@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
@@ -1,2 +1,2 @@
1
- import { a as IngestEvalRunsRequestSchema, c as InsightReportSchema, d as HOSTED_WIRE_VERSION, i as EvalRunEventSchema, l as TraceSpanEventSchema, n as hostedClientFromEnv, o as IngestResponseSchema, r as hostedTenantFromEnv, s as IngestTracesRequestSchema, t as createHostedClient, u as UnixNanoTimestampSchema } from "../client-QwJpsNYU.js";
2
- export { EvalRunEventSchema, HOSTED_WIRE_VERSION, IngestEvalRunsRequestSchema, IngestResponseSchema, IngestTracesRequestSchema, InsightReportSchema, TraceSpanEventSchema, UnixNanoTimestampSchema, createHostedClient, hostedClientFromEnv, hostedTenantFromEnv };
1
+ import { C as TraceSpanEventSchema, S as IngestTracesRequestSchema, T as HOSTED_WIRE_VERSION, _ as admitSearchLedgerBatch, a as HostedRequestError, b as searchLedgerHeadPath, c as hostedTenantFromEnv, d as SEARCH_LEDGER_BATCH_MAX_LINES, f as SEARCH_LEDGER_INGEST_PATH, g as SearchRunKindSchema, h as SearchLedgerHeadSchema, i as startSearchShipper, l as IngestSearchLedgerRequestSchema, m as SearchLedgerConflictSchema, o as createHostedClient, p as SearchBlobPutResponseSchema, r as shipSearchLedger, s as hostedClientFromEnv, t as SearchShipConflictError, u as SEARCH_LEDGER_BATCH_MAX_BYTES, v as searchBlobPath, w as UnixNanoTimestampSchema, x as IngestResponseSchema, y as searchLedgerArtifactRefs } from "../search-shipper-BG182p-U.js";
2
+ export { HOSTED_WIRE_VERSION, HostedRequestError, IngestResponseSchema, IngestSearchLedgerRequestSchema, IngestTracesRequestSchema, SEARCH_LEDGER_BATCH_MAX_BYTES, SEARCH_LEDGER_BATCH_MAX_LINES, SEARCH_LEDGER_INGEST_PATH, SearchBlobPutResponseSchema, SearchLedgerConflictSchema, SearchLedgerHeadSchema, SearchRunKindSchema, SearchShipConflictError, TraceSpanEventSchema, UnixNanoTimestampSchema, admitSearchLedgerBatch, createHostedClient, hostedClientFromEnv, hostedTenantFromEnv, searchBlobPath, searchLedgerArtifactRefs, searchLedgerHeadPath, shipSearchLedger, startSearchShipper };
@@ -306,6 +306,32 @@ declare class FileLedgerJournal<Header extends object, Event extends LedgerEvent
306
306
  * usually reads nothing or one row. */
307
307
  private entryAt;
308
308
  }
309
+ /** Where a memory journal keeps its text: an in-process store keyed by path,
310
+ * such as an in-memory `CampaignStorage`. */
311
+ interface LedgerTextStore {
312
+ read(path: string): string | undefined;
313
+ write(path: string, text: string): void;
314
+ }
315
+ /**
316
+ * A journal for runs without a filesystem. It keeps its rows as text in a
317
+ * store and applies the file journal's rules: the same row parse, hash chain,
318
+ * idempotent append by `eventId`, and projector transition before a row is
319
+ * kept. It has no lock, fsync or trusted head, because its rows live and die
320
+ * with the process that holds the store.
321
+ */
322
+ declare class MemoryLedgerJournal<Header extends object, Event extends LedgerEventBase, Projection> {
323
+ readonly path: string;
324
+ private readonly codec;
325
+ private readonly store;
326
+ private verified;
327
+ constructor(path: string, codec: LedgerJournalCodec<Header, Event, Projection>, store: LedgerTextStore);
328
+ replay(): Promise<Projection>;
329
+ append(event: Event): Promise<LedgerAppendResult<LedgerEntryOf<Header, Event>, Projection>>;
330
+ /** Verify the stored text again whenever it is not the text this instance wrote. */
331
+ private sync;
332
+ /** A refused transition can leave the projector half applied; discard it. */
333
+ private guarded;
334
+ }
309
335
  /** Verify and replay resolved immutable journal bytes without filesystem I/O. */
310
336
  declare function replayLedgerText<Header extends object, Event extends LedgerEventBase, Projection>(text: string, path: string, codec: LedgerJournalCodec<Header, Event, Projection>): Projection;
311
337
  //#endregion
@@ -329,5 +355,5 @@ declare function projectorSnapshotPathFor(journalPath: string): string;
329
355
  * ever explaining why the cache stopped working. */
330
356
  declare function readProjectorSnapshotFile(path: string, context: LedgerFileContext): LedgerProjectorSnapshot | null;
331
357
  //#endregion
332
- export { AtomicFileLockAcquisition as A, FileLockResult as C, withLedgerFileLock as D, tryWithLedgerFileLock as E, probeAtomicFileLock as F, tryAcquireAtomicFileLock as I, AtomicFileLockOptions as M, AtomicFileLockOwner as N, writeLedgerFileAtomically as O, AtomicFileLockUnavailable as P, verifyEntriesAgainstTrustedHead as S, appendLedgerLine as T, LedgerTrustedHead as _, FileLedgerJournalOptions as a, readTrustedHeadFile as b, LedgerChainFields as c, LedgerJournalCodec as d, LedgerLineContext as f, LedgerAnchoredEntry as g, replayLedgerText as h, FileLedgerJournal as i, AtomicFileLockError as j, AtomicFileLock as k, LedgerEntryOf as l, LedgerProjectorSnapshotCodec as m, projectorSnapshotPathFor as n, LedgerAppendOptions as o, LedgerProjector as p, readProjectorSnapshotFile as r, LedgerAppendResult as s, LedgerProjectorSnapshot as t, LedgerEventBase as u, LedgerTrustedHeadRemoval as v, LedgerFileContext as w, trustedHeadPathFor as x, LedgerTrustedHeadSubject as y };
333
- //# sourceMappingURL=index-I4UJGRno.d.ts.map
358
+ export { writeLedgerFileAtomically as A, trustedHeadPathFor as C, appendLedgerLine as D, LedgerFileContext as E, AtomicFileLockOwner as F, AtomicFileLockUnavailable as I, probeAtomicFileLock as L, AtomicFileLockAcquisition as M, AtomicFileLockError as N, tryWithLedgerFileLock as O, AtomicFileLockOptions as P, tryAcquireAtomicFileLock as R, readTrustedHeadFile as S, FileLockResult as T, replayLedgerText as _, FileLedgerJournalOptions as a, LedgerTrustedHeadRemoval as b, LedgerChainFields as c, LedgerJournalCodec as d, LedgerLineContext as f, MemoryLedgerJournal as g, LedgerTextStore as h, FileLedgerJournal as i, AtomicFileLock as j, withLedgerFileLock as k, LedgerEntryOf as l, LedgerProjectorSnapshotCodec as m, projectorSnapshotPathFor as n, LedgerAppendOptions as o, LedgerProjector as p, readProjectorSnapshotFile as r, LedgerAppendResult as s, LedgerProjectorSnapshot as t, LedgerEventBase as u, LedgerAnchoredEntry as v, verifyEntriesAgainstTrustedHead as w, LedgerTrustedHeadSubject as x, LedgerTrustedHead as y };
359
+ //# sourceMappingURL=index-Bj96W0jZ.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index-I4UJGRno.d.ts","names":[],"sources":["../src/ledger-core/atomic-file-lock.ts","../src/ledger-core/journal-file.ts","../src/ledger-core/trusted-head.ts","../src/ledger-core/journal.ts","../src/ledger-core/projector-snapshot.ts"],"mappings":";;;UAiBiB;WACN;WACA;WACA;;UAGM;WACN,OAAO;EAChB;;KAGU;WACG;WAA0B;WAAyB,QAAQ;;WAC3D;WAA0B;;KAE7B;WACG;WAAyB,MAAM;IAC1C;UAEa;WACN;WACA;;cAGE,4BAA4B;WACrB;;;iBASJ,oBACd,SAAS,wBACR;;;;;;;iBAea,yBACd,SAAS,wBACR;;;;;UCpDc;;EAEf;EACA,eAAe,iBAAiB;IAAY;MAAoB;;;;iBAKlD,iBAAiB,cAAc,cAAc,SAAS;;;;iBA6CtD,0BACd,cACA,kBACA,SAAS;iBAqCK,mBAAmB,GAAG,cAAc,SAAS,mBAAmB,WAAW,IAAI;KAQnF,eAAe;EAAO;EAAgB,OAAO;;EAAQ;;iBAEjD,sBAAsB,GACpC,cACA,SAAS,mBACT,WAAW,IACV,eAAe;;;;UCnED;EACf;EACA,WAAW;;;UAII;EACf;EACA,WAAW;;;;iBAKG,mBAAmB;;;;;iBAWnB,oBACd,cACA,SAAS,oBACR;;;;;;;KA2ES;EACN;EAAe,MAAM;EAAmB;;EACxC;EAAe;EAAY;EAAoB;;EAC/C;EAAgB;;;UAsCL;;EAEf;;;EAGA;;;;;;;iBAQc,gCACd,SAAS,cAAc,yCACvB,MAAM,mBACN,SAAS,mBACT,QAAQ;;;;UCzJO;EACf;;;;;UAMe,kBAAkB,cAAc;EAC/C;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD,cAAc,uBAAuB,cAAc,mBAAmB,SAChF,kBAAkB;UAEH;EACf;EACA;;;;;;;;;;UAWe,gBAAgB,OAAO;EACtC,MAAM,OAAO,OAAO;;;;EAIpB,YAAY;;;;;;;;;;UAWG,6BAA6B,OAAO;EACnD,UAAU,YAAY;EACtB,QAAQ,sBAAsB,gBAAgB,OAAO;;UAGtC,mBACf,uBACA,cAAc,iBACd,oBACQ;;EAER,QAAQ;EACR,cAAc,kBAAkB;;;;;;EAMhC,WAAW,cAAc,SAAS,oBAAoB,cAAc,QAAQ;;EAE5E,iBAAiB,OAAO,cAAc,QAAQ,QAAQ;EACtD,mBAAmB,gBAAgB,cAAc,QAAQ,QAAQ;EACjE,qBAAqB,6BAA6B,cAAc,QAAQ,QAAQ;;UAGjE,mBAAmB,OAAO;EACzC,OAAO;;EAEP;EACA,YAAY;;UAGG;;;;;;;;;;;;;;EAcf;;UAGe;;;;;;;EAOf;;;;;;EAMA;IAAsB;;;;;;cA2CX,kBAAkB,uBAAuB,cAAc,iBAAiB;WAC1E;;WAEA;;;WAGA;mBACQ;mBACA;mBACA;mBACA;UACT;EAER,YACE,cACA,OAAO,mBAAmB,QAAQ,OAAO,aACzC,UAAS;EAcL,UAAU,QAAQ;EAIlB,OACJ,OAAO,OACP,UAAS,sBACR,QAAQ,mBAAmB,cAAc,QAAQ,QAAQ;;EA+CtD,eAAe,QAAQ;;;;EAWvB,kBAAkB,QAAQ;;;;;;;;EAkC1B,oBAAoB,QAAQ;;;;UAsB1B;;;;UAgBA;UAwCA;;;;;;;;;;;;;UA2CA;;;;;;;;UAuBA;;;;UAiBA;;;;;;;;;;;;;;;UAwBA;;;;;;UAqIA;;;iBA8CM,iBAAiB,uBAAuB,cAAc,iBAAiB,YACrF,cACA,cACA,OAAO,mBAAmB,QAAQ,OAAO,cACxC;;;;UC7pBc;EACf,MAAM;;;;EAIN;;EAEA;;;iBAIc,yBAAyB;;;;;;iBAYzB,0BACd,cACA,SAAS,oBACR"}
1
+ {"version":3,"file":"index-Bj96W0jZ.d.ts","names":[],"sources":["../src/ledger-core/atomic-file-lock.ts","../src/ledger-core/journal-file.ts","../src/ledger-core/trusted-head.ts","../src/ledger-core/journal.ts","../src/ledger-core/projector-snapshot.ts"],"mappings":";;;UAiBiB;WACN;WACA;WACA;;UAGM;WACN,OAAO;EAChB;;KAGU;WACG;WAA0B;WAAyB,QAAQ;;WAC3D;WAA0B;;KAE7B;WACG;WAAyB,MAAM;IAC1C;UAEa;WACN;WACA;;cAGE,4BAA4B;WACrB;;;iBASJ,oBACd,SAAS,wBACR;;;;;;;iBAea,yBACd,SAAS,wBACR;;;;;UCpDc;;EAEf;EACA,eAAe,iBAAiB;IAAY;MAAoB;;;;iBAKlD,iBAAiB,cAAc,cAAc,SAAS;;;;iBA6CtD,0BACd,cACA,kBACA,SAAS;iBAqCK,mBAAmB,GAAG,cAAc,SAAS,mBAAmB,WAAW,IAAI;KAQnF,eAAe;EAAO;EAAgB,OAAO;;EAAQ;;iBAEjD,sBAAsB,GACpC,cACA,SAAS,mBACT,WAAW,IACV,eAAe;;;;UCnED;EACf;EACA,WAAW;;;UAII;EACf;EACA,WAAW;;;;iBAKG,mBAAmB;;;;;iBAWnB,oBACd,cACA,SAAS,oBACR;;;;;;;KA2ES;EACN;EAAe,MAAM;EAAmB;;EACxC;EAAe;EAAY;EAAoB;;EAC/C;EAAgB;;;UAsCL;;EAEf;;;EAGA;;;;;;;iBAQc,gCACd,SAAS,cAAc,yCACvB,MAAM,mBACN,SAAS,mBACT,QAAQ;;;;UCzJO;EACf;;;;;UAMe,kBAAkB,cAAc;EAC/C;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD,cAAc,uBAAuB,cAAc,mBAAmB,SAChF,kBAAkB;UAEH;EACf;EACA;;;;;;;;;;UAWe,gBAAgB,OAAO;EACtC,MAAM,OAAO,OAAO;;;;EAIpB,YAAY;;;;;;;;;;UAWG,6BAA6B,OAAO;EACnD,UAAU,YAAY;EACtB,QAAQ,sBAAsB,gBAAgB,OAAO;;UAGtC,mBACf,uBACA,cAAc,iBACd,oBACQ;;EAER,QAAQ;EACR,cAAc,kBAAkB;;;;;;EAMhC,WAAW,cAAc,SAAS,oBAAoB,cAAc,QAAQ;;EAE5E,iBAAiB,OAAO,cAAc,QAAQ,QAAQ;EACtD,mBAAmB,gBAAgB,cAAc,QAAQ,QAAQ;EACjE,qBAAqB,6BAA6B,cAAc,QAAQ,QAAQ;;UAGjE,mBAAmB,OAAO;EACzC,OAAO;;EAEP;EACA,YAAY;;UAGG;;;;;;;;;;;;;;EAcf;;UAGe;;;;;;;EAOf;;;;;;EAMA;IAAsB;;;;;;cA2CX,kBAAkB,uBAAuB,cAAc,iBAAiB;WAC1E;;WAEA;;;WAGA;mBACQ;mBACA;mBACA;mBACA;UACT;EAER,YACE,cACA,OAAO,mBAAmB,QAAQ,OAAO,aACzC,UAAS;EAcL,UAAU,QAAQ;EAIlB,OACJ,OAAO,OACP,UAAS,sBACR,QAAQ,mBAAmB,cAAc,QAAQ,QAAQ;;EA+CtD,eAAe,QAAQ;;;;EAWvB,kBAAkB,QAAQ;;;;;;;;EAkC1B,oBAAoB,QAAQ;;;;UAsB1B;;;;UAgBA;UAwCA;;;;;;;;;;;;;UA2CA;;;;;;;;UAuBA;;;;UAiBA;;;;;;;;;;;;;;;UAwBA;;;;;;UAqIA;;;;UA+CO;EACf,KAAK;EACL,MAAM,cAAc;;;;;;;;;cAUT,oBAAoB,uBAAuB,cAAc,iBAAiB;WAC5E;mBACQ;mBACA;UACT;EAMR,YACE,cACA,OAAO,mBAAmB,QAAQ,OAAO,aACzC,OAAO;EAOH,UAAU,QAAQ;EAIlB,OACJ,OAAO,QACN,QAAQ,mBAAmB,cAAc,QAAQ,QAAQ;;UAgCpD;;UAgBM;;;iBAWA,iBAAiB,uBAAuB,cAAc,iBAAiB,YACrF,cACA,cACA,OAAO,mBAAmB,QAAQ,OAAO,cACxC;;;;UC/vBc;EACf,MAAM;;;;EAIN;;EAEA;;;iBAIc,yBAAyB;;;;;;iBAYzB,0BACd,cACA,SAAS,oBACR"}
@@ -1,11 +1,11 @@
1
1
  import { b as CustomTokenPricing, c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
2
- import { i as AnalystFinding, n as AnalystContext } from "./types-Bay1dh2o.js";
2
+ import { i as AnalystFinding, n as AnalystContext } from "./types-CeDfCRXp.js";
3
3
  import { R as LlmThinkingMode, x as ChatClient } from "./types-C0bWoSKC.js";
4
- import { F as ExternalOptimizerCallbackLimits, V as ExternalOptimizerModelCall, W as ExternalOptimizerModelExecutionObservation, q as ExternalOptimizerRunnerCommand } from "./agent-profile-DNV_U6QR.js";
4
+ import { F as ExternalOptimizerCallbackLimits, V as ExternalOptimizerModelCall, W as ExternalOptimizerModelExecutionObservation, q as ExternalOptimizerRunnerCommand } from "./agent-profile-DDfIjdyS.js";
5
5
  import { o as Severity } from "./multi-layer-verifier-BUaQ4C17.js";
6
- import { l as ExactCapableAnalyst } from "./exact-types-CfpP2A4Z.js";
7
- import { i as TraceAnalystDefinition } from "./default-registry-B5HrwrPM.js";
8
- import { n as TraceAnalysisEngine } from "./engine-BkLCnZQX.js";
6
+ import { l as ExactCapableAnalyst } from "./exact-types-KGxBWH4b.js";
7
+ import { i as TraceAnalystDefinition } from "./default-registry-DdqA_-_o.js";
8
+ import { n as TraceAnalysisEngine } from "./engine-C1DIAEpn.js";
9
9
  import { C as SupervisorRunSources, w as SupervisorRunTree } from "./types-DRNlLbBR.js";
10
10
  //#region src/semantic-concept-judge.d.ts
11
11
  /**
@@ -476,4 +476,4 @@ declare const KNOWLEDGE_POISONING_KIND_SPEC: TraceAnalystDefinition;
476
476
  declare const DEFAULT_TRACE_ANALYST_KINDS: readonly TraceAnalystDefinition[];
477
477
  //#endregion
478
478
  export { SemanticConceptJudgeInput as A, DspyRlmTraceEngineOptions as C, ConceptFinding as D, createChatTraceEngine as E, SemanticConceptJudgeResult as M, runSemanticConceptJudge as N, ConceptSpec as O, renderFindingSubject as S, ChatTraceEngineOptions as T, FindingSubject as _, FAILURE_MODE_KIND_SPEC as a, findingSubjectGrammarPromptFor as b, emitControlIntegrityFindings as c, FindingsStore as d, PersistedFinding as f, FINDING_SUBJECT_SYNTAX as g, FINDING_SUBJECT_KINDS as h, IMPROVEMENT_KIND_SPEC as i, SemanticConceptJudgeOptions as j, SEMANTIC_CONCEPT_JUDGE_VERSION as k, DiffPolicy as l, diffFindings as m, KNOWLEDGE_POISONING_KIND_SPEC as n, CONTROL_INTEGRITY_ANALYST as o, defaultIsMaterial as p, KNOWLEDGE_GAP_KIND_SPEC as r, ControlIntegrityAnalyst as s, DEFAULT_TRACE_ANALYST_KINDS as t, FindingsDiff as u, FindingSubjectKind as v, createDspyRlmTraceEngine as w, parseFindingSubject as x, KIND_EXPECTED_SUBJECTS as y };
479
- //# sourceMappingURL=index-2e_MTznj.d.ts.map
479
+ //# sourceMappingURL=index-DBoXFaz0.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index-2e_MTznj.d.ts","names":[],"sources":["../src/semantic-concept-judge.ts","../src/analyst/chat-trace-engine.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts","../src/analyst/finding-subject.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/control-integrity.ts","../src/analyst/kinds/failure-mode.ts","../src/analyst/kinds/improvement.ts","../src/analyst/kinds/knowledge-gap.ts","../src/analyst/kinds/knowledge-poisoning.ts","../src/analyst/kinds/index.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAuDY;UAEK;EACf;;EAEA;;EAEA;;EAEA,aAAa;;UAGE;EACf;EACA;;EAEA;EACA;EACA,UAAU;;UAGK;;EAEf;;EAEA;;EAEA,aAAa;IAAQ;IAAc;;;EAEnC,kBAAkB;;EAElB;EACA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA;;EAEA;EACA;;;;;;;;KASU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,UAAU;EACV,aAAa;EACb;EACA,WAAW;EACX,SAAS;;;;;;EAMT,iBAAiB;;EAEjB,oBAAoB,QAAQ,OAAO;;cAKxB;;;;;;iBAkGS,wBACpB,OAAO,2BACP,SAAS,8BACR,QAAQ;;;UCjMM;;;;;EAKf,MAAM;;EAEN;;EAEA,UAAU;;EAEV;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;EAUA;;;;;;iBAkCc,sBAAsB,SAAS,yBAAyB;;;KCrG5D,0BAA0B;;;UCYrB;;EAEf,MAAM;;EAEN;;EAEA,kBAAkB,aAAa;EAC/B;;EAEA,UAAU;;EAEV;;EAEA;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,kBAAkB,QAAQ;;EAE1B;;;;;;;;;;EAUA;;EAEA,SAAS;;EAET;;;iBAIc,yBAAyB,SAAS,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KCtClE;EAEN;EAAwB;EAAc;;EACtC;EAAyB;;EACzB;EAAuB;;EACvB;EAAyB;;EAGzB;EAAuB;;EACvB;EAAe;;EACf;EAAkB;EAAc;;EAChC;EAAkB;;EAClB;EAAa;EAAgB;;EAC7B;EAAc;;EACd;EAAkB;;EAClB;EAAkB;;EAClB;EAAwB;;EACxB;EAAuB;;EACvB;EAAc;;EACd;EAAa;EAAgB;;EAC7B;EAAgB;;EAChB;EAAqB;;EACrB;EAAuB;;EAEvB;EAA4B;;EAC5B;EAA2B;;EAE3B;EAAiB;;KAEX,qBAAqB;cAEpB,uBAAuB,cAAc;;;;;;;;;;;;;;;;;iBAgGlC,oBAAoB,iCAAiC;;;;;;;iBAoHrD,qBAAqB,GAAG;;;;;;;;;;cA8D3B,wBAAwB,SAAS,OAAO;;;;;;;;;cAoExC,wBAAwB,eAAe,cAAc;;iBAqElD,+BAA+B;;;;;;;;UCxc9B,yBAAyB;EACxC;;cAGW;WAGiB;mBAFX;EAEjB,YAA4B;EAItB,OAAO,eAAe,UAAU,mBAAmB;;EAQzD,WAAW;;EAkBX,QAAQ,gBAAgB;;UAOT;;EAEf,UAAU;;EAEV,aAAa;;EAEb,WAAW;;;;;EAKX,SAAS;IAAQ,UAAU;IAAkB,SAAS;;;UAGvC;;;;;;;;EAQf,cAAc,UAAU,gBAAgB,SAAS;;;;;;iBAOnC,kBAAkB,GAAG,gBAAgB,GAAG;;;;;iBAWxC,aACd,UAAU,oBACV,SAAS,oBACT,SAAQ,aACP;;;;iBC7Fa,6BACd,OAAO,uBAAuB,mBAC9B,qBACC;;cA4BU,mCACA,oBAAoB,uBAAuB;WAE7C;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;;EAGI,QACJ,OAAO,uBAAuB,mBAC9B,KAAK,iBACJ,QAAQ;;cAUA,2BAAyB;;;cCHzB,wBAAwB;;;cC3BxB,uBAAuB;;;cCGvB,yBAAyB;;;cCPzB,+BAA+B;;;;;;;;;;;;cCP/B,sCAAsC"}
1
+ {"version":3,"file":"index-DBoXFaz0.d.ts","names":[],"sources":["../src/semantic-concept-judge.ts","../src/analyst/chat-trace-engine.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts","../src/analyst/finding-subject.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/control-integrity.ts","../src/analyst/kinds/failure-mode.ts","../src/analyst/kinds/improvement.ts","../src/analyst/kinds/knowledge-gap.ts","../src/analyst/kinds/knowledge-poisoning.ts","../src/analyst/kinds/index.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAuDY;UAEK;EACf;;EAEA;;EAEA;;EAEA,aAAa;;UAGE;EACf;EACA;;EAEA;EACA;EACA,UAAU;;UAGK;;EAEf;;EAEA;;EAEA,aAAa;IAAQ;IAAc;;;EAEnC,kBAAkB;;EAElB;EACA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA;;EAEA;EACA;;;;;;;;KASU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,UAAU;EACV,aAAa;EACb;EACA,WAAW;EACX,SAAS;;;;;;EAMT,iBAAiB;;EAEjB,oBAAoB,QAAQ,OAAO;;cAKxB;;;;;;iBAkGS,wBACpB,OAAO,2BACP,SAAS,8BACR,QAAQ;;;UCjMM;;;;;EAKf,MAAM;;EAEN;;EAEA,UAAU;;EAEV;;EAEA;;EAEA,WAAW;;EAEX;;;;;;;;;;EAUA;;;;;;iBAkCc,sBAAsB,SAAS,yBAAyB;;;KCrG5D,0BAA0B;;;UCYrB;;EAEf,MAAM;;EAEN;;EAEA,kBAAkB,aAAa;EAC/B;;EAEA,UAAU;;EAEV;;EAEA;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,kBAAkB,QAAQ;;EAE1B;;;;;;;;;;EAUA;;EAEA,SAAS;;EAET;;;iBAIc,yBAAyB,SAAS,4BAA4B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KCtClE;EAEN;EAAwB;EAAc;;EACtC;EAAyB;;EACzB;EAAuB;;EACvB;EAAyB;;EAGzB;EAAuB;;EACvB;EAAe;;EACf;EAAkB;EAAc;;EAChC;EAAkB;;EAClB;EAAa;EAAgB;;EAC7B;EAAc;;EACd;EAAkB;;EAClB;EAAkB;;EAClB;EAAwB;;EACxB;EAAuB;;EACvB;EAAc;;EACd;EAAa;EAAgB;;EAC7B;EAAgB;;EAChB;EAAqB;;EACrB;EAAuB;;EAEvB;EAA4B;;EAC5B;EAA2B;;EAE3B;EAAiB;;KAEX,qBAAqB;cAEpB,uBAAuB,cAAc;;;;;;;;;;;;;;;;;iBAgGlC,oBAAoB,iCAAiC;;;;;;;iBAoHrD,qBAAqB,GAAG;;;;;;;;;;cA8D3B,wBAAwB,SAAS,OAAO;;;;;;;;;cAoExC,wBAAwB,eAAe,cAAc;;iBAqElD,+BAA+B;;;;;;;;UCxc9B,yBAAyB;EACxC;;cAGW;WAGiB;mBAFX;EAEjB,YAA4B;EAItB,OAAO,eAAe,UAAU,mBAAmB;;EAQzD,WAAW;;EAkBX,QAAQ,gBAAgB;;UAOT;;EAEf,UAAU;;EAEV,aAAa;;EAEb,WAAW;;;;;EAKX,SAAS;IAAQ,UAAU;IAAkB,SAAS;;;UAGvC;;;;;;;;EAQf,cAAc,UAAU,gBAAgB,SAAS;;;;;;iBAOnC,kBAAkB,GAAG,gBAAgB,GAAG;;;;;iBAWxC,aACd,UAAU,oBACV,SAAS,oBACT,SAAQ,aACP;;;;iBC7Fa,6BACd,OAAO,uBAAuB,mBAC9B,qBACC;;cA4BU,mCACA,oBAAoB,uBAAuB;WAE7C;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;;EAGI,QACJ,OAAO,uBAAuB,mBAC9B,KAAK,iBACJ,QAAQ;;cAUA,2BAAyB;;;cCHzB,wBAAwB;;;cC3BxB,uBAAuB;;;cCGvB,yBAAyB;;;cCPzB,+BAA+B;;;;;;;;;;;;cCP/B,sCAAsC"}
@@ -1,14 +1,13 @@
1
1
  import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
2
2
  import { a as RunRecord, c as RunSplitTag, n as RunCostProvenance } from "./run-record-DF-fuu31.js";
3
- import "./types-Bay1dh2o.js";
4
- import "./statistical-heldout-CLpjYffJ.js";
5
- import { _ as GateDecision } from "./types-BvmWNjjs.js";
6
- import "./index-CJv56rhX.js";
7
- import { o as InsightReport } from "./insight-report-CvPkEIcC.js";
8
- import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, o as EvalRunCellScore, s as EvalRunEvent } from "./client-BhEyAdfe.js";
9
- import "./define-agent-eval-DblGedQk.js";
10
- import "./default-registry-B5HrwrPM.js";
11
- import { t as AnalyzeRunsOptions } from "./analyze-runs-BvK4-wSs.js";
3
+ import "./types-CeDfCRXp.js";
4
+ import "./statistical-heldout-C1Ej5Ce8.js";
5
+ import "./types-C1CAD70_.js";
6
+ import "./index-DwQxdWve.js";
7
+ import { p as TraceSpanEvent } from "./client-wFyDisxi.js";
8
+ import { d as InsightReport, t as AnalyzeRunsOptions } from "./analyze-runs-CPiKLJFs.js";
9
+ import "./define-agent-eval-DcludNJS.js";
10
+ import "./default-registry-DdqA_-_o.js";
12
11
  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
12
  //#region src/contract/measured-comparison.d.ts
14
13
  interface SealCandidateBenchmarkSuiteOptions {
@@ -277,99 +276,6 @@ interface EvalReportingSuiteResult {
277
276
  */
278
277
  declare function evalReportingSuite(input: EvalReportingSuiteInput, options?: EvalReportingSuiteOptions): Promise<EvalReportingSuiteResult>;
279
278
  //#endregion
280
- //#region src/contract/diff.d.ts
281
- /** Per-dimension delta. `before` / `after` are null when the judge did not
282
- * emit a value for that side. `delta` is `after - before`; null when
283
- * either side is null. */
284
- interface EvalDimensionDelta {
285
- before: number | null;
286
- after: number | null;
287
- delta: number | null;
288
- }
289
- /** Per-cell delta, keyed on `(scenarioId, rep)`. */
290
- interface EvalCellScoreDelta {
291
- scenarioId: string;
292
- rep: number;
293
- compositeBefore: number | null;
294
- compositeAfter: number | null;
295
- compositeDelta: number | null;
296
- /** Per-judge → per-dimension deltas. Outer key = judge name from
297
- * `EvalRunCellScore.dimensions`; inner key = dimension name. */
298
- dimensions: Record<string, Record<string, EvalDimensionDelta>>;
299
- }
300
- /** Diff between two generation snapshots — the unit the dashboard renders
301
- * for a single "v3 vs v4" comparison. */
302
- interface EvalGenerationDiff {
303
- beforeIndex: number;
304
- afterIndex: number;
305
- beforeSurfaceHash: string;
306
- afterSurfaceHash: string;
307
- surfaceChanged: boolean;
308
- /** Cells present in both snapshots, matched on `(scenarioId, rep)`. */
309
- matched: EvalCellScoreDelta[];
310
- /** Cells present in `before` but missing from `after`. */
311
- removed: EvalRunCellScore[];
312
- /** Cells present in `after` but missing from `before`. */
313
- added: EvalRunCellScore[];
314
- /** Aggregate composite mean, null when that snapshot was unscored. */
315
- compositeBefore: number | null;
316
- compositeAfter: number | null;
317
- compositeDelta: number | null;
318
- costUsdBefore: number;
319
- costUsdAfter: number;
320
- costUsdDelta: number;
321
- durationMsBefore: number;
322
- durationMsAfter: number;
323
- durationMsDelta: number;
324
- }
325
- /** Diff between two full eval-runs. Includes both baseline-vs-baseline and
326
- * winner-vs-winner generation diffs when both sides expose them, plus
327
- * run-level metadata. */
328
- interface EvalRunDiff {
329
- beforeRunId: string;
330
- afterRunId: string;
331
- beforeTimestamp: string;
332
- afterTimestamp: string;
333
- beforeGateDecision: GateDecision | null;
334
- afterGateDecision: GateDecision | null;
335
- beforeHoldoutLift: number | null;
336
- afterHoldoutLift: number | null;
337
- holdoutLiftDelta: number | null;
338
- beforeTotalCostUsd: number;
339
- afterTotalCostUsd: number;
340
- totalCostUsdDelta: number;
341
- beforeTotalDurationMs: number;
342
- afterTotalDurationMs: number;
343
- totalDurationMsDelta: number;
344
- /** Baseline-vs-baseline diff. Null when either run has no baseline. */
345
- baselineDiff: EvalGenerationDiff | null;
346
- /** Highest-index-generation comparison. Null when either run has no
347
- * recorded generations (e.g. baseline-only or errored before any
348
- * generation completed). */
349
- winnersDiff: EvalGenerationDiff | null;
350
- }
351
- /**
352
- * Diff two generation snapshots. Cells are matched on `(scenarioId, rep)`;
353
- * unmatched cells surface in `added` / `removed`. Aggregate fields are
354
- * recomputed from the snapshot's stored fields, not re-derived from cells —
355
- * this keeps the diff consistent with whatever aggregation the substrate
356
- * actually reported.
357
- */
358
- declare function diffGenerations(before: EvalRunGenerationSnapshot, after: EvalRunGenerationSnapshot): EvalGenerationDiff;
359
- /**
360
- * Diff two full eval-runs. Produces baseline-vs-baseline and
361
- * winner-vs-winner generation diffs when both sides expose them, plus
362
- * run-level cost / lift / gate-decision deltas.
363
- */
364
- declare function diffRuns(before: EvalRunEvent, after: EvalRunEvent): EvalRunDiff;
365
- /**
366
- * Within-run baseline → winning-generation diff. The natural "what did the
367
- * improvement loop produce" view for a single run. Returns null when the
368
- * run never reached a generation past baseline (errored early, or the gate
369
- * shipped the baseline as-is).
370
- */
371
- declare function diffRunBaselineToWinner(run: EvalRunEvent): EvalGenerationDiff | null;
372
- //#endregion
373
279
  //#region src/contract/intake/agent-trace.d.ts
374
280
  type AgentTraceContributorType = 'human' | 'ai' | 'mixed' | 'unknown';
375
281
  interface AgentTraceContributor {
@@ -780,5 +686,5 @@ interface FromOtelSpansOptions {
780
686
  }
781
687
  declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
782
688
  //#endregion
783
- export { diffRunBaselineToWinner as $, CodeAgentSessionActionStatus as A, measuredComparisonFromCandidateExperiment as At, AgentTraceFile as B, fromOpenCodeSession as C, EvaluatePairedMeasurementsOptions as Ct, streamCodeAgentJsonlFile as D, RunCandidateExperimentOptions as Dt, parseCodeAgentJsonlFile as E, PairedMeasurementEvaluation as Et, CodeAgentSessionTerminalStatus as F, verifyCandidateExperiment as Ft, PartitionByAuthoringModelResult as G, AgentTraceRange as H, observeCodeAgentSession as I, verifyCandidateExperimentComparison as It, EvalCellScoreDelta as J, parseAgentTrace as K, AgentTraceContributor as L, CodeAgentSessionExecutionReceipt as M, sealCandidateBenchmarkSuite as Mt, CodeAgentSessionObservation as N, sealCandidateBenchmarkTask as Nt, CodeAgentSessionAction as O, SealCandidateBenchmarkSuiteOptions as Ot, CodeAgentSessionSource as P, sealCandidateExperiment as Pt, diffGenerations as Q, AgentTraceContributorType as R, fromKimiCodeSession as S, CompareCandidateExperimentOptions as St, parseCodeAgentJsonl as T, PairedMeasurementAdapter as Tt, AgentTraceRecord as U, AgentTraceIndex as V, AuthoringProvenance as W, EvalGenerationDiff as X, EvalDimensionDelta as Y, EvalRunDiff as Z, CodeAgentSessionIntakeResult as _, sealAgentProfileImprovementSuite as _t, FromFeedbackTableOptions as a, FromRunRecordDirOptions as at, fromClaudeCodeSession as b, CandidateExperimentExecutionInput as bt, CodeAgentStoreObservation as c, fromRunRecordDir as ct, CodeAgentStoreScope as d, CompareAgentProfileImprovementExperimentOptions as dt, diffRuns as et, CodeAgentStoreSession as f, RunAgentProfileImprovementExperimentOptions as ft, CodeAgentSessionIntakeOptions as g, sealAgentProfileImprovementExperiment as gt, CodeAgentSessionDiagnostic as h, runAgentProfileImprovementExperiment as ht, FeedbackTableRow as i, evalReportingSuite as it, CodeAgentSessionActionSurface as j, runCandidateExperiment as jt, CodeAgentSessionActionKind as k, evaluatePairedMeasurements as kt, CodeAgentStoreRejection as l, AgentProfileImprovementExperimentExecutionInput as lt, CodeAgentJsonlLine as m, measuredComparisonFromAgentProfileImprovementExperiment as mt, fromOtelSpans as n, EvalReportingSuiteOptions as nt, FromFeedbackTableResult as o, FromRunRecordDirResult as ot, observeCodeAgentStore as p, SealAgentProfileImprovementSuiteOptions as pt, partitionRunsByAuthoringModel as q, FeedbackTableMeta as r, EvalReportingSuiteResult as rt, fromFeedbackTable as s, RunRecordRejection as st, FromOtelSpansOptions as t, EvalReportingSuiteInput as tt, CodeAgentStoreRejections as u, AgentProfileImprovementExperimentRun as ut, CodeAgentSessionMetrics as v, sealAgentProfileImprovementTask as vt, fromPiSession as w, PairedMeasurement as wt, fromCodexSession as x, CandidateExperimentRun as xt, ParsedCodeAgentJsonl as y, verifyAgentProfileImprovementExperimentComparison as yt, AgentTraceConversation as z };
784
- //# sourceMappingURL=index-CWoIJHmy.d.ts.map
689
+ export { FromRunRecordDirResult as $, CodeAgentSessionActionStatus as A, AgentTraceFile as B, fromOpenCodeSession as C, measuredComparisonFromCandidateExperiment as Ct, streamCodeAgentJsonlFile as D, sealCandidateExperiment as Dt, parseCodeAgentJsonlFile as E, sealCandidateBenchmarkTask as Et, CodeAgentSessionTerminalStatus as F, PartitionByAuthoringModelResult as G, AgentTraceRange as H, observeCodeAgentSession as I, EvalReportingSuiteInput as J, parseAgentTrace as K, AgentTraceContributor as L, CodeAgentSessionExecutionReceipt as M, CodeAgentSessionObservation as N, CodeAgentSessionAction as O, verifyCandidateExperiment as Ot, CodeAgentSessionSource as P, FromRunRecordDirOptions as Q, AgentTraceContributorType as R, fromKimiCodeSession as S, evaluatePairedMeasurements as St, parseCodeAgentJsonl as T, sealCandidateBenchmarkSuite as Tt, AgentTraceRecord as U, AgentTraceIndex as V, AuthoringProvenance as W, EvalReportingSuiteResult as X, EvalReportingSuiteOptions as Y, evalReportingSuite as Z, CodeAgentSessionIntakeResult as _, PairedMeasurement as _t, FromFeedbackTableOptions as a, RunAgentProfileImprovementExperimentOptions as at, fromClaudeCodeSession as b, RunCandidateExperimentOptions as bt, CodeAgentStoreObservation as c, runAgentProfileImprovementExperiment as ct, CodeAgentStoreScope as d, sealAgentProfileImprovementTask as dt, RunRecordRejection as et, CodeAgentStoreSession as f, verifyAgentProfileImprovementExperimentComparison as ft, CodeAgentSessionIntakeOptions as g, EvaluatePairedMeasurementsOptions as gt, CodeAgentSessionDiagnostic as h, CompareCandidateExperimentOptions as ht, FeedbackTableRow as i, CompareAgentProfileImprovementExperimentOptions as it, CodeAgentSessionActionSurface as j, CodeAgentSessionActionKind as k, verifyCandidateExperimentComparison as kt, CodeAgentStoreRejection as l, sealAgentProfileImprovementExperiment as lt, CodeAgentJsonlLine as m, CandidateExperimentRun as mt, fromOtelSpans as n, AgentProfileImprovementExperimentExecutionInput as nt, FromFeedbackTableResult as o, SealAgentProfileImprovementSuiteOptions as ot, observeCodeAgentStore as p, CandidateExperimentExecutionInput as pt, partitionRunsByAuthoringModel as q, FeedbackTableMeta as r, AgentProfileImprovementExperimentRun as rt, fromFeedbackTable as s, measuredComparisonFromAgentProfileImprovementExperiment as st, FromOtelSpansOptions as t, fromRunRecordDir as tt, CodeAgentStoreRejections as u, sealAgentProfileImprovementSuite as ut, CodeAgentSessionMetrics as v, PairedMeasurementAdapter as vt, fromPiSession as w, runCandidateExperiment as wt, fromCodexSession as x, SealCandidateBenchmarkSuiteOptions as xt, ParsedCodeAgentJsonl as y, PairedMeasurementEvaluation as yt, AgentTraceConversation as z };
690
+ //# sourceMappingURL=index-DNVzu9Vv.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"index-DNVzu9Vv.d.ts","names":[],"sources":["../src/contract/measured-comparison.ts","../src/contract/profile-measured-comparison.ts","../src/contract/intake/run-record-dir.ts","../src/contract/eval-reporting-suite.ts","../src/contract/intake/agent-trace.ts","../src/contract/intake/code-agent-observation.ts","../src/contract/intake/code-agent-session.ts","../src/contract/intake/code-agent-store.ts","../src/contract/intake/feedback-table.ts","../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;;UAiCiB;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,eAAe,KAAK,OAAO;EAC3B,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA,kBAAkB;;EAElB,kBAAkB;;;KAIR,8BAA8B,KACxC;EAGA,iBAAiB;EACjB,WAAW;EACX;;;iBAIc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;;;;;;;iBAkEK,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;;iBA8Ya,0CACd,SAAS,oCACR;;iBAuEa,oCACd,iBACC;;;UCjrBc;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;;;;;;iBAsBmB,qCACpB,SAAS,8CACR,QAAQ;;iBA0CK,wDACd,SAAS,kDACR;;iBAuEa,kDACd,iBACC;;;;UC7Oc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;;;;;;;iBAkBoB,iBACpB,cACA,UAAS,0BACR,QAAQ;;;;;KC7CC,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;;;;;;iBAUoB,mBACpB,OAAO,yBACP,UAAS,4BACR,QAAQ;;;KChEC;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;KCjLS;KAEA;KAEA;KAEA;KASA;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,MAAM;EACN,SAAS;EACT;EACA,QAAQ;EACR;EACA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR;EACA;EACA;IACE,QAAQ;IACR;;EAEF,SAAS;;UAGM;EACf,QAAQ;EACR;EACA;EACA,YAAY;;;;;;;iBAeE,wBACd,SAAS,iCACR;;;UCjCc;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;;UAGe;EACf,MAAM;EACN,aAAa;EACb,SAAS;EACT,cAAc;;UAGC;EACf;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;EACA;EACA;EACA;EACA;EACA;;;EAGA,iBAAiB;;;EAGjB,YAAY;;;;KAKF;EACN;EAAe;EAAoB;;EACnC;EAAmB;;iBAcT,oBAAoB,gBAAgB;;;;;;;;;;iBAuB7B,yBAAyB,eAAe,eAAe;;;;;iBA6BxD,wBAAwB,eAAe,QAAQ;iBAUrD,iBACd,SAAS,gCACR;iBAIa,sBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,cACd,SAAS,gCACR;;;;UC/Lc;;;;;WAKN;;WAEA,QAAQ;;;;;WAKR;;;;;WAKA;aACE;;aAEA;;;;KAKD;;UAUK;WACN;WACA;WACA;WACA;WACA,aAAa;;;UAIP;WACN;WACA,UAAU,SAAS,QAAQ,OAAO;;WAElC,QAAQ;aACN;aACA,QAAQ;;;;;;;;;KAUT;WAEG;WACA,OAAO;WACP;WACA,mBAAmB;;WAEnB;;WAEA;;WAEA;WACA,WAAW,SAAS,QAAQ,OAAO;WACnC,YAAY;;WAGZ;WACA,OAAO;WACP;;WAEA;WACA,YAAY;;;;;;;;;iBAaL,sBACpB,OAAO,sBACN,QAAQ;;;UCtHM;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;;;UC3BlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}