.d.ts.map} +1 -1
  96. package/dist/rl.d.ts +1 -1
  97. package/dist/rl.js +1 -1
  98. package/dist/{run-record-DnMFyg_g.js → run-record-tNnlQIrD.js} +2 -10
  99. package/dist/{run-record-DnMFyg_g.js.map → run-record-tNnlQIrD.js.map} +1 -1
  100. package/dist/search-ledger-Ck7tok0h.js +1517 -0
  101. package/dist/search-ledger-Ck7tok0h.js.map +1 -0
  102. package/dist/search-ledger-file-DY0AXX3W.js +27 -0
  103. package/dist/search-ledger-file-DY0AXX3W.js.map +1 -0
  104. package/dist/search-shipper-BG182p-U.js +781 -0
  105. package/dist/search-shipper-BG182p-U.js.map +1 -0
  106. package/dist/{semantic-concept-judge-ClB73cei.js → semantic-concept-judge-Bg-560H0.js} +4 -4
  107. package/dist/{semantic-concept-judge-ClB73cei.js.map → semantic-concept-judge-Bg-560H0.js.map} +1 -1
  108. package/dist/{skillopt-optimization-method-l2vDBtKp.js → skillopt-optimization-method-DoayuM6D.js} +6 -5
  109. package/dist/{skillopt-optimization-method-l2vDBtKp.js.map → skillopt-optimization-method-DoayuM6D.js.map} +1 -1
  110. package/dist/{statistical-heldout-CLpjYffJ.d.ts → statistical-heldout-C1Ej5Ce8.d.ts} +2 -2
  111. package/dist/{statistical-heldout-CLpjYffJ.d.ts.map → statistical-heldout-C1Ej5Ce8.d.ts.map} +1 -1
  112. package/dist/{store-otlp-84DH0OEH.js → store-otlp-BPowgPm6.js} +2 -2
  113. package/dist/{store-otlp-84DH0OEH.js.map → store-otlp-BPowgPm6.js.map} +1 -1
  114. package/dist/{store-tool-spans-DUiqyHvR.js → store-tool-spans-BcC_5ef-.js} +5 -5
  115. package/dist/{store-tool-spans-DUiqyHvR.js.map → store-tool-spans-BcC_5ef-.js.map} +1 -1
  116. package/dist/{store-tool-spans-BBXs8hDP.d.ts → store-tool-spans-ByLdU26j.d.ts} +5 -9
  117. package/dist/store-tool-spans-ByLdU26j.d.ts.map +1 -0
  118. package/dist/{task-failure-attributes-B34TiMBe.js → task-failure-attributes-CEqEfuIY.js} +17 -9
  119. package/dist/task-failure-attributes-CEqEfuIY.js.map +1 -0
  120. package/dist/{tool-groups-W7MJobX_.d.ts → tool-groups-DiBi16ns.d.ts} +3 -3
  121. package/dist/tool-groups-DiBi16ns.d.ts.map +1 -0
  122. package/dist/trace-attributes.d.ts +1 -1
  123. package/dist/trace-attributes.js +23 -75
  124. package/dist/trace-attributes.js.map +1 -1
  125. package/dist/trace-repair/index.d.ts +1 -1
  126. package/dist/traces.d.ts +11 -18
  127. package/dist/traces.d.ts.map +1 -1
  128. package/dist/traces.js +16 -16
  129. package/dist/traces.js.map +1 -1
  130. package/dist/{types-Cvqy7asY.d.ts → types-ByZslIlr.d.ts} +2 -2
  131. package/dist/{types-Cvqy7asY.d.ts.map → types-ByZslIlr.d.ts.map} +1 -1
  132. package/dist/{types-BvmWNjjs.d.ts → types-C1CAD70_.d.ts} +15 -16
  133. package/dist/types-C1CAD70_.d.ts.map +1 -0
  134. package/dist/{types-Bay1dh2o.d.ts → types-CeDfCRXp.d.ts} +11 -16
  135. package/dist/types-CeDfCRXp.d.ts.map +1 -0
  136. package/dist/wire/index.d.ts +1 -1
  137. package/docs/campaign-proposers.md +14 -16
  138. package/docs/hosted-ingest-spec.md +92 -133
  139. package/docs/public-api.md +46 -35
  140. package/docs/search-ledger.md +79 -1
  141. package/package.json +2 -2
  142. package/dist/analyze-runs-BvK4-wSs.d.ts +0 -75
  143. package/dist/analyze-runs-BvK4-wSs.d.ts.map +0 -1
  144. package/dist/attribute-vocabulary-DLJ6303h.d.ts.map +0 -1
  145. package/dist/campaign-B7q74uIh.js.map +0 -1
  146. package/dist/campaign-evidence-D40-LHIL.js.map +0 -1
  147. package/dist/client-BhEyAdfe.d.ts +0 -202
  148. package/dist/client-BhEyAdfe.d.ts.map +0 -1
  149. package/dist/client-QwJpsNYU.js +0 -655
  150. package/dist/client-QwJpsNYU.js.map +0 -1
  151. package/dist/define-agent-eval-DPizhHfy.js.map +0 -1
  152. package/dist/define-agent-eval-DblGedQk.d.ts.map +0 -1
  153. package/dist/external-optimizer-process-BnSmrkgL.js.map +0 -1
  154. package/dist/external-optimizer-subprocess-kbcPu5po.js.map +0 -1
  155. package/dist/index-CJv56rhX.d.ts.map +0 -1
  156. package/dist/index-CWoIJHmy.d.ts.map +0 -1
  157. package/dist/insight-report-CvPkEIcC.d.ts.map +0 -1
  158. package/dist/kind-factory-ByZeXDqF.js.map +0 -1
  159. package/dist/llm-judge-BnxAv7ed.js.map +0 -1
  160. package/dist/otel-spans-CQY1pNiN.js.map +0 -1
  161. package/dist/otlp-attributes-BpC_blLF.js +0 -83
  162. package/dist/otlp-attributes-BpC_blLF.js.map +0 -1
  163. package/dist/store-tool-spans-BBXs8hDP.d.ts.map +0 -1
  164. package/dist/task-failure-attributes-B34TiMBe.js.map +0 -1
  165. package/dist/tool-groups-W7MJobX_.d.ts.map +0 -1
  166. package/dist/types-Bay1dh2o.d.ts.map +0 -1
  167. package/dist/types-BvmWNjjs.d.ts.map +0 -1
@@ -207,14 +207,6 @@ function projectCampaignCellQuality(cell) {
207
207
  }
208
208
  };
209
209
  }
210
- /** Read the canonical task score without recomputing cell quality. */
211
- function campaignCellTaskScore(cell) {
212
- return projectCampaignCellQuality(cell).score;
213
- }
214
- /** Read canonical successful judge dimensions without recomputing cell quality. */
215
- function campaignCellJudgeDimensions(cell) {
216
- return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {};
217
- }
218
210
  function finiteMetrics(metrics) {
219
211
  const finite = {};
220
212
  for (const [key, value] of Object.entries(metrics ?? {})) if (Number.isFinite(value)) finite[key] = value;
@@ -224,6 +216,6 @@ function mean(values) {
224
216
  return values.reduce((sum, value) => sum + value, 0) / values.length;
225
217
  }
226
218
  //#endregion
227
- export { campaignCellToRunRecord as a, campaignCellTaskScore as i, campaignCellExecutionEvidence as n, projectCampaignCellQuality as o, campaignCellJudgeDimensions as r, campaignCellCostProvenance as t };
219
+ export { campaignCellToRunRecord as n, projectCampaignCellQuality as r, campaignCellCostProvenance as t };
228
220
 
229
- //# sourceMappingURL=run-record-DnMFyg_g.js.map
221
+ //# sourceMappingURL=run-record-tNnlQIrD.js.map
@@ -1 +1 @@
1
- {"version":3,"file":"run-record-DnMFyg_g.js","names":[],"sources":["../src/campaign/run-record.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n JudgeScoresRecord,\n RunCostProvenance,\n RunOutcome,\n RunRecord,\n RunSearchCoordinates,\n RunSplitTag,\n RunTerminalOutcome,\n RunTraceRef,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n /** With `search`, this must be `searchCellRunId(search)`. */\n runId: string\n experimentId: string\n candidateId: string\n model: string\n promptHash: string\n configHash: string\n commitSha: string\n splitTag: RunSplitTag\n seed?: number\n scenarioId?: string\n /** Estimated total for a cell that proved no spend at all. A cell that\n * proved part of its spend keeps that floor instead. */\n defaultCostUsd?: number\n agentProfile?: AgentProfileCell\n raw?: Record<string, number>\n search?: RunSearchCoordinates\n traceRef?: RunTraceRef\n}\n\nexport interface CampaignCellQualityProjection {\n score?: number\n judgeScores?: JudgeScoresRecord\n successfulJudgeScores: Record<string, JudgeScore>\n failedJudges: string[]\n raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n constructor(judgeName: string) {\n super(\n `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n `({ judgeName, dimension, score }) is a different type with the same name.`,\n )\n this.name = 'CampaignJudgeScoreShapeError'\n }\n}\n\nexport interface CampaignCellExecutionEvidence {\n terminalOutcome: RunTerminalOutcome\n executionErrorCount?: number\n judgeErrorCount?: number\n unclassifiedErrorCount?: number\n terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n options: CampaignCellRunRecordOptions,\n): RunRecord {\n const quality = projectCampaignCellQuality(cell)\n const execution = campaignCellExecutionEvidence(cell)\n const judgeErrorCount = Math.max(\n quality.raw.judge_error_count ?? 0,\n execution.judgeErrorCount ?? 0,\n )\n const costProvenance = campaignCellRunCostProvenance(cell, options.defaultCostUsd)\n const costUsd = costProvenance.usd\n const raw: Record<string, number> = {\n ...finiteMetrics(options.raw),\n ...quality.raw,\n rep: cell.rep,\n duration_ms: cell.durationMs,\n ...(costUsd === null ? {} : { cost_usd: costUsd }),\n tokens_input: cell.tokenUsage.input,\n tokens_output: cell.tokenUsage.output,\n tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n latency_ms: cell.durationMs,\n ...(execution.executionErrorCount === undefined\n ? {}\n : { execution_error_count: execution.executionErrorCount }),\n ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n ...(execution.unclassifiedErrorCount === undefined\n ? {}\n : { unclassified_error_count: execution.unclassifiedErrorCount }),\n }\n if (typeof cell.generation === 'number') raw.generation = cell.generation\n if (cell.tokenUsage.reasoning !== undefined) {\n raw.tokens_reasoning = cell.tokenUsage.reasoning\n }\n if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n if (cell.tokenUsage.cacheWrite !== undefined) {\n raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n }\n if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n }\n if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n raw.cost_per_quality = costUsd / quality.score\n }\n\n const outcome: RunOutcome = {\n raw,\n ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n }\n if (quality.score !== undefined) {\n if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n else outcome.searchScore = quality.score\n }\n\n return validateRunRecord({\n runId: options.runId,\n experimentId: options.experimentId,\n candidateId: options.candidateId,\n seed: options.seed ?? cell.seed,\n model: options.model,\n promptHash: options.promptHash,\n configHash: options.configHash,\n commitSha: options.commitSha,\n wallMs: cell.durationMs,\n costUsd,\n costProvenance,\n tokenUsage: { ...cell.tokenUsage },\n terminalOutcome: execution.terminalOutcome,\n ...(execution.terminalFailureReason\n ? { terminalFailureReason: execution.terminalFailureReason }\n : {}),\n outcome,\n splitTag: options.splitTag,\n scenarioId: options.scenarioId ?? cell.scenarioId,\n ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n ...(options.search ? { search: options.search } : {}),\n ...(options.traceRef ? { traceRef: options.traceRef } : {}),\n })\n}\n\n/**\n * A cell's cost as a RunRecord states it. An uncaptured cell's `costUsd` is\n * the spend its receipts proved, so a positive one is the run's floor. The\n * caller's `defaultCostUsd` stands in only when nothing was proved: a proven\n * floor is never replaced by a guess.\n */\nfunction campaignCellRunCostProvenance<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n defaultCostUsd: number | undefined,\n): RunCostProvenance {\n const provenance = campaignCellCostProvenance(cell)\n if (provenance.kind !== 'uncaptured') return provenance\n if (cell.costUsd > 0) return { kind: 'lower-bound', usd: null, knownLowerBoundUsd: cell.costUsd }\n return defaultCostUsd === undefined ? provenance : { kind: 'estimated', usd: defaultCostUsd }\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n }\n const provenance = cell.costProvenance\n if (!provenance || typeof provenance !== 'object') {\n throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n }\n if (provenance.kind === 'uncaptured') {\n if (provenance.usd !== null) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n }\n return { kind: 'uncaptured', usd: null }\n }\n if (\n (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n !Number.isFinite(provenance.usd) ||\n provenance.usd < 0\n ) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n }\n if (provenance.usd !== cell.costUsd) {\n throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n }\n return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n if (cell.errorStage === 'dispatch') {\n return {\n terminalOutcome: 'failed',\n executionErrorCount: 1,\n ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n }\n }\n if (cell.errorStage === 'judge') {\n return {\n terminalOutcome: 'succeeded',\n executionErrorCount: 0,\n judgeErrorCount: 1,\n }\n }\n if (!cell.error) {\n return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n }\n return {\n terminalOutcome: 'unknown',\n unclassifiedErrorCount: 1,\n }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n if (cell.errorStage === 'dispatch') {\n return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n }\n\n const perJudge: Record<string, Record<string, number>> = {}\n const successfulJudgeScores: Record<string, JudgeScore> = {}\n const dimensionValues = new Map<string, number[]>()\n const composites: number[] = []\n const notes: string[] = []\n const failedJudges = new Set<string>(\n cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n )\n const raw: Record<string, number> = {}\n\n for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n if (\n typeof dimensionsShape !== 'object' ||\n dimensionsShape === null ||\n Array.isArray(dimensionsShape)\n ) {\n throw new CampaignJudgeScoreShapeError(judgeName)\n }\n const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n failedJudges.add(judgeName)\n continue\n }\n\n composites.push(score.composite)\n successfulJudgeScores[judgeName] = score\n const dimensions = { ...score.dimensions }\n perJudge[judgeName] = dimensions\n for (const [dimension, value] of Object.entries(dimensions)) {\n raw[`${judgeName}.${dimension}`] = value\n const values = dimensionValues.get(dimension) ?? []\n values.push(value)\n dimensionValues.set(dimension, values)\n }\n if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n for (const failedJudge of score.failedJudges ?? []) {\n failedJudges.add(`${judgeName}/${failedJudge}`)\n }\n }\n\n if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n const sortedFailedJudges = [...failedJudges].sort()\n if (composites.length === 0) {\n return {\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n raw,\n }\n }\n\n const composite = mean(composites)\n const perDimMean = Object.fromEntries(\n [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n )\n const complete =\n cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n if (complete) raw.composite = composite\n\n return {\n ...(complete ? { score: composite } : {}),\n raw,\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n judgeScores: {\n perJudge,\n perDimMean,\n composite,\n ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n },\n }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n const finite: Record<string, number> = {}\n for (const [key, value] of Object.entries(metrics ?? {})) {\n if (Number.isFinite(value)) finite[key] = value\n }\n return finite\n}\n\nfunction mean(values: number[]): number {\n return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;;;;;;;;AAsDA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,iBAAiB,8BAA8B,MAAM,QAAQ,cAAc;CACjF,MAAM,UAAU,eAAe;CAC/B,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAChD,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;EACrE,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;EACnD,GAAI,QAAQ,WAAW,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;CAC3D,CAAC;AACH;;;;;;;AAQA,SAAS,8BACP,MACA,gBACmB;CACnB,MAAM,aAAa,2BAA2B,IAAI;CAClD,IAAI,WAAW,SAAS,cAAc,OAAO;CAC7C,IAAI,KAAK,UAAU,GAAG,OAAO;EAAE,MAAM;EAAe,KAAK;EAAM,oBAAoB,KAAK;CAAQ;CAChG,OAAO,mBAAmB,KAAA,IAAY,aAAa;EAAE,MAAM;EAAa,KAAK;CAAe;AAC9F;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;;AAGA,SAAgB,sBACd,MACoB;CACpB,OAAO,2BAA2B,IAAI,CAAC,CAAC;AAC1C;;AAGA,SAAgB,4BACd,MACwC;CACxC,OAAO,2BAA2B,IAAI,CAAC,CAAC,aAAa,YAAY,CAAC;AACpE;AAEA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}
1
+ {"version":3,"file":"run-record-tNnlQIrD.js","names":[],"sources":["../src/campaign/run-record.ts"],"sourcesContent":["import type { AgentProfileCell } from '../agent-profile-cell'\nimport type { CostProvenance } from '../cost-ledger'\nimport type {\n JudgeScoresRecord,\n RunCostProvenance,\n RunOutcome,\n RunRecord,\n RunSearchCoordinates,\n RunSplitTag,\n RunTerminalOutcome,\n RunTraceRef,\n} from '../run-record'\nimport { validateRunRecord } from '../run-record'\nimport type { CampaignCellResult, JudgeScore } from './types'\n\nexport interface CampaignCellRunRecordOptions {\n /** With `search`, this must be `searchCellRunId(search)`. */\n runId: string\n experimentId: string\n candidateId: string\n model: string\n promptHash: string\n configHash: string\n commitSha: string\n splitTag: RunSplitTag\n seed?: number\n scenarioId?: string\n /** Estimated total for a cell that proved no spend at all. A cell that\n * proved part of its spend keeps that floor instead. */\n defaultCostUsd?: number\n agentProfile?: AgentProfileCell\n raw?: Record<string, number>\n search?: RunSearchCoordinates\n traceRef?: RunTraceRef\n}\n\nexport interface CampaignCellQualityProjection {\n score?: number\n judgeScores?: JudgeScoresRecord\n successfulJudgeScores: Record<string, JudgeScore>\n failedJudges: string[]\n raw: Record<string, number>\n}\n\n/**\n * A campaign cell carried a judge score without a `dimensions` record.\n *\n * Two exported types share the name `JudgeScore`: the campaign verdict\n * (`{ dimensions, composite, notes }` from `@tangle-network/agent-eval/campaign`)\n * and the root export's flat per-dimension row (`{ judgeName, dimension, score }`\n * from `@tangle-network/agent-eval`). Campaign aggregation accepts only the\n * campaign shape; the flat shape previously crashed here with an opaque\n * TypeError deep inside aggregation.\n */\nexport class CampaignJudgeScoreShapeError extends TypeError {\n constructor(judgeName: string) {\n super(\n `campaign cell judge '${judgeName}' carries a score without a 'dimensions' record. ` +\n `Use the campaign JudgeScore ({ dimensions, composite, notes }) from ` +\n `'@tangle-network/agent-eval/campaign'; the root export's JudgeScore ` +\n `({ judgeName, dimension, score }) is a different type with the same name.`,\n )\n this.name = 'CampaignJudgeScoreShapeError'\n }\n}\n\nexport interface CampaignCellExecutionEvidence {\n terminalOutcome: RunTerminalOutcome\n executionErrorCount?: number\n judgeErrorCount?: number\n unclassifiedErrorCount?: number\n terminalFailureReason?: string\n}\n\n/**\n * Project one campaign cell into the canonical run format.\n *\n * A dispatch error establishes terminal execution failure. A judge error only\n * establishes that quality measurement failed after dispatch completed.\n * Failures without a stage remain unknown. No failure becomes a zero-quality\n * label.\n */\nexport function campaignCellToRunRecord<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n options: CampaignCellRunRecordOptions,\n): RunRecord {\n const quality = projectCampaignCellQuality(cell)\n const execution = campaignCellExecutionEvidence(cell)\n const judgeErrorCount = Math.max(\n quality.raw.judge_error_count ?? 0,\n execution.judgeErrorCount ?? 0,\n )\n const costProvenance = campaignCellRunCostProvenance(cell, options.defaultCostUsd)\n const costUsd = costProvenance.usd\n const raw: Record<string, number> = {\n ...finiteMetrics(options.raw),\n ...quality.raw,\n rep: cell.rep,\n duration_ms: cell.durationMs,\n ...(costUsd === null ? {} : { cost_usd: costUsd }),\n tokens_input: cell.tokenUsage.input,\n tokens_output: cell.tokenUsage.output,\n tokens_known: cell.tokenUsage.tokensKnown === false ? 0 : 1,\n latency_ms: cell.durationMs,\n ...(execution.executionErrorCount === undefined\n ? {}\n : { execution_error_count: execution.executionErrorCount }),\n ...(judgeErrorCount > 0 ? { judge_error_count: judgeErrorCount } : {}),\n ...(execution.unclassifiedErrorCount === undefined\n ? {}\n : { unclassified_error_count: execution.unclassifiedErrorCount }),\n }\n if (typeof cell.generation === 'number') raw.generation = cell.generation\n if (cell.tokenUsage.reasoning !== undefined) {\n raw.tokens_reasoning = cell.tokenUsage.reasoning\n }\n if (cell.tokenUsage.cached !== undefined) raw.tokens_cached = cell.tokenUsage.cached\n if (cell.tokenUsage.cacheWrite !== undefined) {\n raw.tokens_cache_write = cell.tokenUsage.cacheWrite\n }\n if (cell.tokenUsage.tokensKnown !== false && costUsd !== null && costUsd > 0) {\n raw.tokens_per_dollar = (cell.tokenUsage.input + cell.tokenUsage.output) / costUsd\n }\n if (costUsd !== null && quality.score !== undefined && quality.score > 0.01) {\n raw.cost_per_quality = costUsd / quality.score\n }\n\n const outcome: RunOutcome = {\n raw,\n ...(quality.judgeScores ? { judgeScores: quality.judgeScores } : {}),\n }\n if (quality.score !== undefined) {\n if (options.splitTag === 'holdout') outcome.holdoutScore = quality.score\n else outcome.searchScore = quality.score\n }\n\n return validateRunRecord({\n runId: options.runId,\n experimentId: options.experimentId,\n candidateId: options.candidateId,\n seed: options.seed ?? cell.seed,\n model: options.model,\n promptHash: options.promptHash,\n configHash: options.configHash,\n commitSha: options.commitSha,\n wallMs: cell.durationMs,\n costUsd,\n costProvenance,\n tokenUsage: { ...cell.tokenUsage },\n terminalOutcome: execution.terminalOutcome,\n ...(execution.terminalFailureReason\n ? { terminalFailureReason: execution.terminalFailureReason }\n : {}),\n outcome,\n splitTag: options.splitTag,\n scenarioId: options.scenarioId ?? cell.scenarioId,\n ...(options.agentProfile ? { agentProfile: options.agentProfile } : {}),\n ...(options.search ? { search: options.search } : {}),\n ...(options.traceRef ? { traceRef: options.traceRef } : {}),\n })\n}\n\n/**\n * A cell's cost as a RunRecord states it. An uncaptured cell's `costUsd` is\n * the spend its receipts proved, so a positive one is the run's floor. The\n * caller's `defaultCostUsd` stands in only when nothing was proved: a proven\n * floor is never replaced by a guess.\n */\nfunction campaignCellRunCostProvenance<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n defaultCostUsd: number | undefined,\n): RunCostProvenance {\n const provenance = campaignCellCostProvenance(cell)\n if (provenance.kind !== 'uncaptured') return provenance\n if (cell.costUsd > 0) return { kind: 'lower-bound', usd: null, knownLowerBoundUsd: cell.costUsd }\n return defaultCostUsd === undefined ? provenance : { kind: 'estimated', usd: defaultCostUsd }\n}\n\n/**\n * Validate the cost fields that cross campaign cache and RunRecord boundaries.\n * `costUsd` is a known subtotal for uncaptured cells, but it must equal the\n * authoritative total whenever that total is observed or estimated.\n */\nexport function campaignCellCostProvenance<TArtifact>(\n cell: Pick<CampaignCellResult<TArtifact>, 'cellId' | 'costUsd' | 'costProvenance'>,\n): CostProvenance {\n if (!Number.isFinite(cell.costUsd) || cell.costUsd < 0) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costUsd`)\n }\n const provenance = cell.costProvenance\n if (!provenance || typeof provenance !== 'object') {\n throw new Error(`campaign cell '${cell.cellId}' has no costProvenance`)\n }\n if (provenance.kind === 'uncaptured') {\n if (provenance.usd !== null) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid uncaptured costProvenance`)\n }\n return { kind: 'uncaptured', usd: null }\n }\n if (\n (provenance.kind !== 'observed' && provenance.kind !== 'estimated') ||\n !Number.isFinite(provenance.usd) ||\n provenance.usd < 0\n ) {\n throw new Error(`campaign cell '${cell.cellId}' has invalid costProvenance`)\n }\n if (provenance.usd !== cell.costUsd) {\n throw new Error(`campaign cell '${cell.cellId}' has costUsd inconsistent with costProvenance`)\n }\n return { kind: provenance.kind, usd: provenance.usd }\n}\n\nexport function campaignCellExecutionEvidence<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellExecutionEvidence {\n if (cell.errorStage === 'dispatch') {\n return {\n terminalOutcome: 'failed',\n executionErrorCount: 1,\n ...(cell.error ? { terminalFailureReason: cell.error } : {}),\n }\n }\n if (cell.errorStage === 'judge') {\n return {\n terminalOutcome: 'succeeded',\n executionErrorCount: 0,\n judgeErrorCount: 1,\n }\n }\n if (!cell.error) {\n return { terminalOutcome: 'succeeded', executionErrorCount: 0 }\n }\n return {\n terminalOutcome: 'unknown',\n unclassifiedErrorCount: 1,\n }\n}\n\n/**\n * Produce the only task-quality view used by campaign aggregates and exports.\n *\n * Successful judge results remain available for diagnosis after another judge\n * fails, but a task score exists only for an error-free cell whose reported\n * judge values are all finite.\n */\nexport function projectCampaignCellQuality<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): CampaignCellQualityProjection {\n if (cell.errorStage === 'dispatch') {\n return { successfulJudgeScores: {}, failedJudges: [], raw: {} }\n }\n\n const perJudge: Record<string, Record<string, number>> = {}\n const successfulJudgeScores: Record<string, JudgeScore> = {}\n const dimensionValues = new Map<string, number[]>()\n const composites: number[] = []\n const notes: string[] = []\n const failedJudges = new Set<string>(\n cell.errorStage === 'judge' ? [cell.errorJudge ?? 'unknown-judge'] : [],\n )\n const raw: Record<string, number> = {}\n\n for (const [judgeName, score] of Object.entries(cell.judgeScores)) {\n const dimensionsShape = (score as { dimensions?: unknown }).dimensions\n if (\n typeof dimensionsShape !== 'object' ||\n dimensionsShape === null ||\n Array.isArray(dimensionsShape)\n ) {\n throw new CampaignJudgeScoreShapeError(judgeName)\n }\n const finiteDimensions = Object.values(score.dimensions).every(Number.isFinite)\n if (score.failed || !Number.isFinite(score.composite) || !finiteDimensions) {\n failedJudges.add(judgeName)\n continue\n }\n\n composites.push(score.composite)\n successfulJudgeScores[judgeName] = score\n const dimensions = { ...score.dimensions }\n perJudge[judgeName] = dimensions\n for (const [dimension, value] of Object.entries(dimensions)) {\n raw[`${judgeName}.${dimension}`] = value\n const values = dimensionValues.get(dimension) ?? []\n values.push(value)\n dimensionValues.set(dimension, values)\n }\n if (score.notes) notes.push(`${judgeName}: ${score.notes}`)\n for (const failedJudge of score.failedJudges ?? []) {\n failedJudges.add(`${judgeName}/${failedJudge}`)\n }\n }\n\n if (failedJudges.size > 0) raw.judge_error_count = failedJudges.size\n const sortedFailedJudges = [...failedJudges].sort()\n if (composites.length === 0) {\n return {\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n raw,\n }\n }\n\n const composite = mean(composites)\n const perDimMean = Object.fromEntries(\n [...dimensionValues.entries()].map(([dimension, values]) => [dimension, mean(values)]),\n )\n const complete =\n cell.error === undefined && cell.errorStage === undefined && failedJudges.size === 0\n if (complete) raw.composite = composite\n\n return {\n ...(complete ? { score: composite } : {}),\n raw,\n successfulJudgeScores,\n failedJudges: sortedFailedJudges,\n judgeScores: {\n perJudge,\n perDimMean,\n composite,\n ...(sortedFailedJudges.length > 0 ? { failedJudges: sortedFailedJudges } : {}),\n ...(notes.length > 0 ? { notes: notes.join(' | ') } : {}),\n },\n }\n}\n\n/** Read the canonical task score without recomputing cell quality. */\nexport function campaignCellTaskScore<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): number | undefined {\n return projectCampaignCellQuality(cell).score\n}\n\n/** Read canonical successful judge dimensions without recomputing cell quality. */\nexport function campaignCellJudgeDimensions<TArtifact>(\n cell: CampaignCellResult<TArtifact>,\n): Record<string, Record<string, number>> {\n return projectCampaignCellQuality(cell).judgeScores?.perJudge ?? {}\n}\n\nfunction finiteMetrics(metrics: Record<string, number> | undefined): Record<string, number> {\n const finite: Record<string, number> = {}\n for (const [key, value] of Object.entries(metrics ?? {})) {\n if (Number.isFinite(value)) finite[key] = value\n }\n return finite\n}\n\nfunction mean(values: number[]): number {\n return values.reduce((sum, value) => sum + value, 0) / values.length\n}\n"],"mappings":";;;;;;;;;;;;AAsDA,IAAa,+BAAb,cAAkD,UAAU;CAC1D,YAAY,WAAmB;EAC7B,MACE,wBAAwB,UAAU,mQAIpC;EACA,KAAK,OAAO;CACd;AACF;;;;;;;;;AAkBA,SAAgB,wBACd,MACA,SACW;CACX,MAAM,UAAU,2BAA2B,IAAI;CAC/C,MAAM,YAAY,8BAA8B,IAAI;CACpD,MAAM,kBAAkB,KAAK,IAC3B,QAAQ,IAAI,qBAAqB,GACjC,UAAU,mBAAmB,CAC/B;CACA,MAAM,iBAAiB,8BAA8B,MAAM,QAAQ,cAAc;CACjF,MAAM,UAAU,eAAe;CAC/B,MAAM,MAA8B;EAClC,GAAG,cAAc,QAAQ,GAAG;EAC5B,GAAG,QAAQ;EACX,KAAK,KAAK;EACV,aAAa,KAAK;EAClB,GAAI,YAAY,OAAO,CAAC,IAAI,EAAE,UAAU,QAAQ;EAChD,cAAc,KAAK,WAAW;EAC9B,eAAe,KAAK,WAAW;EAC/B,cAAc,KAAK,WAAW,gBAAgB,QAAQ,IAAI;EAC1D,YAAY,KAAK;EACjB,GAAI,UAAU,wBAAwB,KAAA,IAClC,CAAC,IACD,EAAE,uBAAuB,UAAU,oBAAoB;EAC3D,GAAI,kBAAkB,IAAI,EAAE,mBAAmB,gBAAgB,IAAI,CAAC;EACpE,GAAI,UAAU,2BAA2B,KAAA,IACrC,CAAC,IACD,EAAE,0BAA0B,UAAU,uBAAuB;CACnE;CACA,IAAI,OAAO,KAAK,eAAe,UAAU,IAAI,aAAa,KAAK;CAC/D,IAAI,KAAK,WAAW,cAAc,KAAA,GAChC,IAAI,mBAAmB,KAAK,WAAW;CAEzC,IAAI,KAAK,WAAW,WAAW,KAAA,GAAW,IAAI,gBAAgB,KAAK,WAAW;CAC9E,IAAI,KAAK,WAAW,eAAe,KAAA,GACjC,IAAI,qBAAqB,KAAK,WAAW;CAE3C,IAAI,KAAK,WAAW,gBAAgB,SAAS,YAAY,QAAQ,UAAU,GACzE,IAAI,qBAAqB,KAAK,WAAW,QAAQ,KAAK,WAAW,UAAU;CAE7E,IAAI,YAAY,QAAQ,QAAQ,UAAU,KAAA,KAAa,QAAQ,QAAQ,KACrE,IAAI,mBAAmB,UAAU,QAAQ;CAG3C,MAAM,UAAsB;EAC1B;EACA,GAAI,QAAQ,cAAc,EAAE,aAAa,QAAQ,YAAY,IAAI,CAAC;CACpE;CACA,IAAI,QAAQ,UAAU,KAAA,GACpB,IAAI,QAAQ,aAAa,WAAW,QAAQ,eAAe,QAAQ;MAC9D,QAAQ,cAAc,QAAQ;CAGrC,OAAO,kBAAkB;EACvB,OAAO,QAAQ;EACf,cAAc,QAAQ;EACtB,aAAa,QAAQ;EACrB,MAAM,QAAQ,QAAQ,KAAK;EAC3B,OAAO,QAAQ;EACf,YAAY,QAAQ;EACpB,YAAY,QAAQ;EACpB,WAAW,QAAQ;EACnB,QAAQ,KAAK;EACb;EACA;EACA,YAAY,EAAE,GAAG,KAAK,WAAW;EACjC,iBAAiB,UAAU;EAC3B,GAAI,UAAU,wBACV,EAAE,uBAAuB,UAAU,sBAAsB,IACzD,CAAC;EACL;EACA,UAAU,QAAQ;EAClB,YAAY,QAAQ,cAAc,KAAK;EACvC,GAAI,QAAQ,eAAe,EAAE,cAAc,QAAQ,aAAa,IAAI,CAAC;EACrE,GAAI,QAAQ,SAAS,EAAE,QAAQ,QAAQ,OAAO,IAAI,CAAC;EACnD,GAAI,QAAQ,WAAW,EAAE,UAAU,QAAQ,SAAS,IAAI,CAAC;CAC3D,CAAC;AACH;;;;;;;AAQA,SAAS,8BACP,MACA,gBACmB;CACnB,MAAM,aAAa,2BAA2B,IAAI;CAClD,IAAI,WAAW,SAAS,cAAc,OAAO;CAC7C,IAAI,KAAK,UAAU,GAAG,OAAO;EAAE,MAAM;EAAe,KAAK;EAAM,oBAAoB,KAAK;CAAQ;CAChG,OAAO,mBAAmB,KAAA,IAAY,aAAa;EAAE,MAAM;EAAa,KAAK;CAAe;AAC9F;;;;;;AAOA,SAAgB,2BACd,MACgB;CAChB,IAAI,CAAC,OAAO,SAAS,KAAK,OAAO,KAAK,KAAK,UAAU,GACnD,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,sBAAsB;CAEtE,MAAM,aAAa,KAAK;CACxB,IAAI,CAAC,cAAc,OAAO,eAAe,UACvC,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wBAAwB;CAExE,IAAI,WAAW,SAAS,cAAc;EACpC,IAAI,WAAW,QAAQ,MACrB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,wCAAwC;EAExF,OAAO;GAAE,MAAM;GAAc,KAAK;EAAK;CACzC;CACA,IACG,WAAW,SAAS,cAAc,WAAW,SAAS,eACvD,CAAC,OAAO,SAAS,WAAW,GAAG,KAC/B,WAAW,MAAM,GAEjB,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,6BAA6B;CAE7E,IAAI,WAAW,QAAQ,KAAK,SAC1B,MAAM,IAAI,MAAM,kBAAkB,KAAK,OAAO,+CAA+C;CAE/F,OAAO;EAAE,MAAM,WAAW;EAAM,KAAK,WAAW;CAAI;AACtD;AAEA,SAAgB,8BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,GAAI,KAAK,QAAQ,EAAE,uBAAuB,KAAK,MAAM,IAAI,CAAC;CAC5D;CAEF,IAAI,KAAK,eAAe,SACtB,OAAO;EACL,iBAAiB;EACjB,qBAAqB;EACrB,iBAAiB;CACnB;CAEF,IAAI,CAAC,KAAK,OACR,OAAO;EAAE,iBAAiB;EAAa,qBAAqB;CAAE;CAEhE,OAAO;EACL,iBAAiB;EACjB,wBAAwB;CAC1B;AACF;;;;;;;;AASA,SAAgB,2BACd,MAC+B;CAC/B,IAAI,KAAK,eAAe,YACtB,OAAO;EAAE,uBAAuB,CAAC;EAAG,cAAc,CAAC;EAAG,KAAK,CAAC;CAAE;CAGhE,MAAM,WAAmD,CAAC;CAC1D,MAAM,wBAAoD,CAAC;CAC3D,MAAM,kCAAkB,IAAI,IAAsB;CAClD,MAAM,aAAuB,CAAC;CAC9B,MAAM,QAAkB,CAAC;CACzB,MAAM,eAAe,IAAI,IACvB,KAAK,eAAe,UAAU,CAAC,KAAK,cAAc,eAAe,IAAI,CAAC,CACxE;CACA,MAAM,MAA8B,CAAC;CAErC,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,KAAK,WAAW,GAAG;EACjE,MAAM,kBAAmB,MAAmC;EAC5D,IACE,OAAO,oBAAoB,YAC3B,oBAAoB,QACpB,MAAM,QAAQ,eAAe,GAE7B,MAAM,IAAI,6BAA6B,SAAS;EAElD,MAAM,mBAAmB,OAAO,OAAO,MAAM,UAAU,CAAC,CAAC,MAAM,OAAO,QAAQ;EAC9E,IAAI,MAAM,UAAU,CAAC,OAAO,SAAS,MAAM,SAAS,KAAK,CAAC,kBAAkB;GAC1E,aAAa,IAAI,SAAS;GAC1B;EACF;EAEA,WAAW,KAAK,MAAM,SAAS;EAC/B,sBAAsB,aAAa;EACnC,MAAM,aAAa,EAAE,GAAG,MAAM,WAAW;EACzC,SAAS,aAAa;EACtB,KAAK,MAAM,CAAC,WAAW,UAAU,OAAO,QAAQ,UAAU,GAAG;GAC3D,IAAI,GAAG,UAAU,GAAG,eAAe;GACnC,MAAM,SAAS,gBAAgB,IAAI,SAAS,KAAK,CAAC;GAClD,OAAO,KAAK,KAAK;GACjB,gBAAgB,IAAI,WAAW,MAAM;EACvC;EACA,IAAI,MAAM,OAAO,MAAM,KAAK,GAAG,UAAU,IAAI,MAAM,OAAO;EAC1D,KAAK,MAAM,eAAe,MAAM,gBAAgB,CAAC,GAC/C,aAAa,IAAI,GAAG,UAAU,GAAG,aAAa;CAElD;CAEA,IAAI,aAAa,OAAO,GAAG,IAAI,oBAAoB,aAAa;CAChE,MAAM,qBAAqB,CAAC,GAAG,YAAY,CAAC,CAAC,KAAK;CAClD,IAAI,WAAW,WAAW,GACxB,OAAO;EACL;EACA,cAAc;EACd;CACF;CAGF,MAAM,YAAY,KAAK,UAAU;CACjC,MAAM,aAAa,OAAO,YACxB,CAAC,GAAG,gBAAgB,QAAQ,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,YAAY,CAAC,WAAW,KAAK,MAAM,CAAC,CAAC,CACvF;CACA,MAAM,WACJ,KAAK,UAAU,KAAA,KAAa,KAAK,eAAe,KAAA,KAAa,aAAa,SAAS;CACrF,IAAI,UAAU,IAAI,YAAY;CAE9B,OAAO;EACL,GAAI,WAAW,EAAE,OAAO,UAAU,IAAI,CAAC;EACvC;EACA;EACA,cAAc;EACd,aAAa;GACX;GACA;GACA;GACA,GAAI,mBAAmB,SAAS,IAAI,EAAE,cAAc,mBAAmB,IAAI,CAAC;GAC5E,GAAI,MAAM,SAAS,IAAI,EAAE,OAAO,MAAM,KAAK,KAAK,EAAE,IAAI,CAAC;EACzD;CACF;AACF;AAgBA,SAAS,cAAc,SAAqE;CAC1F,MAAM,SAAiC,CAAC;CACxC,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,WAAW,CAAC,CAAC,GACrD,IAAI,OAAO,SAAS,KAAK,GAAG,OAAO,OAAO;CAE5C,OAAO;AACT;AAEA,SAAS,KAAK,QAA0B;CACtC,OAAO,OAAO,QAAQ,KAAK,UAAU,MAAM,OAAO,CAAC,IAAI,OAAO;AAChE"}