@tangle-network/agent-eval 0.144.4 → 0.144.6

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 (140) hide show
  1. package/CHANGELOG.md +14 -0
  2. package/dist/{agent-profile-cell-Cw0PVwDr.d.ts → agent-profile-cell-BOP-iA9Q.d.ts} +2 -2
  3. package/dist/{agent-profile-cell-Cw0PVwDr.d.ts.map → agent-profile-cell-BOP-iA9Q.d.ts.map} +1 -1
  4. package/dist/analyst/index.d.ts +366 -19
  5. package/dist/analyst/index.d.ts.map +1 -1
  6. package/dist/analyst/index.js +8 -8
  7. package/dist/{benchmark-CYtcIF2V.js → benchmark-B181aMF9.js} +2 -2
  8. package/dist/{benchmark-CYtcIF2V.js.map → benchmark-B181aMF9.js.map} +1 -1
  9. package/dist/{benchmark-CP6kWfj8.d.ts → benchmark-J9Qe6j2_.d.ts} +3 -3
  10. package/dist/{benchmark-CP6kWfj8.d.ts.map → benchmark-J9Qe6j2_.d.ts.map} +1 -1
  11. package/dist/{benchmark-command-9FTgq6Fg.js → benchmark-command-CQd78YHt.js} +922 -36
  12. package/dist/benchmark-command-CQd78YHt.js.map +1 -0
  13. package/dist/benchmarks/index.d.ts +1 -1
  14. package/dist/benchmarks/index.js +1 -1
  15. package/dist/{benchmarks-CkG1bWFa.js → benchmarks-BEOkuvIg.js} +3 -3
  16. package/dist/{benchmarks-CkG1bWFa.js.map → benchmarks-BEOkuvIg.js.map} +1 -1
  17. package/dist/campaign/index.d.ts +6 -6
  18. package/dist/campaign/index.js +3 -3
  19. package/dist/{campaign-DjGFyPxH.js → campaign-CXsdyym7.js} +4 -4
  20. package/dist/{campaign-DjGFyPxH.js.map → campaign-CXsdyym7.js.map} +1 -1
  21. package/dist/cli.js +2 -2
  22. package/dist/{client-Bbht4xxl.d.ts → client-0JI64ovJ.d.ts} +4 -4
  23. package/dist/{client-Bbht4xxl.d.ts.map → client-0JI64ovJ.d.ts.map} +1 -1
  24. package/dist/{completion-verifier-EJERfFwF.d.ts → completion-verifier-CBiee74w.d.ts} +5 -5
  25. package/dist/{completion-verifier-EJERfFwF.d.ts.map → completion-verifier-CBiee74w.d.ts.map} +1 -1
  26. package/dist/contract/index.d.ts +10 -10
  27. package/dist/contract/index.js +5 -5
  28. package/dist/control.d.ts +2 -2
  29. package/dist/{cost-ledger-FuQvHxPm.d.ts → cost-ledger-Bv_e8XHY.d.ts} +2 -2
  30. package/dist/{cost-ledger-FuQvHxPm.d.ts.map → cost-ledger-Bv_e8XHY.d.ts.map} +1 -1
  31. package/dist/{dataset-v_Y5902-.d.ts → dataset-C8xaLXdY.d.ts} +2 -2
  32. package/dist/{dataset-v_Y5902-.d.ts.map → dataset-C8xaLXdY.d.ts.map} +1 -1
  33. package/dist/{default-registry-SOyHB6qG.js → default-registry-Dta70shL.js} +4 -4
  34. package/dist/{default-registry-SOyHB6qG.js.map → default-registry-Dta70shL.js.map} +1 -1
  35. package/dist/{default-registry-iXfu2trt.d.ts → default-registry-J9m-_tya.d.ts} +5 -5
  36. package/dist/{default-registry-iXfu2trt.d.ts.map → default-registry-J9m-_tya.d.ts.map} +1 -1
  37. package/dist/{dspy-rlm-engine-IRCG8kdi.js → dspy-rlm-engine-19FQEMBK.js} +2 -2
  38. package/dist/{dspy-rlm-engine-IRCG8kdi.js.map → dspy-rlm-engine-19FQEMBK.js.map} +1 -1
  39. package/dist/{errors-DkfjIDvD.d.ts → errors-CKPfb2aH.d.ts} +2 -2
  40. package/dist/{errors-DkfjIDvD.d.ts.map → errors-CKPfb2aH.d.ts.map} +1 -1
  41. package/dist/errors-D-LKuDhb.js.map +1 -1
  42. package/dist/{eval-campaign-lZcDIwQM.js → eval-campaign-CfLQQs9B.js} +2 -2
  43. package/dist/{eval-campaign-lZcDIwQM.js.map → eval-campaign-CfLQQs9B.js.map} +1 -1
  44. package/dist/{exact-types-BygCBR4L.d.ts → exact-types-CBYF5MGd.d.ts} +2 -2
  45. package/dist/{exact-types-BygCBR4L.d.ts.map → exact-types-CBYF5MGd.d.ts.map} +1 -1
  46. package/dist/{external-optimizer-contracts-CdmX2K2S.d.ts → external-optimizer-contracts-Q9c0L3eY.d.ts} +3 -3
  47. package/dist/{external-optimizer-contracts-CdmX2K2S.d.ts.map → external-optimizer-contracts-Q9c0L3eY.d.ts.map} +1 -1
  48. package/dist/{extract-usage-7l1Xq5ti.js → extract-usage-BW27f3XW.js} +2 -2
  49. package/dist/{extract-usage-7l1Xq5ti.js.map → extract-usage-BW27f3XW.js.map} +1 -1
  50. package/dist/{feedback-trajectory-CSIkRLQX.d.ts → feedback-trajectory-GgoS0-MK.d.ts} +4 -4
  51. package/dist/{feedback-trajectory-CSIkRLQX.d.ts.map → feedback-trajectory-GgoS0-MK.d.ts.map} +1 -1
  52. package/dist/fuzz.d.ts +1 -1
  53. package/dist/hosted/index.d.ts +3 -3
  54. package/dist/{index-D_qTihaQ.d.ts → index-4XwggC10.d.ts} +5 -5
  55. package/dist/{index-D_qTihaQ.d.ts.map → index-4XwggC10.d.ts.map} +1 -1
  56. package/dist/{index-CNOCxBLh.d.ts → index-B6-B0zTB.d.ts} +2 -2
  57. package/dist/{index-CNOCxBLh.d.ts.map → index-B6-B0zTB.d.ts.map} +1 -1
  58. package/dist/{index-BuHs_OnD.d.ts → index-BIL5vxxt.d.ts} +12 -12
  59. package/dist/{index-BuHs_OnD.d.ts.map → index-BIL5vxxt.d.ts.map} +1 -1
  60. package/dist/{index-DEb46kc6.d.ts → index-CvSN3IG1.d.ts} +2 -2
  61. package/dist/{index-DEb46kc6.d.ts.map → index-CvSN3IG1.d.ts.map} +1 -1
  62. package/dist/{index-DGIzNtRv.d.ts → index-Dx1kF3Ez.d.ts} +3 -3
  63. package/dist/{index-DGIzNtRv.d.ts.map → index-Dx1kF3Ez.d.ts.map} +1 -1
  64. package/dist/index.d.ts +68 -77
  65. package/dist/index.d.ts.map +1 -1
  66. package/dist/index.js +62 -126
  67. package/dist/index.js.map +1 -1
  68. package/dist/{insight-report-D5m1z0_n.d.ts → insight-report-DqEsugpr.d.ts} +4 -4
  69. package/dist/{insight-report-D5m1z0_n.d.ts.map → insight-report-DqEsugpr.d.ts.map} +1 -1
  70. package/dist/{integrity-DRXobPEs.d.ts → integrity-CNGUaGBY.d.ts} +3 -3
  71. package/dist/{integrity-DRXobPEs.d.ts.map → integrity-CNGUaGBY.d.ts.map} +1 -1
  72. package/dist/{kind-factory-Bvwe3pup.js → kind-factory-BHIgPmzS.js} +2 -2
  73. package/dist/{kind-factory-Bvwe3pup.js.map → kind-factory-BHIgPmzS.js.map} +1 -1
  74. package/dist/ledger-core/index.d.ts +1 -1
  75. package/dist/{llm-client-D3EoChAU.js → llm-client-Dv5BiKLE.js} +289 -7
  76. package/dist/llm-client-Dv5BiKLE.js.map +1 -0
  77. package/dist/meta-eval/index.d.ts +2 -2
  78. package/dist/multishot/index.d.ts +2 -2
  79. package/dist/multishot/index.js +1 -1
  80. package/dist/multishot/index.js.map +1 -1
  81. package/dist/openapi.json +1 -1
  82. package/dist/profile-cell.d.ts +1 -1
  83. package/dist/{release-report-BZRWdq_t.d.ts → release-report-ChOgpIoQ.d.ts} +4 -4
  84. package/dist/{release-report-BZRWdq_t.d.ts.map → release-report-ChOgpIoQ.d.ts.map} +1 -1
  85. package/dist/{replay-CqOsGjzU.js → replay-GW61ezMW.js} +4 -4
  86. package/dist/{replay-CqOsGjzU.js.map → replay-GW61ezMW.js.map} +1 -1
  87. package/dist/{replay-DQ-55DC_.d.ts → replay-Krvb114g.d.ts} +7 -7
  88. package/dist/{replay-DQ-55DC_.d.ts.map → replay-Krvb114g.d.ts.map} +1 -1
  89. package/dist/reporting.d.ts +4 -4
  90. package/dist/{researcher-BSiCoM1s.d.ts → researcher-xLeNcpKX.d.ts} +6 -6
  91. package/dist/{researcher-BSiCoM1s.d.ts.map → researcher-xLeNcpKX.d.ts.map} +1 -1
  92. package/dist/{reward-hacking-DFgkEY4p.d.ts → reward-hacking-RZgnGWlx.d.ts} +2 -2
  93. package/dist/{reward-hacking-DFgkEY4p.d.ts.map → reward-hacking-RZgnGWlx.d.ts.map} +1 -1
  94. package/dist/rl.d.ts +5 -5
  95. package/dist/rl.js +1 -1
  96. package/dist/rollout/index.d.ts +1 -1
  97. package/dist/{rubric-predictive-validity-C4r4Y-q8.d.ts → rubric-predictive-validity-9qAwzkZm.d.ts} +2 -2
  98. package/dist/{rubric-predictive-validity-C4r4Y-q8.d.ts.map → rubric-predictive-validity-9qAwzkZm.d.ts.map} +1 -1
  99. package/dist/{run-evidence-H1vRpIdT.d.ts → run-evidence-C6G41MSI.d.ts} +3 -3
  100. package/dist/{run-evidence-H1vRpIdT.d.ts.map → run-evidence-C6G41MSI.d.ts.map} +1 -1
  101. package/dist/{run-record-ooo9FWns.d.ts → run-record-DdSa93_W.d.ts} +4 -4
  102. package/dist/{run-record-ooo9FWns.d.ts.map → run-record-DdSa93_W.d.ts.map} +1 -1
  103. package/dist/{semantic-concept-judge-Do5aM9wP.js → semantic-concept-judge-DwF6n05O.js} +3 -3
  104. package/dist/{semantic-concept-judge-Do5aM9wP.js.map → semantic-concept-judge-DwF6n05O.js.map} +1 -1
  105. package/dist/{server-Df00sdwz.js → server-D6XJQHw7.js} +2 -2
  106. package/dist/{server-Df00sdwz.js.map → server-D6XJQHw7.js.map} +1 -1
  107. package/dist/{skill-usage-DtpLou9L.d.ts → skill-usage-GlOphAhX.d.ts} +9 -9
  108. package/dist/{skill-usage-DtpLou9L.d.ts.map → skill-usage-GlOphAhX.d.ts.map} +1 -1
  109. package/dist/{skillopt-optimization-method-CvSJGdm3.d.ts → skillopt-optimization-method-7S43rbDB.d.ts} +12 -12
  110. package/dist/{skillopt-optimization-method-CvSJGdm3.d.ts.map → skillopt-optimization-method-7S43rbDB.d.ts.map} +1 -1
  111. package/dist/{skillopt-optimization-method-C4FX42dy.js → skillopt-optimization-method-Bfb-vBKe.js} +2 -2
  112. package/dist/{skillopt-optimization-method-C4FX42dy.js.map → skillopt-optimization-method-Bfb-vBKe.js.map} +1 -1
  113. package/dist/{statistics-B4u_CiFd.d.ts → statistics-C-dm-J6H.d.ts} +2 -2
  114. package/dist/{statistics-B4u_CiFd.d.ts.map → statistics-C-dm-J6H.d.ts.map} +1 -1
  115. package/dist/{store-otlp-D4I90_vR.js → store-otlp-CKtTpRhv.js} +2 -2
  116. package/dist/{store-otlp-D4I90_vR.js.map → store-otlp-CKtTpRhv.js.map} +1 -1
  117. package/dist/{summary-report-BOM6dfP7.d.ts → summary-report-B0cAyA7N.d.ts} +3 -3
  118. package/dist/{summary-report-BOM6dfP7.d.ts.map → summary-report-B0cAyA7N.d.ts.map} +1 -1
  119. package/dist/{tool-groups-CMmsgTzj.d.ts → tool-groups-CK0JCkqO.d.ts} +9 -9
  120. package/dist/{tool-groups-CMmsgTzj.d.ts.map → tool-groups-CK0JCkqO.d.ts.map} +1 -1
  121. package/dist/traces.d.ts +6 -6
  122. package/dist/traces.js +4 -4
  123. package/dist/{types-y8jrxXWd.d.ts → types-BhP9q0Fq.d.ts} +22 -4
  124. package/dist/{types-y8jrxXWd.d.ts.map → types-BhP9q0Fq.d.ts.map} +1 -1
  125. package/dist/{types-DcJxgsLy.d.ts → types-DOZyvsFU.d.ts} +5 -5
  126. package/dist/{types-DcJxgsLy.d.ts.map → types-DOZyvsFU.d.ts.map} +1 -1
  127. package/dist/{types-BjMFz88h.d.ts → types-XMVEdrE_.d.ts} +207 -6
  128. package/dist/types-XMVEdrE_.d.ts.map +1 -0
  129. package/dist/{usage-receipt-CgxMEBZq.js → usage-receipt-EVI8B8Xu.js} +8 -1
  130. package/dist/usage-receipt-EVI8B8Xu.js.map +1 -0
  131. package/dist/wire/index.d.ts +3 -3
  132. package/dist/wire/index.js +1 -1
  133. package/docs/building-doctrine.md +15 -0
  134. package/docs/prime-analyst.md +120 -0
  135. package/docs/trace-analysis.md +1 -1
  136. package/package.json +3 -3
  137. package/dist/benchmark-command-9FTgq6Fg.js.map +0 -1
  138. package/dist/llm-client-D3EoChAU.js.map +0 -1
  139. package/dist/types-BjMFz88h.d.ts.map +0 -1
  140. package/dist/usage-receipt-CgxMEBZq.js.map +0 -1
package/dist/cli.js CHANGED
@@ -1,7 +1,7 @@
1
1
  #!/usr/bin/env node
2
2
  import { o as runRolloutReleaseCli } from "./hf-dataset-XggBupCr.js";
3
- import { n as runAnalystBenchmarkCommand } from "./benchmark-command-9FTgq6Fg.js";
4
- import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-Df00sdwz.js";
3
+ import { n as runAnalystBenchmarkCommand } from "./benchmark-command-CQd78YHt.js";
4
+ import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-D6XJQHw7.js";
5
5
  import { writeFileSync } from "node:fs";
6
6
  //#region src/cli-config.ts
7
7
  function resolveCliLlmConfig(env = process.env) {
@@ -1,6 +1,6 @@
1
- import { l as RunTerminalOutcome } from "./run-record-ooo9FWns.js";
2
- import { _ as GateDecision, j as MutableSurface } from "./types-DcJxgsLy.js";
3
- import { o as InsightReport } from "./insight-report-D5m1z0_n.js";
1
+ import { l as RunTerminalOutcome } from "./run-record-DdSa93_W.js";
2
+ import { _ as GateDecision, j as MutableSurface } from "./types-DOZyvsFU.js";
3
+ import { o as InsightReport } from "./insight-report-DqEsugpr.js";
4
4
  //#region src/hosted/types.d.ts
5
5
  declare const HOSTED_WIRE_VERSION: '2026-07-24.v1';
6
6
  type HostedWireVersion = typeof HOSTED_WIRE_VERSION;
@@ -199,4 +199,4 @@ declare function hostedClientFromEnv(overrides?: Partial<HostedTenant> & {
199
199
  }): HostedClient | undefined;
200
200
  //#endregion
201
201
  export { UnixNanoTimestamp as _, hostedTenantFromEnv as a, EvalRunGenerationSnapshot as c, HostedIngestHeaders as d, HostedWireVersion as f, TraceSpanEvent as g, IngestTracesRequest as h, hostedClientFromEnv as i, EvalRunStatus as l, IngestResponse as m, HostedTenant as n, EvalRunCellScore as o, IngestEvalRunsRequest as p, createHostedClient as r, EvalRunEvent as s, HostedClient as t, HOSTED_WIRE_VERSION as u };
202
- //# sourceMappingURL=client-Bbht4xxl.d.ts.map
202
+ //# sourceMappingURL=client-0JI64ovJ.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"client-Bbht4xxl.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}
1
+ {"version":3,"file":"client-0JI64ovJ.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}
@@ -1,8 +1,8 @@
1
1
  import { t as DefaultVerdict } from "./verdict-Dps8_okt.js";
2
- import { c as CostLedgerHandle } from "./cost-ledger-FuQvHxPm.js";
3
- import { A as ChatClient, _t as RawProviderSink } from "./types-BjMFz88h.js";
4
- import { c as AnalystRunInputs, i as AnalystFinding, l as AnalystRunResult, n as AnalystContext, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-y8jrxXWd.js";
5
- import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-BygCBR4L.js";
2
+ import { c as CostLedgerHandle } from "./cost-ledger-Bv_e8XHY.js";
3
+ import { A as ChatClient, Pt as RawProviderSink } from "./types-XMVEdrE_.js";
4
+ import { c as AnalystRunInputs, i as AnalystFinding, l as AnalystRunResult, n as AnalystContext, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary } from "./types-BhP9q0Fq.js";
5
+ import { i as ExactAnalystRunEvent, o as ExactAnalystRunResult, u as ExactExecutionComponentIdentity } from "./exact-types-CBYF5MGd.js";
6
6
  //#region src/analyst/registry.d.ts
7
7
  interface AnalystHooks {
8
8
  /** Legacy runs may mutate ctx; exact runs provide a frozen observational context. */
@@ -411,4 +411,4 @@ declare function createTokenRecallChecker(opts?: {
411
411
  }): CorrectnessChecker;
412
412
  //#endregion
413
413
  export { ExactAnalystRunExecutionError as A, jsonHasKeys as C, AnalystRegistryOptions as D, AnalystRegistry as E, RegistryRunOpts as M, assertExactRegistryRunOpts as N, BudgetPolicy as O, containsAll as S, AnalystHooks as T, ValidationContext as _, ProducedProposal as a, byteLengthRange as b, SatisfiedBy as c, createLlmCorrectnessChecker as d, createTokenRecallChecker as f, ArtifactValidator as g, Artifact as h, LlmCorrectnessCheckerOpts as i, ExactRegistryRunOpts as j, ExactAnalystBudgetPolicy as k, TaskGold as l, verifyCompletion as m, CompletionVerdict as n, ProducedState as o, parseCorrectnessResponse as p, CorrectnessChecker as r, RequirementCheck as s, CompletionRequirement as t, completionVerdict as u, ValidationIssue as v, regexMatch as w, composeValidators as x, ValidationResult as y };
414
- //# sourceMappingURL=completion-verifier-EJERfFwF.d.ts.map
414
+ //# sourceMappingURL=completion-verifier-CBiee74w.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"completion-verifier-EJERfFwF.d.ts","names":[],"sources":["../src/analyst/registry.ts","../src/artifact-validator.ts","../src/completion-verifier.ts"],"mappings":";;;;;;UAqDiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cAqCE,sCAAsC;WACxC;WACA,QAAQ;EAEjB,YAAY,iBAAiB,QAAQ,uBAAuB,UAAU;;cAU3D;mBACM;mBACA;EAEjB,YAAY,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe;UAIV;UA8BA;UAwFO;UA2aP;UAaA;UAUA;;;iBAiGM,2BAA2B,yBAAyB,SAAS;;;;;;;;;;;;;;;;UC79B5D;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,WAAW;;UAGI;EACf;EACA;;EAEA,iBAAiB;;EAEjB,QAAQ;;UAGO;EACf;EACA;;EAEA;;UAGe;EACf;;EAEA;EACA,QAAQ;;EAER,WAAW;;UAGI;;EAEf;;EAEA;;EAEA,SAAS,UAAU,UAAU,SAAS,oBAAoB,QAAQ;;;;;;iBAWpD,kBACd,YAAY,qBACZ;EAAY;EAAe;IAC1B;;iBAgCa,WAAW,cAAc,SAAS,SAAS;;iBAkB3C,YAAY,cAAc,0BAA0B;;iBAuCpD,gBAAgB,cAAc,aAAa,cAAc;;iBAoBzD,YACd,cACA,oBACA;EAAY;IACX;;;;KCnJS;UAEK;;EAEf;;EAEA;;EAEA;;EAEA,cAAc;;UAGC;EACf;EACA,cAAc;;UAGC;EACf;EACA;EACA;;EAEA;;;UAIe;;EAEf,WAAW;;EAEX,WAAW;;EAEX;;UAGe;EACf;EACA;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;EASA;;EAEA;;EAEA;;;;;UAMe,0BAA0B;EACzC;EACA,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;iBAUc,kBAAkB;EAChC;EACA,cAAc;IACZ;;;;;;KAoCQ,sBACV,aAAa,uBACb,oBACG;EAAU;EAAkB;;;;;;;;;iBAqLX,iBACpB,MAAM,UACN,OAAO,eACP,kBAAkB,qBACjB,QAAQ;UAoFM;EACf;;EAEA,aAAa;EACb;EACA,WAAW;EACX,SAAS;;EAET;;;;;;EAMA;;;;;;EAMA,UAAU;;;;;;;;;iBAUI,yBAAyB;EAAgB;EAAkB;;;;;;;;iBAiC3D,4BACd,MAAM,YACN,OAAM,4BACL;;;;;;;;;;;;;;;iBA6Ia,yBACd;EAAQ;EAAoB;IAC3B"}
1
+ {"version":3,"file":"completion-verifier-CBiee74w.d.ts","names":[],"sources":["../src/analyst/registry.ts","../src/artifact-validator.ts","../src/completion-verifier.ts"],"mappings":";;;;;;UAqDiB;;EAEf,iBAAiB;IACf,SAAS;IACT,KAAK;IACL;aACS;;EAEX,gBAAgB;IACd,SAAS;IACT,SAAS;IACT,UAAU;IACV;aACS;;;;;;EAMX,SAAS;IACP,SAAS;IACT,OAAO;IACP;MACE,+BAA+B,QAAQ;;EAE3C,YAAY;IAAQ,QAAQ;aAA4B;;UAGzC;;EAEf;;EAEA,UAAU;;;;;;;EAOV,YAAY;IACV,SAAS;IACT;IACA;IACA;;;UAIa;;EAEf,OAAO;;EAEP,OAAO,aAAa,SAAS;;EAE7B,QAAQ;;EAER,gBAAgB;;EAEhB,gBAAgB;;EAEhB,eAAe;;UAGA;;EAEf;;EAEA;;EAEA,SAAS;;EAET;;EAEA,SAAS;;EAET,aAAa;;EAEb;;EAEA,OAAO;;;;;;;;;EASP,gBAAgB,cAAc,kBAAkB,eAAe,cAAc;;;;;;EAM7E;;;KAIU;WAEG;WACA;;WAGA;WACA;;WAEA,SAAS,SAAS;;;;;;;;;;UAWhB;WACN;WACA,QAAQ;WACR;WACA,QAAQ;WACR,YAAY;WACZ,oBAAoB;WACpB;WACA,MAAM,SAAS;WACf,eACL,cAAc,kBACd,SAAS,eAAe,cAAc;WAEjC;WACA;WACA;WACA;;;cAqCE,sCAAsC;WACxC;WACA,QAAQ;EAEjB,YAAY,iBAAiB,QAAQ,uBAAuB,UAAU;;cAU3D;mBACM;mBACA;EAEjB,YAAY,UAAS;EAIrB,SAAS,SAAS;EAsBlB,QAAQ;IACN;IACA;IACA;IACA,MAAM;;EAUF,IACJ,eACA,QAAQ,kBACR,UAAS,kBACR,QAAQ;;EAUL,SACJ,eACA,QAAQ,kBACR,SAAS,uBACR,QAAQ;;EAQJ,eACL,eACA,QAAQ,kBACR,SAAS,uBACR,eAAe;;;;;;;;;;;;EAmBX,UACL,eACA,QAAQ,kBACR,UAAS,kBACR,eAAe;UAIV;UA8BA;UAwFO;UA2aP;UAaA;UAUA;;;iBAiGM,2BAA2B,yBAAyB,SAAS;;;;;;;;;;;;;;;;UC79B5D;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,WAAW;;UAGI;EACf;EACA;;EAEA,iBAAiB;;EAEjB,QAAQ;;UAGO;EACf;EACA;;EAEA;;UAGe;EACf;;EAEA;EACA,QAAQ;;EAER,WAAW;;UAGI;;EAEf;;EAEA;;EAEA,SAAS,UAAU,UAAU,SAAS,oBAAoB,QAAQ;;;;;;iBAWpD,kBACd,YAAY,qBACZ;EAAY;EAAe;IAC1B;;iBAgCa,WAAW,cAAc,SAAS,SAAS;;iBAkB3C,YAAY,cAAc,0BAA0B;;iBAuCpD,gBAAgB,cAAc,aAAa,cAAc;;iBAoBzD,YACd,cACA,oBACA;EAAY;IACX;;;;KCnJS;UAEK;;EAEf;;EAEA;;EAEA;;EAEA,cAAc;;UAGC;EACf;EACA,cAAc;;UAGC;EACf;EACA;EACA;;EAEA;;;UAIe;;EAEf,WAAW;;EAEX,WAAW;;EAEX;;UAGe;EACf;EACA;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;EASA;;EAEA;;EAEA;;;;;UAMe,0BAA0B;EACzC;EACA,cAAc;;EAEd;;EAEA;;EAEA;;;;;;;;;iBAUc,kBAAkB;EAChC;EACA,cAAc;IACZ;;;;;;KAoCQ,sBACV,aAAa,uBACb,oBACG;EAAU;EAAkB;;;;;;;;;iBAqLX,iBACpB,MAAM,UACN,OAAO,eACP,kBAAkB,qBACjB,QAAQ;UAoFM;EACf;;EAEA,aAAa;EACb;EACA,WAAW;EACX,SAAS;;EAET;;;;;;EAMA;;;;;;EAMA,UAAU;;;;;;;;;iBAUI,yBAAyB;EAAgB;EAAkB;;;;;;;;iBAiC3D,4BACd,MAAM,YACN,OAAM,4BACL;;;;;;;;;;;;;;;iBA6Ia,yBACd;EAAQ;EAAoB;IAC3B"}
@@ -1,13 +1,13 @@
1
- import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from "../cost-ledger-FuQvHxPm.js";
2
- import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-ooo9FWns.js";
3
- import { A as ChatClient, F as CreateChatClientOpts, V as createChatClient } from "../types-BjMFz88h.js";
4
- import { h as ProposalFindingOrigin, i as AnalystFinding, m as ProposalFinding, v as makeProposalFinding } from "../types-y8jrxXWd.js";
5
- import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-iXfu2trt.js";
6
- import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-DcJxgsLy.js";
7
- import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-D5m1z0_n.js";
8
- import { A as summarizeExecution, D as ExecutionReport, E as AnalyzeRunsOptions, O as SummarizeExecutionOptions, k as analyzeRuns } from "../tool-groups-CMmsgTzj.js";
9
- import { A as RunEvalOptions, An as LlmJudgeDimension, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, D as RunOptimizationOptions, Dn as ReferenceEquivalenceScenario, E as PremeasuredOptimizationBaseline, En as ReferenceEquivalenceJudgeResult, F as GepaOptimizationMethodConfig, Ft as ExternalTextOptimizerContext, G as ObjectiveSource, Gt as OptimizationMethodProvenance, H as AxisVerdict, Ht as OptimizationMethodComparison, I as GepaOptimizationRecipe, It as ExternalTextOptimizerResult, J as PromotionPolicy, K as ParetoSignificanceGateOptions, Kt as OptimizationMethodResult, L as GepaRunnerCommand, Lt as ExternalOptimizationExample, M as GepaAdaptiveEngineRun, Mn as llmJudge, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as createReferenceEquivalenceJudge, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, Qt as compareOptimizationMethods, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, T as runImprovementLoop, Tn as ReferenceEquivalenceJudgeOptions, U as BuildEvidenceVectorOptions, Ut as OptimizationMethodInput, V as AxisEvidence, Vt as OptimizationMethod, W as EvidenceVector, X as paretoPolicy, Xt as OptimizationTokenUsage, Y as buildEvidenceVector, Yt as OptimizationPackageSource, Z as paretoSignificanceGate, _n as campaignSplitDigest, an as RunCampaignOptions, dt as DefaultProductionGateCheck, fn as fsCampaignStorage, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, j as runEval, jn as LlmJudgeOptions, kn as runReferenceEquivalenceJudge, lt as HeldOutGateOptions, mt as defaultProductionGate, n as SkillOptRunnerCommand, ot as PowerPreflight, p as LoopProvenanceRecord, pn as inMemoryCampaignStorage, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as runCampaign, t as SkillOptOptimizationMethodConfig, tn as CampaignCellFailureReceipt, un as CampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceJudgeInput, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-CvSJGdm3.js";
10
- import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-Bbht4xxl.js";
1
+ import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from "../cost-ledger-Bv_e8XHY.js";
2
+ import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-DdSa93_W.js";
3
+ import { A as ChatClient, F as CreateChatClientOpts, V as createChatClient } from "../types-XMVEdrE_.js";
4
+ import { h as ProposalFindingOrigin, i as AnalystFinding, m as ProposalFinding, v as makeProposalFinding } from "../types-BhP9q0Fq.js";
5
+ import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-J9m-_tya.js";
6
+ import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-DOZyvsFU.js";
7
+ import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-DqEsugpr.js";
8
+ import { A as summarizeExecution, D as ExecutionReport, E as AnalyzeRunsOptions, O as SummarizeExecutionOptions, k as analyzeRuns } from "../tool-groups-CK0JCkqO.js";
9
+ import { A as RunEvalOptions, An as LlmJudgeDimension, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, D as RunOptimizationOptions, Dn as ReferenceEquivalenceScenario, E as PremeasuredOptimizationBaseline, En as ReferenceEquivalenceJudgeResult, F as GepaOptimizationMethodConfig, Ft as ExternalTextOptimizerContext, G as ObjectiveSource, Gt as OptimizationMethodProvenance, H as AxisVerdict, Ht as OptimizationMethodComparison, I as GepaOptimizationRecipe, It as ExternalTextOptimizerResult, J as PromotionPolicy, K as ParetoSignificanceGateOptions, Kt as OptimizationMethodResult, L as GepaRunnerCommand, Lt as ExternalOptimizationExample, M as GepaAdaptiveEngineRun, Mn as llmJudge, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as createReferenceEquivalenceJudge, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, Qt as compareOptimizationMethods, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, T as runImprovementLoop, Tn as ReferenceEquivalenceJudgeOptions, U as BuildEvidenceVectorOptions, Ut as OptimizationMethodInput, V as AxisEvidence, Vt as OptimizationMethod, W as EvidenceVector, X as paretoPolicy, Xt as OptimizationTokenUsage, Y as buildEvidenceVector, Yt as OptimizationPackageSource, Z as paretoSignificanceGate, _n as campaignSplitDigest, an as RunCampaignOptions, dt as DefaultProductionGateCheck, fn as fsCampaignStorage, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, j as runEval, jn as LlmJudgeOptions, kn as runReferenceEquivalenceJudge, lt as HeldOutGateOptions, mt as defaultProductionGate, n as SkillOptRunnerCommand, ot as PowerPreflight, p as LoopProvenanceRecord, pn as inMemoryCampaignStorage, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as runCampaign, t as SkillOptOptimizationMethodConfig, tn as CampaignCellFailureReceipt, un as CampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceJudgeInput, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-7S43rbDB.js";
10
+ import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-0JI64ovJ.js";
11
11
  import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
12
12
  import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementCost, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
13
13
  //#region src/contract/self-improve.d.ts
@@ -1,16 +1,16 @@
1
1
  import { c as ValidationError } from "../errors-D-LKuDhb.js";
2
2
  import { i as estimateCost, o as isModelPriced } from "../metrics-C9YY1OcL.js";
3
3
  import { i as CostLedger } from "../cost-ledger-DMFxsLKr.js";
4
- import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../default-registry-SOyHB6qG.js";
4
+ import { C as createChatClient, t as buildDefaultAnalystRegistry } from "../default-registry-Dta70shL.js";
5
5
  import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js";
6
- import { nt as classifyOtlpSpanRole, rt as isOtlpModelCall } from "../kind-factory-Bvwe3pup.js";
7
- import { s as makeProposalFinding } from "../usage-receipt-CgxMEBZq.js";
6
+ import { nt as classifyOtlpSpanRole, rt as isOtlpModelCall } from "../kind-factory-BHIgPmzS.js";
7
+ import { s as makeProposalFinding } from "../usage-receipt-EVI8B8Xu.js";
8
8
  import { i as inMemoryCampaignStorage, n as createRunCostLedger, r as fsCampaignStorage } from "../single-run-lock-D5iN0Xzb.js";
9
9
  import { m as mapConcurrentRange } from "../ledger-core-DXZIqu17.js";
10
- import { A as assertOptimizationResult, B as surfaceContentHash, Ct as decidePairedPromotion, M as compareOptimizationMethods, O as composeGate, S as defaultProductionGate, T as heldoutSignificance, V as surfaceHash, X as resolveRunDir, Y as runCampaign, _ as buildEvidenceVector, a as emitLoopProvenance, at as REFERENCE_EQUIVALENCE_JUDGE_VERSION, b as powerPreflight, d as runImprovementLoop, g as gepaOptimizationMethod, gt as llmJudge, it as REFERENCE_EQUIVALENCE_INPUT_LIMITS, k as externalTextOptimizationMethod, o as loopProvenanceArgsFromResult, ot as createReferenceEquivalenceJudge, p as runEval, st as runReferenceEquivalenceJudge, t as skillOptOptimizationMethod, tt as campaignSplitDigest, v as paretoPolicy, x as heldOutGate, y as paretoSignificanceGate } from "../skillopt-optimization-method-C4FX42dy.js";
10
+ import { A as assertOptimizationResult, B as surfaceContentHash, Ct as decidePairedPromotion, M as compareOptimizationMethods, O as composeGate, S as defaultProductionGate, T as heldoutSignificance, V as surfaceHash, X as resolveRunDir, Y as runCampaign, _ as buildEvidenceVector, a as emitLoopProvenance, at as REFERENCE_EQUIVALENCE_JUDGE_VERSION, b as powerPreflight, d as runImprovementLoop, g as gepaOptimizationMethod, gt as llmJudge, it as REFERENCE_EQUIVALENCE_INPUT_LIMITS, k as externalTextOptimizationMethod, o as loopProvenanceArgsFromResult, ot as createReferenceEquivalenceJudge, p as runEval, st as runReferenceEquivalenceJudge, t as skillOptOptimizationMethod, tt as campaignSplitDigest, v as paretoPolicy, x as heldOutGate, y as paretoSignificanceGate } from "../skillopt-optimization-method-Bfb-vBKe.js";
11
11
  import { E as pairedBootstrap } from "../statistics-ByxzSiOM.js";
12
12
  import { i as parseRunRecordSafe, r as modelHasSnapshot } from "../run-record-CWN8-VsV.js";
13
- import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-7l1Xq5ti.js";
13
+ import { a as recordAggregateMeasurements, i as readTaskFailureLabels, o as summarizeExecutionMeasurements, s as summarizeTraceErrors, t as extractUsage } from "../extract-usage-BW27f3XW.js";
14
14
  import { n as summarizeExecution, t as analyzeRuns } from "../analyze-runs-DWIvOAGk.js";
15
15
  import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "../reward-hacking-CyuzxKly.js";
16
16
  import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
package/dist/control.d.ts CHANGED
@@ -1,3 +1,3 @@
1
- import { $ as ControlRunResult, J as ControlActionOutcome, Q as ControlEvalResult, X as ControlContext, Y as ControlBudget, Z as ControlDecision, at as StopDecision, ct as runAgentControlLoop, dt as subjectiveEval, et as ControlRuntimeConfig, it as ControlStopPolicies, lt as stopOnNoProgress, nt as ControlSeverity, ot as allCriticalPassed, q as ControlActionFailureMode, rt as ControlStep, st as objectiveEval, tt as ControlRuntimeError, ut as stopOnRepeatedAction } from "./feedback-trajectory-CSIkRLQX.js";
2
- import { A as ActionExecutionPolicy, M as evaluateActionPolicy, _ as ProposeReviewReport, a as ProposeReviewControlAction, c as ProposeReviewControlState, g as ProposeReviewConfig, i as scoreFromEvals, j as ActionPolicyDecision, k as runProposeReview, n as RunEvidenceMetadata, o as ProposeReviewControlConfig, r as controlRunToRunRecord, s as ProposeReviewControlResult, t as ControlRunToRunRecordOptions, u as runProposeReviewAsControlLoop } from "./run-evidence-H1vRpIdT.js";
1
+ import { $ as ControlRunResult, J as ControlActionOutcome, Q as ControlEvalResult, X as ControlContext, Y as ControlBudget, Z as ControlDecision, at as StopDecision, ct as runAgentControlLoop, dt as subjectiveEval, et as ControlRuntimeConfig, it as ControlStopPolicies, lt as stopOnNoProgress, nt as ControlSeverity, ot as allCriticalPassed, q as ControlActionFailureMode, rt as ControlStep, st as objectiveEval, tt as ControlRuntimeError, ut as stopOnRepeatedAction } from "./feedback-trajectory-GgoS0-MK.js";
2
+ import { A as ActionExecutionPolicy, M as evaluateActionPolicy, _ as ProposeReviewReport, a as ProposeReviewControlAction, c as ProposeReviewControlState, g as ProposeReviewConfig, i as scoreFromEvals, j as ActionPolicyDecision, k as runProposeReview, n as RunEvidenceMetadata, o as ProposeReviewControlConfig, r as controlRunToRunRecord, s as ProposeReviewControlResult, t as ControlRunToRunRecordOptions, u as runProposeReviewAsControlLoop } from "./run-evidence-C6G41MSI.js";
3
3
  export { type ActionExecutionPolicy, type ActionPolicyDecision, type ControlActionFailureMode, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRunToRunRecordOptions, type ControlRuntimeConfig, type ControlRuntimeError, type ControlSeverity, type ControlStep, type ControlStopPolicies, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type RunEvidenceMetadata, type StopDecision, allCriticalPassed, controlRunToRunRecord, evaluateActionPolicy, objectiveEval, runAgentControlLoop, runProposeReview, runProposeReviewAsControlLoop, scoreFromEvals, stopOnNoProgress, stopOnRepeatedAction, subjectiveEval };
@@ -1,4 +1,4 @@
1
- import { l as ValidationError } from "./errors-DkfjIDvD.js";
1
+ import { l as ValidationError } from "./errors-CKPfb2aH.js";
2
2
  //#region src/cost-ledger.d.ts
3
3
  type CostChannel = 'agent' | 'judge' | 'verifier' | 'analyst' | 'driver' | (string & {});
4
4
  /** How a USD amount was obtained. */
@@ -259,4 +259,4 @@ declare function costForUsage(model: string, usage: CostUsage): CostResult;
259
259
  declare function costForTokenPricing(pricing: CustomTokenPricing, usage: Pick<CostUsage, 'inputTokens' | 'outputTokens' | 'cachedTokens' | 'cacheWriteTokens'>): number;
260
260
  //#endregion
261
261
  export { PendingCostCall as C, costForUsage as D, costForTokenPricing as E, modelPriceKey as O, PaidCallResult as S, RunPaidCallInput as T, CostReservationExceededError as _, CostChannel as a, CustomTokenPricing as b, CostLedgerHandle as c, CostLedgerPersistenceError as d, CostLedgerSummary as f, CostReceiptInput as g, CostReceiptCaptureError as h, CostCeilingReachedError as i, CostLedgerOptions as l, CostReceipt as m, CostAccountingIncompleteError as n, CostLedger as o, CostProvenance as p, CostCallConflictError as r, CostLedgerFilter as s, ChannelRollup as t, CostLedgerPersistence as u, CostResult as v, PendingCostCallView as w, MaximumCharge as x, CostUsage as y };
262
- //# sourceMappingURL=cost-ledger-FuQvHxPm.d.ts.map
262
+ //# sourceMappingURL=cost-ledger-Bv_e8XHY.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"cost-ledger-FuQvHxPm.d.ts","names":[],"sources":["../src/cost-ledger.ts"],"mappings":";;KAIY;;KAGA;EACN;EAAkB;;EAClB;EAAmB;;EACnB;EAAoB;;UAET;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGQ;EACR;EACA,SAAS;EACT;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,wBAAwB;EACvC;;UAGe,4BAA4B;EAC3C;;UAGe,oBAAoB,cAAc;EACjD;EACA;EACA;EACA;;EAEA;IACE;IACA;IACA;IACA;;;EAGF;;EAEA;;;EAGA;;UAKe,yBAAyB;EACxC;;EAEA,qBAAqB;EACrB;;EAEA;EACA;EACA;;;UAIe;;EAEf;;EAEA;;EAEA;EACA;;KAGU;EACN;;EAEA,oBAAoB;IAClB,KAAK;EACN;IAAkB;UAER,iBAAiB;EAChC;EACA,SAAS;EACT;EACA;;EAEA;EACA,OAAO;EACP,SAAS;;EAET,gBAAgB;;EAEhB,QAAQ,QAAQ,aAAa,iBAAiB,QAAQ;EACtD,QAAQ,OAAO,IAAI;EACnB,kBAAkB,OAAO,QAAQ;;KAGvB,eAAe;EACrB;EAAiB;EAAgB,OAAO;EAAG,SAAS;;EACpD;EAAkB;EAAiB,OAAO;EAAO,UAAU;;UAEhD;EACf,SAAS;EACT;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,gBAAgB;EAChB,WAAW;EACX;EACA;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA,OAAO;;UAGQ;;EAEf;;EAEA,SAAS;;;UAIM;EACf;IAAU;IAAkB;;EAC5B,OAAO,0BAA0B;;UAGlB;EACf;EACA,cAAc;;EAEd,oBAAoB;;cAGT,gCAAgC;EAC3C,YACE,oBACA,sBACA,sBACA,eACA;;cAQS,sCAAsC;cAEtC,qCAAqC;EAChD,YAAY,eAAe,mBAAmB;;cAOnC,8BAA8B;WAChC;WACA,UAAU;EAEnB,YACE,iBACA;IAAW;IAAiB,UAAU;IAAa;;;cAQ1C,mCAAmC;WACrC;WACA,UAAU;EAEnB,YAAY,gBAAgB,iBAAiB,UAAU;;cAU5C,gCAAgC;WAClC;WACA,UAAU;WACV,cAAc;EAEvB,YAAY,gBAAgB,gBAAgB,uBAAuB,UAAU;;;cAgClE;mBACM;mBACA;mBACA;mBACA;UACT;UACA;UACA;WACC;mBACQ;EAEjB,YAAY,iBAAiB;EAwDvB,YAAY,GAAG,OAAO,iBAAiB,KAAK,QAAQ,eAAe;;EA2GnE,YAAY,UAAS,wBAA6B;;EA0BxD,UACE,gBACA,UAAU,kBACV;IAAW;IAAkB;MAC5B;EAkBH,KAAK,SAAS,mBAAmB;;EAQjC,YAAY,SAAS,mBAAmB;EAcxC,QAAQ,SAAS,mBAAmB;EAkFpC,cAAc;EAYd;UAIc;UAoDN;UAsCA;UAKM;UAoBN;UAmBA;UAkBA;UAcA;UA+BA;;;;;;UAcA;UASA;UAYA;UAMA;;;;;;;;KA0BE,mBAAmB,KAC7B,YACA,cAAc,8CAEd,QAAQ,KAAK;;iBAGC,cAAc;UAIb;EACf;EACA;;iBAGc,aAAa,eAAe,OAAO,YAAY;;iBAc/C,oBACd,SAAS,oBACT,OAAO,KAAK"}
1
+ {"version":3,"file":"cost-ledger-Bv_e8XHY.d.ts","names":[],"sources":["../src/cost-ledger.ts"],"mappings":";;KAIY;;KAGA;EACN;EAAkB;;EAClB;EAAmB;;EACnB;EAAoB;;UAET;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;UAGQ;EACR;EACA,SAAS;EACT;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,wBAAwB;EACvC;;UAGe,4BAA4B;EAC3C;;UAGe,oBAAoB,cAAc;EACjD;EACA;EACA;EACA;;EAEA;IACE;IACA;IACA;IACA;;;EAGF;;EAEA;;;EAGA;;UAKe,yBAAyB;EACxC;;EAEA,qBAAqB;EACrB;;EAEA;EACA;EACA;;;UAIe;;EAEf;;EAEA;;EAEA;EACA;;KAGU;EACN;;EAEA,oBAAoB;IAClB,KAAK;EACN;IAAkB;UAER,iBAAiB;EAChC;EACA,SAAS;EACT;EACA;;EAEA;EACA,OAAO;EACP,SAAS;;EAET,gBAAgB;;EAEhB,QAAQ,QAAQ,aAAa,iBAAiB,QAAQ;EACtD,QAAQ,OAAO,IAAI;EACnB,kBAAkB,OAAO,QAAQ;;KAGvB,eAAe;EACrB;EAAiB;EAAgB,OAAO;EAAG,SAAS;;EACpD;EAAkB;EAAiB,OAAO;EAAO,UAAU;;UAEhD;EACf,SAAS;EACT;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,gBAAgB;EAChB,WAAW;EACX;EACA;EACA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA,OAAO;;UAGQ;;EAEf;;EAEA,SAAS;;;UAIM;EACf;IAAU;IAAkB;;EAC5B,OAAO,0BAA0B;;UAGlB;EACf;EACA,cAAc;;EAEd,oBAAoB;;cAGT,gCAAgC;EAC3C,YACE,oBACA,sBACA,sBACA,eACA;;cAQS,sCAAsC;cAEtC,qCAAqC;EAChD,YAAY,eAAe,mBAAmB;;cAOnC,8BAA8B;WAChC;WACA,UAAU;EAEnB,YACE,iBACA;IAAW;IAAiB,UAAU;IAAa;;;cAQ1C,mCAAmC;WACrC;WACA,UAAU;EAEnB,YAAY,gBAAgB,iBAAiB,UAAU;;cAU5C,gCAAgC;WAClC;WACA,UAAU;WACV,cAAc;EAEvB,YAAY,gBAAgB,gBAAgB,uBAAuB,UAAU;;;cAgClE;mBACM;mBACA;mBACA;mBACA;UACT;UACA;UACA;WACC;mBACQ;EAEjB,YAAY,iBAAiB;EAwDvB,YAAY,GAAG,OAAO,iBAAiB,KAAK,QAAQ,eAAe;;EA2GnE,YAAY,UAAS,wBAA6B;;EA0BxD,UACE,gBACA,UAAU,kBACV;IAAW;IAAkB;MAC5B;EAkBH,KAAK,SAAS,mBAAmB;;EAQjC,YAAY,SAAS,mBAAmB;EAcxC,QAAQ,SAAS,mBAAmB;EAkFpC,cAAc;EAYd;UAIc;UAoDN;UAsCA;UAKM;UAoBN;UAmBA;UAkBA;UAcA;UA+BA;;;;;;UAcA;UASA;UAYA;UAMA;;;;;;;;KA0BE,mBAAmB,KAC7B,YACA,cAAc,8CAEd,QAAQ,KAAK;;iBAGC,cAAc;UAIb;EACf;EACA;;iBAGc,aAAa,eAAe,OAAO,YAAY;;iBAc/C,oBACd,SAAS,oBACT,OAAO,KAAK"}
@@ -1,4 +1,4 @@
1
- import { l as ValidationError } from "./errors-DkfjIDvD.js";
1
+ import { l as ValidationError } from "./errors-CKPfb2aH.js";
2
2
  //#region src/dataset.d.ts
3
3
  /**
4
4
  * Dataset — versioned, sliceable, content-hashed scenario collection.
@@ -112,4 +112,4 @@ declare class Dataset {
112
112
  declare function hashScenarios(scenarios: DatasetScenario[]): Promise<string>;
113
113
  //#endregion
114
114
  export { DatasetScenario as a, SliceOptions as c, DatasetProvenance as i, hashScenarios as l, DatasetDifficulty as n, DatasetSplit as o, DatasetManifest as r, HoldoutLockedError as s, Dataset as t };
115
- //# sourceMappingURL=dataset-v_Y5902-.d.ts.map
115
+ //# sourceMappingURL=dataset-C8xaLXdY.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"dataset-v_Y5902-.d.ts","names":[],"sources":["../src/dataset.ts"],"mappings":";;;;;;;;;;;;;;;;;;KAiBY;KACA;UAEK;EACf;;EAEA;EACA,QAAQ;EACR,aAAa;;EAEb;;;;;;;;;;EAUA;EACA,OAAO;;UAGQ;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA,YAAY;;EAEZ;EACA;EACA,aAAa,OAAO;;UAGL;EACf,QAAQ;EACR,aAAa;;EAEb;EACA;;EAEA,UAAU,UAAU;;EAEpB;;;cAMW,2BAA2B;EACtC,YAAY;;cAOD;WACF;WACA,YAAY;UACb;UACA;EAER,YAAY;IACV;IACA,YAAY;IACZ,WAAW;IACX;;;EASF,gBAAgB;MAIZ;;;;;EAQJ,MAAM,UAAS,eAAoB;;;;;EAqB7B,YAAY,QAAQ;;EAgB1B,MAAM,YAAW;IAAU;IAAc;OAA0B;EAWnE;EAIA,IAAI,UAAU;EAQd,OAAO;;;;;EAWP;SAQO,UACL,eACA,UAAU,KAAK,oEACd;;iBAaiB,cAAc,WAAW,oBAAoB"}
1
+ {"version":3,"file":"dataset-C8xaLXdY.d.ts","names":[],"sources":["../src/dataset.ts"],"mappings":";;;;;;;;;;;;;;;;;;KAiBY;KACA;UAEK;EACf;;EAEA;EACA,QAAQ;EACR,aAAa;;EAEb;;;;;;;;;;EAUA;EACA,OAAO;;UAGQ;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA,YAAY;;EAEZ;EACA;EACA,aAAa,OAAO;;UAGL;EACf,QAAQ;EACR,aAAa;;EAEb;EACA;;EAEA,UAAU,UAAU;;EAEpB;;;cAMW,2BAA2B;EACtC,YAAY;;cAOD;WACF;WACA,YAAY;UACb;UACA;EAER,YAAY;IACV;IACA,YAAY;IACZ,WAAW;IACX;;;EASF,gBAAgB;MAIZ;;;;;EAQJ,MAAM,UAAS,eAAoB;;;;;EAqB7B,YAAY,QAAQ;;EAgB1B,MAAM,YAAW;IAAU;IAAc;OAA0B;EAWnE;EAIA,IAAI,UAAU;EAQd,OAAO;;;;;EAWP;SAQO,UACL,eACA,UAAU,KAAK,oEACd;;iBAaiB,cAAc,WAAW,oBAAoB"}
@@ -1,8 +1,8 @@
1
- import { n as LlmClient } from "./llm-client-D3EoChAU.js";
1
+ import { n as LlmClient } from "./llm-client-Dv5BiKLE.js";
2
2
  import { LLM_CONTEXT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_OUTPUT_TOKEN_ATTR_KEYS, TOOL_NAME_ATTR_KEYS } from "./trace-attributes.js";
3
3
  import { t as executionTrackByLane } from "./execution-tracks-CpgFPpS5.js";
4
- import { $ as spanEpochMillis, H as snapshotExactExecutionPlan, R as findingSubjectGrammarPromptFor, U as deepFreezeCanonicalJson, V as snapshotExactExecutionComponentIdentity, t as createTraceAnalyst } from "./kind-factory-Bvwe3pup.js";
5
- import { i as validateUsageSettlementTimeout, o as makeFinding, t as assertValidAnalystUsageReceipt } from "./usage-receipt-CgxMEBZq.js";
4
+ import { $ as spanEpochMillis, H as snapshotExactExecutionPlan, R as findingSubjectGrammarPromptFor, U as deepFreezeCanonicalJson, V as snapshotExactExecutionComponentIdentity, t as createTraceAnalyst } from "./kind-factory-BHIgPmzS.js";
5
+ import { i as validateUsageSettlementTimeout, o as makeFinding, t as assertValidAnalystUsageReceipt } from "./usage-receipt-EVI8B8Xu.js";
6
6
  import { i as hashCanonical, r as canonicalString } from "./canonical-D011XM8r.js";
7
7
  import { t as analyzeSupervisorRunIntegrity } from "./integrity-DY6tIbl0.js";
8
8
  import { o as combineAbortSignals } from "./proposal-findings-2GIUo1et.js";
@@ -2449,4 +2449,4 @@ function buildDefaultAnalystRegistry(options = {}) {
2449
2449
  //#endregion
2450
2450
  export { createChatClient as C, computeTraceMetrics as S, CONTROL_INTEGRITY_ANALYST as _, analystFindingDigest as a, behavioralAnalyst as b, completedAnalystReviewQuality as c, validateAnalystReviewDecisions as d, DEFAULT_TRACE_ANALYST_KINDS as f, FAILURE_MODE_KIND_SPEC as g, IMPROVEMENT_KIND_SPEC as h, assertExactRegistryRunOpts as i, readAnalystReview as l, KNOWLEDGE_GAP_KIND_SPEC as m, AnalystRegistry as n, analystRunDigest as o, KNOWLEDGE_POISONING_KIND_SPEC as p, ExactAnalystRunExecutionError as r, assertUniqueFindingIds as s, buildDefaultAnalystRegistry as t, snapshotAnalystRun as u, ControlIntegrityAnalyst as v, deriveEfficiencyFindings as x, emitControlIntegrityFindings as y };
2451
2451
 
2452
- //# sourceMappingURL=default-registry-SOyHB6qG.js.map
2452
+ //# sourceMappingURL=default-registry-Dta70shL.js.map