@tangle-network/agent-eval 0.144.5 → 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.
- package/CHANGELOG.md +6 -0
- package/dist/{agent-profile-cell-Cw0PVwDr.d.ts → agent-profile-cell-BOP-iA9Q.d.ts} +2 -2
- package/dist/{agent-profile-cell-Cw0PVwDr.d.ts.map → agent-profile-cell-BOP-iA9Q.d.ts.map} +1 -1
- package/dist/analyst/index.d.ts +12 -12
- package/dist/analyst/index.d.ts.map +1 -1
- package/dist/analyst/index.js +3 -3
- package/dist/{benchmark-Fmo42QVE.d.ts → benchmark-J9Qe6j2_.d.ts} +3 -3
- package/dist/{benchmark-Fmo42QVE.d.ts.map → benchmark-J9Qe6j2_.d.ts.map} +1 -1
- package/dist/{benchmark-command-CA_NFOmy.js → benchmark-command-CQd78YHt.js} +6 -4
- package/dist/{benchmark-command-CA_NFOmy.js.map → benchmark-command-CQd78YHt.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +1 -1
- package/dist/benchmarks/index.js +1 -1
- package/dist/{benchmarks-CWbsj0t4.js → benchmarks-BEOkuvIg.js} +3 -3
- package/dist/{benchmarks-CWbsj0t4.js.map → benchmarks-BEOkuvIg.js.map} +1 -1
- package/dist/campaign/index.d.ts +6 -6
- package/dist/campaign/index.js +2 -2
- package/dist/{campaign-ClpnD7Ug.js → campaign-CXsdyym7.js} +3 -3
- package/dist/{campaign-ClpnD7Ug.js.map → campaign-CXsdyym7.js.map} +1 -1
- package/dist/cli.js +2 -2
- package/dist/{client-DAb7MWtL.d.ts → client-0JI64ovJ.d.ts} +4 -4
- package/dist/{client-DAb7MWtL.d.ts.map → client-0JI64ovJ.d.ts.map} +1 -1
- package/dist/{completion-verifier-VvpHRu78.d.ts → completion-verifier-CBiee74w.d.ts} +5 -5
- package/dist/{completion-verifier-VvpHRu78.d.ts.map → completion-verifier-CBiee74w.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +10 -10
- package/dist/contract/index.js +2 -2
- package/dist/control.d.ts +2 -2
- package/dist/{cost-ledger-FuQvHxPm.d.ts → cost-ledger-Bv_e8XHY.d.ts} +2 -2
- package/dist/{cost-ledger-FuQvHxPm.d.ts.map → cost-ledger-Bv_e8XHY.d.ts.map} +1 -1
- package/dist/{dataset-v_Y5902-.d.ts → dataset-C8xaLXdY.d.ts} +2 -2
- package/dist/{dataset-v_Y5902-.d.ts.map → dataset-C8xaLXdY.d.ts.map} +1 -1
- package/dist/{default-registry-RLNNoeEP.js → default-registry-Dta70shL.js} +2 -2
- package/dist/{default-registry-RLNNoeEP.js.map → default-registry-Dta70shL.js.map} +1 -1
- package/dist/{default-registry-BwbZ9N9v.d.ts → default-registry-J9m-_tya.d.ts} +5 -5
- package/dist/{default-registry-BwbZ9N9v.d.ts.map → default-registry-J9m-_tya.d.ts.map} +1 -1
- package/dist/{errors-DkfjIDvD.d.ts → errors-CKPfb2aH.d.ts} +2 -2
- package/dist/{errors-DkfjIDvD.d.ts.map → errors-CKPfb2aH.d.ts.map} +1 -1
- package/dist/errors-D-LKuDhb.js.map +1 -1
- package/dist/{eval-campaign-lZcDIwQM.js → eval-campaign-CfLQQs9B.js} +2 -2
- package/dist/{eval-campaign-lZcDIwQM.js.map → eval-campaign-CfLQQs9B.js.map} +1 -1
- package/dist/{exact-types-BQ7W90C4.d.ts → exact-types-CBYF5MGd.d.ts} +2 -2
- package/dist/{exact-types-BQ7W90C4.d.ts.map → exact-types-CBYF5MGd.d.ts.map} +1 -1
- package/dist/{external-optimizer-contracts-CdmX2K2S.d.ts → external-optimizer-contracts-Q9c0L3eY.d.ts} +3 -3
- package/dist/{external-optimizer-contracts-CdmX2K2S.d.ts.map → external-optimizer-contracts-Q9c0L3eY.d.ts.map} +1 -1
- package/dist/{feedback-trajectory-WK7x4mhy.d.ts → feedback-trajectory-GgoS0-MK.d.ts} +4 -4
- package/dist/{feedback-trajectory-WK7x4mhy.d.ts.map → feedback-trajectory-GgoS0-MK.d.ts.map} +1 -1
- package/dist/fuzz.d.ts +1 -1
- package/dist/hosted/index.d.ts +3 -3
- package/dist/{index-CsuAo2-J.d.ts → index-4XwggC10.d.ts} +5 -5
- package/dist/{index-CsuAo2-J.d.ts.map → index-4XwggC10.d.ts.map} +1 -1
- package/dist/{index-CNOCxBLh.d.ts → index-B6-B0zTB.d.ts} +2 -2
- package/dist/{index-CNOCxBLh.d.ts.map → index-B6-B0zTB.d.ts.map} +1 -1
- package/dist/{index-CpxZSlB7.d.ts → index-BIL5vxxt.d.ts} +12 -12
- package/dist/{index-CpxZSlB7.d.ts.map → index-BIL5vxxt.d.ts.map} +1 -1
- package/dist/{index-DEb46kc6.d.ts → index-CvSN3IG1.d.ts} +2 -2
- package/dist/{index-DEb46kc6.d.ts.map → index-CvSN3IG1.d.ts.map} +1 -1
- package/dist/{index-DGIzNtRv.d.ts → index-Dx1kF3Ez.d.ts} +3 -3
- package/dist/{index-DGIzNtRv.d.ts.map → index-Dx1kF3Ez.d.ts.map} +1 -1
- package/dist/index.d.ts +68 -77
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +56 -120
- package/dist/index.js.map +1 -1
- package/dist/{insight-report-D5m1z0_n.d.ts → insight-report-DqEsugpr.d.ts} +4 -4
- package/dist/{insight-report-D5m1z0_n.d.ts.map → insight-report-DqEsugpr.d.ts.map} +1 -1
- package/dist/{integrity-DRXobPEs.d.ts → integrity-CNGUaGBY.d.ts} +3 -3
- package/dist/{integrity-DRXobPEs.d.ts.map → integrity-CNGUaGBY.d.ts.map} +1 -1
- package/dist/ledger-core/index.d.ts +1 -1
- package/dist/{llm-client-D3EoChAU.js → llm-client-Dv5BiKLE.js} +289 -7
- package/dist/llm-client-Dv5BiKLE.js.map +1 -0
- package/dist/meta-eval/index.d.ts +2 -2
- package/dist/multishot/index.d.ts +2 -2
- package/dist/openapi.json +1 -1
- package/dist/profile-cell.d.ts +1 -1
- package/dist/{release-report-BZRWdq_t.d.ts → release-report-ChOgpIoQ.d.ts} +4 -4
- package/dist/{release-report-BZRWdq_t.d.ts.map → release-report-ChOgpIoQ.d.ts.map} +1 -1
- package/dist/{replay-B7S7Pdbw.d.ts → replay-Krvb114g.d.ts} +7 -7
- package/dist/{replay-B7S7Pdbw.d.ts.map → replay-Krvb114g.d.ts.map} +1 -1
- package/dist/reporting.d.ts +4 -4
- package/dist/{researcher-BSiCoM1s.d.ts → researcher-xLeNcpKX.d.ts} +6 -6
- package/dist/{researcher-BSiCoM1s.d.ts.map → researcher-xLeNcpKX.d.ts.map} +1 -1
- package/dist/{reward-hacking-DFgkEY4p.d.ts → reward-hacking-RZgnGWlx.d.ts} +2 -2
- package/dist/{reward-hacking-DFgkEY4p.d.ts.map → reward-hacking-RZgnGWlx.d.ts.map} +1 -1
- package/dist/rl.d.ts +5 -5
- package/dist/rl.js +1 -1
- package/dist/rollout/index.d.ts +1 -1
- package/dist/{rubric-predictive-validity-C4r4Y-q8.d.ts → rubric-predictive-validity-9qAwzkZm.d.ts} +2 -2
- package/dist/{rubric-predictive-validity-C4r4Y-q8.d.ts.map → rubric-predictive-validity-9qAwzkZm.d.ts.map} +1 -1
- package/dist/{run-evidence-j5Ynww6L.d.ts → run-evidence-C6G41MSI.d.ts} +3 -3
- package/dist/{run-evidence-j5Ynww6L.d.ts.map → run-evidence-C6G41MSI.d.ts.map} +1 -1
- package/dist/{run-record-ooo9FWns.d.ts → run-record-DdSa93_W.d.ts} +4 -4
- package/dist/{run-record-ooo9FWns.d.ts.map → run-record-DdSa93_W.d.ts.map} +1 -1
- package/dist/{semantic-concept-judge-DKRtp2sY.js → semantic-concept-judge-DwF6n05O.js} +2 -2
- package/dist/{semantic-concept-judge-DKRtp2sY.js.map → semantic-concept-judge-DwF6n05O.js.map} +1 -1
- package/dist/{server-Df00sdwz.js → server-D6XJQHw7.js} +2 -2
- package/dist/{server-Df00sdwz.js.map → server-D6XJQHw7.js.map} +1 -1
- package/dist/{skill-usage-DUvvudWR.d.ts → skill-usage-GlOphAhX.d.ts} +9 -9
- package/dist/{skill-usage-DUvvudWR.d.ts.map → skill-usage-GlOphAhX.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-CGz9ywhM.d.ts → skillopt-optimization-method-7S43rbDB.d.ts} +12 -12
- package/dist/{skillopt-optimization-method-CGz9ywhM.d.ts.map → skillopt-optimization-method-7S43rbDB.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-C4FX42dy.js → skillopt-optimization-method-Bfb-vBKe.js} +2 -2
- package/dist/{skillopt-optimization-method-C4FX42dy.js.map → skillopt-optimization-method-Bfb-vBKe.js.map} +1 -1
- package/dist/{statistics-B4u_CiFd.d.ts → statistics-C-dm-J6H.d.ts} +2 -2
- package/dist/{statistics-B4u_CiFd.d.ts.map → statistics-C-dm-J6H.d.ts.map} +1 -1
- package/dist/{summary-report-BOM6dfP7.d.ts → summary-report-B0cAyA7N.d.ts} +3 -3
- package/dist/{summary-report-BOM6dfP7.d.ts.map → summary-report-B0cAyA7N.d.ts.map} +1 -1
- package/dist/{tool-groups-ByZiqpVk.d.ts → tool-groups-CK0JCkqO.d.ts} +9 -9
- package/dist/{tool-groups-ByZiqpVk.d.ts.map → tool-groups-CK0JCkqO.d.ts.map} +1 -1
- package/dist/traces.d.ts +6 -6
- package/dist/{types-CZt1PBIk.d.ts → types-BhP9q0Fq.d.ts} +4 -4
- package/dist/{types-CZt1PBIk.d.ts.map → types-BhP9q0Fq.d.ts.map} +1 -1
- package/dist/{types-Dcoaqcsc.d.ts → types-DOZyvsFU.d.ts} +5 -5
- package/dist/{types-Dcoaqcsc.d.ts.map → types-DOZyvsFU.d.ts.map} +1 -1
- package/dist/{types-BjMFz88h.d.ts → types-XMVEdrE_.d.ts} +207 -6
- package/dist/types-XMVEdrE_.d.ts.map +1 -0
- package/dist/wire/index.d.ts +3 -3
- package/dist/wire/index.js +1 -1
- package/docs/building-doctrine.md +15 -0
- package/package.json +3 -3
- package/dist/llm-client-D3EoChAU.js.map +0 -1
- package/dist/types-BjMFz88h.d.ts.map +0 -1
package/dist/profile-cell.d.ts
CHANGED
|
@@ -1,2 +1,2 @@
|
|
|
1
|
-
import { C as verifyAgentProfileCell, S as validateAgentProfileCell, _ as buildAgentInterfaceProfileCell, a as AgentProfileCellSchemaVersion, b as requireAgentProfileCell, c as AgentProfileHarness, d as AgentProfileKind, f as AgentProfileSource, g as assertRunAgentProfileCell, h as agentProfileCellKey, i as AgentProfileCellInput, l as AgentProfileJson, m as agentProfileCellHashMaterial, n as AgentInterfaceProfileLike, o as AgentProfileCellValidationError, p as AgentProfileSourceInput, r as AgentProfileCell, s as AgentProfileDimensionValue, t as AGENT_PROFILE_KINDS, u as AgentProfileJsonObject, v as buildAgentProfileCell, x as toAgentProfileJson, y as groupRunsByAgentProfileCell } from "./agent-profile-cell-
|
|
1
|
+
import { C as verifyAgentProfileCell, S as validateAgentProfileCell, _ as buildAgentInterfaceProfileCell, a as AgentProfileCellSchemaVersion, b as requireAgentProfileCell, c as AgentProfileHarness, d as AgentProfileKind, f as AgentProfileSource, g as assertRunAgentProfileCell, h as agentProfileCellKey, i as AgentProfileCellInput, l as AgentProfileJson, m as agentProfileCellHashMaterial, n as AgentInterfaceProfileLike, o as AgentProfileCellValidationError, p as AgentProfileSourceInput, r as AgentProfileCell, s as AgentProfileDimensionValue, t as AGENT_PROFILE_KINDS, u as AgentProfileJsonObject, v as buildAgentProfileCell, x as toAgentProfileJson, y as groupRunsByAgentProfileCell } from "./agent-profile-cell-BOP-iA9Q.js";
|
|
2
2
|
export { AGENT_PROFILE_KINDS, AgentInterfaceProfileLike, AgentProfileCell, AgentProfileCellInput, AgentProfileCellSchemaVersion, AgentProfileCellValidationError, AgentProfileDimensionValue, AgentProfileHarness, AgentProfileJson, AgentProfileJsonObject, AgentProfileKind, AgentProfileSource, AgentProfileSourceInput, agentProfileCellHashMaterial, agentProfileCellKey, assertRunAgentProfileCell, buildAgentInterfaceProfileCell, buildAgentProfileCell, groupRunsByAgentProfileCell, requireAgentProfileCell, toAgentProfileJson, validateAgentProfileCell, verifyAgentProfileCell };
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { o as FailureClass } from "./schema-BtVldJ3T.js";
|
|
2
|
-
import { a as RunRecord, s as RunSplitTag } from "./run-record-
|
|
3
|
-
import { a as DatasetScenario, o as DatasetSplit, r as DatasetManifest } from "./dataset-
|
|
4
|
-
import { x as GateDecision } from "./summary-report-
|
|
2
|
+
import { a as RunRecord, s as RunSplitTag } from "./run-record-DdSa93_W.js";
|
|
3
|
+
import { a as DatasetScenario, o as DatasetSplit, r as DatasetManifest } from "./dataset-C8xaLXdY.js";
|
|
4
|
+
import { x as GateDecision } from "./summary-report-B0cAyA7N.js";
|
|
5
5
|
//#region src/release-confidence.d.ts
|
|
6
6
|
/** Severity of an actionable finding attached to a run/trace. */
|
|
7
7
|
type AsiSeverity = 'info' | 'warning' | 'error' | 'critical';
|
|
@@ -241,4 +241,4 @@ interface RenderReleaseReportOptions {
|
|
|
241
241
|
declare function renderReleaseReport(scorecard: ReleaseConfidenceScorecard, options?: RenderReleaseReportOptions): string;
|
|
242
242
|
//#endregion
|
|
243
243
|
export { ReleaseConfidenceStatus as _, JudgeReplayGateArgs as a, assertReleaseConfidence as b, judgeReplayGate as c, ReleaseConfidenceAxis as d, ReleaseConfidenceAxisName as f, ReleaseConfidenceScorecard as g, ReleaseConfidenceMetrics as h, BootstrapResult as i, ActionableSideInfo as l, ReleaseConfidenceIssue as m, renderReleaseReport as n, Verdict as o, ReleaseConfidenceInput as p, BootstrapOptions as r, bootstrapCi as s, RenderReleaseReportOptions as t, AsiSeverity as u, ReleaseConfidenceThresholds as v, evaluateReleaseConfidence as x, ReleaseTraceEvidence as y };
|
|
244
|
-
//# sourceMappingURL=release-report-
|
|
244
|
+
//# sourceMappingURL=release-report-ChOgpIoQ.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"release-report-
|
|
1
|
+
{"version":3,"file":"release-report-ChOgpIoQ.d.ts","names":[],"sources":["../src/release-confidence.ts","../src/promotion-gate.ts","../src/release-report.ts"],"mappings":";;;;;;KAsBY;;UAGK;;EAEf;;EAEA;EACA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;EACA,WAAW;;KAGD;KACA;UAQK;EACf;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,eAAe;EACf,MAAM;EACN,WAAW;;UAGI;;EAEf;EACA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,UAAU;EACV,qBAAqB;EACrB,gBAAgB;EAChB,kBAAkB;EAClB,eAAe;EACf,aAAa;;UAGE;EACf,MAAM;EACN,QAAQ;EACR;EACA;;UAGe;EACf,MAAM;EACN;EACA;EACA;;UAGe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,aAAa,OAAO;EACpB,cAAc;EACd,oBAAoB,QAAQ,OAAO;EACnC,0BAA0B;;;;;;;EAO1B;;UAGe;EACf;EACA;EACA;EACA,QAAQ;EACR;EACA,MAAM;EACN,QAAQ;EACR,SAAS;EACT,SAAS;EACT,cAAc;EACd;;iBAkBc,0BACd,OAAO,yBACN;iBA4Ha,wBAAwB,OAAO,yBAAyB;;;;;;;;;;;;;;;;;;;;;;;;;;;KCxR5D;UAEK;EACf;EACA;;EAEA;;EAEA;;EAEA;;EAEA;EACA;EACA,SAAS;;UAGM;;EAEf;;EAEA;;;;;;EAMA;;EAEA;;;;;;;;;iBAUc,YACd,oBACA,qBACA,UAAS,mBACR;;;;;;;;;;;;;;;;UAsHc,oBAAoB;EACnC,iBAAiB;EACjB,kBAAkB;;EAElB,QAAQ,QAAQ,YAAY;EAC5B;EACA;;EAEA;;EAEA;;;;;iBAMoB,gBAAgB,SACpC,MAAM,oBAAoB,WACzB,QAAQ;EAAoB;EAAyB;;;;UCxMvC;EACf;EACA,gBAAgB;EAChB;EACA;EACA;;iBAGc,oBACd,WAAW,4BACX,UAAS"}
|
|
@@ -1,12 +1,12 @@
|
|
|
1
|
-
import { c as ReplayError, l as ValidationError, o as LimitExceededError, r as CaptureIntegrityError, s as NotFoundError, t as AgentEvalError } from "./errors-
|
|
2
|
-
import { c as CostLedgerHandle } from "./cost-ledger-
|
|
1
|
+
import { c as ReplayError, l as ValidationError, o as LimitExceededError, r as CaptureIntegrityError, s as NotFoundError, t as AgentEvalError } from "./errors-CKPfb2aH.js";
|
|
2
|
+
import { c as CostLedgerHandle } from "./cost-ledger-Bv_e8XHY.js";
|
|
3
3
|
import { S as ToolSpan, f as Run } from "./schema-BtVldJ3T.js";
|
|
4
|
-
import { a as RunRecord, l as RunTerminalOutcome, s as RunSplitTag, u as RunTokenUsage } from "./run-record-
|
|
5
|
-
import {
|
|
4
|
+
import { a as RunRecord, l as RunTerminalOutcome, s as RunSplitTag, u as RunTokenUsage } from "./run-record-DdSa93_W.js";
|
|
5
|
+
import { Nt as RawProviderEvent, Pt as RawProviderSink } from "./types-XMVEdrE_.js";
|
|
6
6
|
import { s as TraceStore } from "./store-CT9YIIve.js";
|
|
7
7
|
import { n as RunCompleteHookContext, t as RunCompleteHook } from "./emitter-DGQGoLyj.js";
|
|
8
|
-
import { D as SearchSpanResult, E as QueryTracesPage, F as TraceAnalystSpanStatus, L as ViewSpansResult, M as TraceAnalystFilters, O as SearchTraceResult, P as TraceAnalystSpanKind, S as TraceAnalysisStoreContext, f as AnalystUsageReceipt, i as AnalystFinding, w as DatasetOverview, x as TraceAnalysisStore, y as BoundedTraceAnalysisStoreOptions, z as ViewTraceResult } from "./types-
|
|
9
|
-
import { i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, s as TraceAnalystLimits, t as TraceToolGroupName } from "./tool-groups-
|
|
8
|
+
import { D as SearchSpanResult, E as QueryTracesPage, F as TraceAnalystSpanStatus, L as ViewSpansResult, M as TraceAnalystFilters, O as SearchTraceResult, P as TraceAnalystSpanKind, S as TraceAnalysisStoreContext, f as AnalystUsageReceipt, i as AnalystFinding, w as DatasetOverview, x as TraceAnalysisStore, y as BoundedTraceAnalysisStoreOptions, z as ViewTraceResult } from "./types-BhP9q0Fq.js";
|
|
9
|
+
import { i as TraceAnalysisEngine, o as TraceAnalysisEngineResult, s as TraceAnalystLimits, t as TraceToolGroupName } from "./tool-groups-CK0JCkqO.js";
|
|
10
10
|
//#region src/trace/otel.d.ts
|
|
11
11
|
declare const OTEL_AGENT_EVAL_SCOPE: {
|
|
12
12
|
name: string;
|
|
@@ -830,4 +830,4 @@ declare function iterateRawCalls(sink: RawProviderSink, filter?: {
|
|
|
830
830
|
}): AsyncGenerator<ReplayCacheEntry>;
|
|
831
831
|
//#endregion
|
|
832
832
|
export { traceAnalystOnRunComplete as $, stringField as A, OtlpSpanRoleInput as At, TraceInsightReadiness as B, exportRunAsOtlp as Bt, ProjectedOtlpSpan as C, createOtelTracingStore as Ct, inferOtlpKind as D, OtelExporter as Dt, firstStringAttr as E, OtelExportConfig as Et, TraceInsightFinding as F, traceSpanKindToOpenInferenceKind as Ft, defaultTraceInsightPanel as G, TraceInsightTask as H, TraceInsightPanelRole as I, OTEL_AGENT_EVAL_SCOPE as It, inferDomainKeywords as J, describeTraceInsightScope as K, TraceInsightPromptInput as L, OtlpExport as Lt, flattenOtlpExportToNdjson as M, applyToolSpanOtlpAttributes as Mt, OtlpFlatLine as N, classifyOtlpSpanRole as Nt, projectOtlpFlatLine as O, createOtelExporter as Ot, TraceInsightContext as P, isOtlpModelCall as Pt, TraceAnalystHookOptions as Q, TraceInsightQualityGate as R, OtlpResourceSpans as Rt, otlpToTraceRunRecords as S, redactValue as St, extractOtlpAttributes as T, ExportableSpan as Tt, buildTraceInsightContext as U, TraceInsightSuite as V, buildTraceInsightPrompt as W, scoreTraceInsightReadiness as X, planTraceInsightQuestions as Y, tokenizeDomainWords as Z, OtlpTraceRunRecord as _, DEFAULT_REDACTION_RULES as _t, ReplayFetchOptions as a, TraceFileMissingError as at, otlpRowsToTraceRunRecords as b, RedactionRule as bt, ToolTraceMissingError as c, AnalyzeTracesInput as ct, OtlpFileTraceStoreOptions as d, analyzeTraces as dt, SpanNotFoundError as et, ToolSpansToTraceAnalysisStoreOptions as f, createBoundedTraceAnalysisStore as ft, OtlpToRunRecordsOptions as g, convertTraceStoresToOtlp as gt, TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION as h, TracesToOtlpResult as ht, ReplayCacheStats as i, TraceFileMalformedError as it, FlattenOtlpOptions as j, ToolSpanOtlpInput as jt, readOtlpStatus as k, OtlpSpanRole as kt, toolSpansToTraceAnalysisStore as l, AnalyzeTracesOptions as lt, TRACE_ANALYST_ACTOR_DESCRIPTION as m, TraceStoreToOtlpOptions as mt, ReplayCacheEntry as n, TraceAnalysisStoreContractError as nt, createReplayFetch as o, TraceFileTooLargeError as ot, otlpTextToTraceAnalysisStore as p, TraceStoreSource as pt, domainEvidencePattern as q, ReplayCacheMissError as r, TraceAnalysisValidationError as rt, iterateRawCalls as s, TraceNotFoundError as st, ReplayCache as t, TraceAnalysisLimitError as tt, OtlpFileTraceStore as u, AnalyzeTracesResult as ut, TraceAggregate as v, REDACTION_VERSION as vt, asString as w, otelRunCompleteHook as wt, otlpToRunRecords as x, redactString as xt, otlpRowsToRunRecords as y, RedactionReport as yt, TraceInsightQuestion as z, OtlpSpan as zt };
|
|
833
|
-
//# sourceMappingURL=replay-
|
|
833
|
+
//# sourceMappingURL=replay-Krvb114g.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"replay-
|
|
1
|
+
{"version":3,"file":"replay-Krvb114g.d.ts","names":[],"sources":["../src/trace/otel.ts","../src/trace/otlp-attributes.ts","../src/trace/otel-export.ts","../src/trace/otel-bridge.ts","../src/trace/redact.ts","../src/trace/store-to-otlp.ts","../src/trace-analyst/store-boundary.ts","../src/trace-analyst/analyst.ts","../src/trace-analyst/errors.ts","../src/trace-analyst/hook.ts","../src/trace-analyst/insights.ts","../src/trace/otlp-flat.ts","../src/trace-analyst/otlp-flatten.ts","../src/trace-analyst/otlp-span.ts","../src/trace-analyst/otlp-to-run-records.ts","../src/trace-analyst/prompts.ts","../src/trace-analyst/store-otlp.ts","../src/trace-analyst/store-tool-spans.ts","../src/replay.ts"],"mappings":";;;;;;;;;;cAqBa;EAA0B;EAAoC;;UAE1D;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,YAAY;IACV;IACA;MAAS;MAAsB;MAAmB;MAAsB;;;EAE1E,SAAS;IAAQ;IAAsB;IAAc,aAAa;;EAClE;IAAW;IAAc;;;UAGV;EACf;IAAY,YAAY;;EACxB,YAAY;IAAQ,cAAc;IAAuB,OAAO;;;UAGjD;EACf,eAAe;;;iBAIK,gBACpB,OAAO,YACP,eACA,gBAAe,4CACd,QAAQ;;;KC7BC,oBAAoB,KAC9B;KAIU;UAUK;EACf;EACA,YAAY;EACZ;;;;;;;iBAkBc,qBAAqB,OAAO,oBAAoB;iBA2BhD,gBAAgB,OAAO;iBAkBvB,4BACd,YAAY,yBACZ,MAAM;iBAMQ,iCAAiC;;;UC9FhC;;EAEf;;EAEA,UAAU;;EAEV;;EAEA;;EAEA,qBAAqB;;EAErB;;UAGe;;EAEf,WAAW,MAAM;;EAEjB,SAAS;;EAET,YAAY;;UAGG;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP,aAAa;;;;;;iBAOC,mBAAmB,SAAS,mBAAmB;;;;;;;iBC/C/C,oBAAoB,UAAU,eAAe;;;;;;;;;iBAoB7C,uBACd,OAAO,YACP,UAAU,cACV,kBACC;;;;;;;;;;;;;;;;UC/Bc;EACf;EACA,SAAS;;EAET;;UAGe;EACf;EACA,QAAQ;;;cAIG,yBAAyB;cAezB;;;;;iBAMG,aACd,eACA,QAAO;EACJ;EAAgB,QAAQ;;;;;;;iBAuBb,YACd,gBACA,QAAO,iBACP,SAAQ;EACL;EAAgB,QAAQ;;;;UCzCZ;;EAEf;;EAEA;;EAEA;;;EAGA;;;;;;;;;;;;UAae;;EAEf;;EAEA;;EAEA;;;;UAKe;;;EAGf;;;EAGA,sBAAsB,KAAK,QAAQ;;EAEnC,iBAAiB,KAAK,QAAQ;;;;;;;;;;;iBAYhB,yBACd,iBAAiB,4BAA4B,oBAC7C,iBACA,OAAM,0BACL;;;;iBCjEa,gCACd,QAAQ,oBACR,UAAS,mCACR;;;UClBc;EACf;EACA;EACA;EACA;;KAGU,sBAAsB;UAEjB;;EAEf,iBAAiB;;EAEjB,QAAQ;;EAER;;EAEA,YAAY;EACZ,SAAS,QAAQ;EACjB;EACA;EACA,aAAa;EACb;EACA,yBAAyB;EACzB,4BAA4B;EAC5B,eAAe,SAAS;EACxB,OAAO;EACP,OAAO,iBAAiB,SAAS;EACjC,SAAS;;;;;;;;iBASW,cACpB,OAAO,oBACP,SAAS,uBACR,QAAQ;;;;cC7CE,qCAAqC;WAC/B;;;cAIN,gCAAgC;WAC1B;WACR;WACA;WACA;EAET,YAAY,mBAAmB,gBAAgB,eAAe;;;cASnD,wCAAwC;WAClC;WACR;EAET,YAAY,mBAAmB,iBAAiB;IAAY;;;cAMjD,8BAA8B;WACxB;WACR;EAET,YAAY;;cAMD,+BAA+B;WACjC;WACA;WACA;EAET,YAAY,cAAc,oBAAoB;;cAcnC,gCAAgC;WAClC;WACA;WACA;EAET,YAAY,cAAc,qBAAqB,qBAAqB;;cAQzD,2BAA2B;WACrB;WACR;EAET,YAAY;;cAMD,0BAA0B;WACpB;WACR;WACA;EAET,YAAY,kBAAkB;;;;UC1Ef;;;;;;EAMf,SAAS,KAAK;IAAoC,SAAS;;;;;;;EAM3D;;;;;;EAMA,QAAQ,QAAQ,qBAAqB,KAAK,2BAA2B;;;;;EAKrE,aAAa,KAAK;;;;;;EAMlB,UAAU,QAAQ,qBAAqB,KAAK;;iBAM9B,0BAA0B,MAAM,0BAA0B;;;UCzDzD;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,OAAO;;UAGQ;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf,OAAO;EACP,WAAW;EACX,QAAQ;EACR,SAAS;EACT;;UAGe;EACf,OAAO;EACP;EACA;EACA,WAAW;EACX,OAAO;EACP,UAAU;EACV,OAAO;EACP,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,OAAO;;iBA0BO,oBAAoB;iBAMpB,oBAAoB,OAAO;iBAuB3B,sBAAsB,qBAAqB;iBAS3C,0BAA0B,OAAO;iBAsBjC,0BAA0B,OAAO,0BAA0B;iBAsD3D,yBAAyB,OAAO,0BAA0B;iBAa1D,2BAA2B,SAAS,sBAAsB;iBA2D1D,4BAA4B;iBA6B5B,wBAAwB,OAAO;;;KChTnC;UAEK;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IACE,MAAM;IACN;;EAEF;IAAY,YAAY;;EACxB,YAAY;EACZ,SAAS;IAAQ;IAAc;IAAuB,aAAa;;;;;UCEpD;;;;EAIf;;EAEA,UAAU,QAAQ;;iBA+DJ,0BACd,YAAY,YACZ,OAAM,qBACL;;;;;;;;;UC1Dc;EACf;EACA;EACA;EACA;EACA,MAAM;EACN;EACA;EACA;EACA,QAAQ;EACR;EACA;EACA;EACA;EACA;;EAEA,YAAY;;;;;;iBAOE,oBAAoB,KAAK,0BAA0B;iBAkEnD,eAAe,KAAK;EAClC,MAAM;EACN;;iBAkBc,cAAc,OAAO,0BAA0B;;;;;;iBAwB/C,sBAAsB,KAAK,0BAA0B;iBAoBrD,YAAY,KAAK,yBAAyB;iBAK1C,SAAS;;iBAKT,gBACd,OAAO,yBACP;;;UCnIe;;EAEf;;;EAGA;;;EAGA,WAAW;;EAEX;;EAEA;;EAEA;;EAEA;;;;;;;EAOA;;;;;;EAMA;;;;;;;;;EASA,wBAAwB;;;;;;;;;EASxB,iBAAiB,eAAe,MAAM;;;;;EAKtC,yBAAyB,kBAAkB;;;;;;UAO5B;EACf,QAAQ;;EAER;;EAEA;;;UAIe;EACf;;EAEA;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,YAAY;;EAEZ;EACA;EACA;EACA;EACA;;EAEA,iBAAiB;;;;;;;iBAeH,iBAAiB,mBAAmB,MAAM,0BAA0B;;;;;;iBASpE,qBACd,MAAM,kBACN,MAAM,0BACL;;iBAKa,sBACd,mBACA,MAAM,0BACL;;iBAQa,0BACd,MAAM,kBACN,MAAM,0BACL;;;;cC/MU;cAcA;;;UC0HI;;EAEf;;EAEA;;EAEA;;EAEA;;KAGG,gCAAgC;UAEpB,kCAAkC;;EAEjD;;;;;;;;EAQA;;uBAMa,kCAAkC;mBAC9B;mBACA;mBACA;mBACA;UACT;UACA;UACA;UACA;EAER,YAAY,MAAM;EAyBZ,SAAS,kBAAkB,UAAU,4BAA4B;EAOjE,SACJ;IAAQ;IAAkB;KAC1B,UAAU,4BACT;EAeG,YACJ,UAAU,qBACV,UAAU,4BACT,QAAQ;EAoDL,YACJ;IAAQ,UAAU;IAAqB;IAAe;KACtD,UAAU,4BACT,QAAQ;EAmBL,YACJ,UAAU,qBACV,UAAU,4BACT;EAMG,UACJ;IAAQ;IAAkB;KAC1B,UAAU,4BACT,QAAQ;EAuCL,UACJ;IAAQ;IAAkB;IAA6B;KACvD,UAAU,4BACT,QAAQ;EAsFL,YACJ;IAAQ;IAAkB;IAAuB;KACjD,UAAU,4BACT,QAAQ;EA4CL,WACJ;IAAQ;IAAkB;IAAiB;IAAuB;KAClE,UAAU,4BACT,QAAQ;;;EA6CL,cAAc,UAAU,4BAA4B;UAI5C;qBAmBK,WAAW,SAAS,cAAc,QAAQ;qBAE1C;UAEL;UAmBA;UAsHA;UA+DN;UAkBA;UAwDA;UAmBM;;cA+CH,2BAA2B;mBACrB;mBACA;EAEjB,YAAY,MAAM;;;YAYF,WAAW,SAAS,cAAc,QAAQ;YAkBhD;;;iBAwBI,6BACd,cACA,OAAM,uCACL;;;;cC/7BU,8BAA8B;EACzC;;;;;;iBASc,8BACd,gBAAgB,+BAChB,UAAS,uCACR;;;cCSU,6BAA6B;WAEtB;WACA;EAFlB,YACkB,aACA,oBAChB;;UAMa;EACf,SAAS;EACT,UAAU;;UAGK;EACf;EACA,YAAY;EACZ,SAAS;;EAET;;;;;;;;;;cAWW;UACH;UACA;UACA;UACA;;;;;SAMK,SACX,MAAM,iBACN;IAAU;IAAgB;MACzB,QAAQ;;SASE,WAAW,QAAQ,qBAAqB,QAAQ;;EA4B7D;EAIA,SAAS;;EAUR,WAAW,iBAAiB;;;;;;EASvB,OAAO,uBAAuB,QAAQ;;UAM7B;;;;;;;EAOf;EACA,uBAAuB;;EAEvB,SAAS;IAAQ;IAAa;IAAkB;;;EAEhD,gBAAgB;IAAQ;IAAa;;;;;;;;;;;;;iBAavB,kBAAkB,OAAO,aAAa,OAAM,4BAAiC;;;;;;iBAsDtE,gBACrB,MAAM,iBACN;EAAU;EAAgB;IACzB,eAAe"}
|
package/dist/reporting.d.ts
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { N as benjaminiHochberg, Q as pairedBootstrap, gt as wilcoxonSignedRank, v as PairedBootstrapOptions, y as PairedBootstrapResult } from "./statistics-
|
|
2
|
-
import { _ as paretoChart, a as ParetoPoint, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, y as summaryTable } from "./summary-report-
|
|
3
|
-
import { _ as ReleaseConfidenceStatus, a as JudgeReplayGateArgs, b as assertReleaseConfidence, c as judgeReplayGate, d as ReleaseConfidenceAxis, f as ReleaseConfidenceAxisName, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceMetrics, i as BootstrapResult, m as ReleaseConfidenceIssue, n as renderReleaseReport, o as Verdict, p as ReleaseConfidenceInput, r as BootstrapOptions, s as bootstrapCi, t as RenderReleaseReportOptions, v as ReleaseConfidenceThresholds, x as evaluateReleaseConfidence, y as ReleaseTraceEvidence } from "./release-report-
|
|
1
|
+
import { N as benjaminiHochberg, Q as pairedBootstrap, gt as wilcoxonSignedRank, v as PairedBootstrapOptions, y as PairedBootstrapResult } from "./statistics-C-dm-J6H.js";
|
|
2
|
+
import { _ as paretoChart, a as ParetoPoint, c as ResearchReportCandidate, d as ResearchReportOptions, f as ResearchReportRecommendation, g as gainHistogram, h as SummaryTableRow, i as ParetoFigureSpec, l as ResearchReportDecision, m as SummaryTableOptions, n as GainDistributionFigureSpec, o as RESEARCH_REPORT_HARD_PAIR_FLOOR, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, u as ResearchReportMethodology, v as researchReport, y as summaryTable } from "./summary-report-B0cAyA7N.js";
|
|
3
|
+
import { _ as ReleaseConfidenceStatus, a as JudgeReplayGateArgs, b as assertReleaseConfidence, c as judgeReplayGate, d as ReleaseConfidenceAxis, f as ReleaseConfidenceAxisName, g as ReleaseConfidenceScorecard, h as ReleaseConfidenceMetrics, i as BootstrapResult, m as ReleaseConfidenceIssue, n as renderReleaseReport, o as Verdict, p as ReleaseConfidenceInput, r as BootstrapOptions, s as bootstrapCi, t as RenderReleaseReportOptions, v as ReleaseConfidenceThresholds, x as evaluateReleaseConfidence, y as ReleaseTraceEvidence } from "./release-report-ChOgpIoQ.js";
|
|
4
4
|
import { a as PairedEvalueStep, c as pairedEvalueSequence, i as PairedEvalueSequence, n as InterimReleaseConfidenceInput, o as SequentialDecision, r as PairedEvalueOptions, s as evaluateInterimReleaseConfidence, t as InterimReleaseConfidence } from "./sequential-CYwq6Ff_.js";
|
|
5
|
-
import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "./rubric-predictive-validity-
|
|
5
|
+
import { a as rubricPredictiveValidity, i as RubricRanking, n as RubricPredictiveValidityInput, r as RubricPredictiveValidityReport, t as RubricOutcomePair } from "./rubric-predictive-validity-9qAwzkZm.js";
|
|
6
6
|
export { type BootstrapOptions, type BootstrapResult, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, type JudgeReplayGateArgs, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type ParetoFigureSpec, type ParetoPoint, RESEARCH_REPORT_HARD_PAIR_FLOOR, type ReleaseConfidenceAxis, type ReleaseConfidenceAxisName, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseConfidenceStatus, type ReleaseConfidenceThresholds, type ReleaseTraceEvidence, type RenderReleaseReportOptions, type ResearchReport, type ResearchReportCandidate, type ResearchReportDecision, type ResearchReportMethodology, type ResearchReportOptions, type ResearchReportRecommendation, type RubricOutcomePair, type RubricPredictiveValidityInput, type RubricPredictiveValidityReport, type RubricRanking, type SequentialDecision, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type Verdict, assertReleaseConfidence, benjaminiHochberg, bootstrapCi, evaluateInterimReleaseConfidence, evaluateReleaseConfidence, gainHistogram, judgeReplayGate, pairedBootstrap, pairedEvalueSequence, paretoChart, renderReleaseReport, researchReport, rubricPredictiveValidity, summaryTable, wilcoxonSignedRank };
|
|
@@ -1,10 +1,10 @@
|
|
|
1
|
-
import { i as AgentProfileCellInput, r as AgentProfileCell } from "./agent-profile-cell-
|
|
2
|
-
import { a as RunRecord, c as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, s as RunSplitTag, t as JudgeScoresRecord, u as RunTokenUsage } from "./run-record-
|
|
3
|
-
import {
|
|
1
|
+
import { i as AgentProfileCellInput, r as AgentProfileCell } from "./agent-profile-cell-BOP-iA9Q.js";
|
|
2
|
+
import { a as RunRecord, c as RunTaskFailure, n as RunCostProvenance, r as RunJudgeMetadata, s as RunSplitTag, t as JudgeScoresRecord, u as RunTokenUsage } from "./run-record-DdSa93_W.js";
|
|
3
|
+
import { Pt as RawProviderSink, Z as LlmRouteRequirements, q as LlmClientOptions } from "./types-XMVEdrE_.js";
|
|
4
4
|
import { s as TraceStore } from "./store-CT9YIIve.js";
|
|
5
5
|
import { i as TraceEmitter, t as RunCompleteHook } from "./emitter-DGQGoLyj.js";
|
|
6
|
-
import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-
|
|
7
|
-
import { d as ResearchReportOptions, s as ResearchReport, x as GateDecision } from "./summary-report-
|
|
6
|
+
import { a as RunIntegrityReport, n as RunIntegrityExpectations } from "./integrity-CNGUaGBY.js";
|
|
7
|
+
import { d as ResearchReportOptions, s as ResearchReport, x as GateDecision } from "./summary-report-B0cAyA7N.js";
|
|
8
8
|
//#region src/eval-campaign.d.ts
|
|
9
9
|
interface CampaignVariant<V> {
|
|
10
10
|
id: string;
|
|
@@ -311,4 +311,4 @@ declare class NoopResearcher implements Researcher {
|
|
|
311
311
|
}
|
|
312
312
|
//#endregion
|
|
313
313
|
export { EvalCampaignResult as _, FailureMode as a, SteeringChange as c, CampaignRunContext as d, CampaignRunOutcome as f, EvalCampaignOptions as g, CampaignVariant as h, ExperimentResult as i, CampaignFactoryParams as l, CampaignScenario as m, CallbackResearcherOptions as n, NoopResearcher as o, CampaignRunner as p, ExperimentPlan as r, Researcher as s, CallbackResearcher as t, CampaignIntegrityPolicy as u, FailedRun as v, runEvalCampaign as y };
|
|
314
|
-
//# sourceMappingURL=researcher-
|
|
314
|
+
//# sourceMappingURL=researcher-xLeNcpKX.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"researcher-
|
|
1
|
+
{"version":3,"file":"researcher-xLeNcpKX.d.ts","names":[],"sources":["../src/eval-campaign.ts","../src/researcher.ts"],"mappings":";;;;;;;;UAyEiB,gBAAgB;EAC/B;EACA,SAAS;;UAGM;EACf;;EAEA,OAAO;;UAGQ,mBAAmB;;EAElC;;EAEA;EACA,SAAS;EACT;EACA;EACA,cAAc;EACd;EACA,UAAU;;;;;;;EAOV,SAAS;EACT,OAAO;EACP,SAAS;;;;;;EAMT,SAAS;;UAGD;;EAER;;EAEA;;EAEA;;EAEA,gBAAgB;EAChB,YAAY;;EAEZ;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,gBAAgB;;;;;;EAMhB,cAAc;;;;;;EAMd,eAAe,mBAAmB;;;KAIxB,qBAAqB,2BAA2B;KAEhD,eAAe,MAAM,KAAK,mBAAmB,OAAO,QAAQ;KAE5D;UAEK,oBAAoB;;;;;EAKnC;EACA,UAAU,gBAAgB;EAC1B,WAAW;;EAEX;;EAEA,WAAW;;EAEX;;;;;;EAMA,SAAS;;;;;;EAMT,oBAAoB;;;;;;EAMpB,eAAe,QAAQ,0BAA0B;;;;;;;EAOjD,kBAAkB,QAAQ,0BAA0B;;;;;EAKpD;;;;;EAKA,gBAAgB;;;;;;EAMhB,YAAY;;EAEZ,qBAAqB;;;;;EAKrB,QAAQ,eAAe;;;;;EAKvB;IAAW;MAAwB,KACjC;;;;;EAOF;;EAEA;;;;EAIA;;EAEA,SAAS,QAAQ;;;;;;;EAOjB,eACI,mBACA,0BAEE,QAAQ;IACN,SAAS;IACT,cAAc;QAGd,mBACA,wBACA,QAAQ,mBAAmB;;UAGpB;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;EAEA,MAAM;;EAEN,kBAAkB;EAClB,YAAY;;EAEZ,SAAS;EACT;EACA;;iBAgBoB,gBAAgB,GACpC,MAAM,oBAAoB,KACzB,QAAQ;;;;UC5QM;;;EAGf;;EAEA;EACA;;;IAGE;;IAEA;;;;UAKa;EACf;;;;EAIA;;;EAGA;;EAEA;;;UAIe;EACf;EACA;EACA,SAAS;;;EAGT;;EAEA;IAAU;IAAkB;;;;UAIb;EACf,MAAM;EACN,MAAM;EACN,cAAc;;;;;;;;;;;;;;;;;;;UAoBC;EACf,gBAAgB,MAAM,cAAc,QAAQ;EAC5C,cAAc,UAAU,gBAAgB,QAAQ;EAChD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAC1E,eAAe,MAAM,iBAAiB,QAAQ;;UAG/B;EACf,iBAAiB;EACjB,eAAe;EACf,aAAa;EACb,gBAAgB;;;;;;cAOL,8BAA8B;mBACZ;EAA7B,YAA6B,WAAW;EAExC,gBAAgB,MAAM,cAAc,QAAQ;EAI5C,cAAc,UAAU,gBAAgB,QAAQ;EAIhD,YAAY,SAAS,kBAAkB,UAAU,iBAAiB,QAAQ;EAI1E,eAAe,MAAM,iBAAiB,QAAQ;;;;;;;;;cAYnC,0BAA0B;mBACpB;EAEjB,YAAY;EAIN,gBAAgB,OAAO,cAAc,QAAQ;EAI7C,cAAc,WAAW,gBAAgB,QAAQ;EAIjD,YACJ,UAAU,kBACV,WAAW,iBACV,QAAQ;EAIL,eAAe,OAAO,iBAAiB,QAAQ"}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { s as VerificationReport } from "./multi-layer-verifier-BHY1gWAc.js";
|
|
2
|
-
import { a as RunRecord } from "./run-record-
|
|
2
|
+
import { a as RunRecord } from "./run-record-DdSa93_W.js";
|
|
3
3
|
//#region src/rl/verifiable-reward.d.ts
|
|
4
4
|
type VerifiableRewardSource = 'compile' | 'test' | 'schema' | 'sandbox' | 'judge' | 'composite';
|
|
5
5
|
interface VerifiableReward {
|
|
@@ -221,4 +221,4 @@ interface DetectRewardHackingInput {
|
|
|
221
221
|
declare function detectRewardHacking(input: DetectRewardHackingInput): RewardHackingReport;
|
|
222
222
|
//#endregion
|
|
223
223
|
export { detectRewardHacking as a, VerifiableRewardSource as c, filterDeterministicallyRewarded as d, RewardHackingSignal as i, extractVerifiableReward as l, RewardHackingFinding as n, VerifiableReward as o, RewardHackingReport as r, VerifiableRewardExtractionOptions as s, DetectRewardHackingInput as t, extractVerifiableRewardsFromRecords as u };
|
|
224
|
-
//# sourceMappingURL=reward-hacking-
|
|
224
|
+
//# sourceMappingURL=reward-hacking-RZgnGWlx.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"reward-hacking-
|
|
1
|
+
{"version":3,"file":"reward-hacking-RZgnGWlx.d.ts","names":[],"sources":["../src/rl/verifiable-reward.ts","../src/rl/reward-hacking.ts"],"mappings":";;;KAiCY;UAQK;;EAEf;;EAEA,QAAQ;;;;;;;;EAQR;;;;;;EAMA;;EAEA;;;;;;;EAOA,YAAY;;;;;;;;;;;EAWZ;;;;;;;;;;;;;;;;;;EAkBA;;UAGe;;;;;;;;EAQf;;;;EAIA,aAAa,sBAAsB;;;;;;;EAOnC;;;;;EAKA;;;;;;;;;;;;EAYA;;;;;;;;;;;;;;;iBA6Cc,wBACd,QAAQ,oBACR,OAAM,oCACL;;;;;;;;;;;;;;;;;;;;;;;iBAqGa,oCACd,MAAM,aACN,OAAM,oCACL;EAAQ;EAAe,QAAQ;;;;;;;;;;;iBA4GlB,gCACd,MAAM,aACN,OAAM,oCACL;EAAQ,KAAK;EAAW,QAAQ;;;;KChWvB;UAMK;EACf,QAAQ;;EAER;EACA;;EAEA,QAAQ;;UAGO;EACf,UAAU;;EAEV,kBAAkB;;;;;;EAMlB;;EAEA;;EAEA;;UAGe;;;;;EAKf,MAAM;;;;;EAKN,WAAW,KAAK;;;;;;EAMhB,WAAW,KAAK;;;;;;EAMhB,qBAAqB,KAAK;;;;;EAK1B;;;;;EAKA;IAAe;IAAkB;;;;;EAIjC,0BAA0B;;iBAgBZ,oBAAoB,OAAO,2BAA2B"}
|
package/dist/rl.d.ts
CHANGED
|
@@ -1,15 +1,15 @@
|
|
|
1
1
|
import { s as VerificationReport } from "./multi-layer-verifier-BHY1gWAc.js";
|
|
2
2
|
import { _ as Span } from "./schema-BtVldJ3T.js";
|
|
3
|
-
import { a as RunRecord, s as RunSplitTag } from "./run-record-
|
|
3
|
+
import { a as RunRecord, s as RunSplitTag } from "./run-record-DdSa93_W.js";
|
|
4
4
|
import { s as TraceStore } from "./store-CT9YIIve.js";
|
|
5
|
-
import { a as CampaignResult } from "./types-
|
|
5
|
+
import { a as CampaignResult } from "./types-DOZyvsFU.js";
|
|
6
6
|
import { b as RolloutSplit, o as MintedRolloutLine } from "./schema-Cef2cFmb.js";
|
|
7
|
-
import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-
|
|
7
|
+
import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-RZgnGWlx.js";
|
|
8
8
|
import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "./outcome-store-BYHIuO0e.js";
|
|
9
9
|
import { t as AdversarialMutation } from "./adversarial-smnADNFS.js";
|
|
10
|
-
import { _ as EvalCampaignResult, a as FailureMode, c as SteeringChange, g as EvalCampaignOptions, i as ExperimentResult, r as ExperimentPlan, s as Researcher, y as runEvalCampaign } from "./researcher-
|
|
10
|
+
import { _ as EvalCampaignResult, a as FailureMode, c as SteeringChange, g as EvalCampaignOptions, i as ExperimentResult, r as ExperimentPlan, s as Researcher, y as runEvalCampaign } from "./researcher-xLeNcpKX.js";
|
|
11
11
|
import { t as InterimReleaseConfidence } from "./sequential-CYwq6Ff_.js";
|
|
12
|
-
import { r as RubricPredictiveValidityReport } from "./rubric-predictive-validity-
|
|
12
|
+
import { r as RubricPredictiveValidityReport } from "./rubric-predictive-validity-9qAwzkZm.js";
|
|
13
13
|
//#region src/rl/active-curriculum.d.ts
|
|
14
14
|
interface CellObservation {
|
|
15
15
|
variantId: string;
|
package/dist/rl.js
CHANGED
|
@@ -7,7 +7,7 @@ import { t as isSplitEligible } from "./exporters-q9iL-2Jf.js";
|
|
|
7
7
|
import { t as mintRolloutRows } from "./mint-DD-0oQTA.js";
|
|
8
8
|
import { a as InMemoryTraceStore } from "./integrity-fdt8XPAv.js";
|
|
9
9
|
import { i as filterDeterministicallyRewarded, l as campaignCellToRunRecord, n as extractVerifiableReward, r as extractVerifiableRewardsFromRecords, t as detectRewardHacking } from "./reward-hacking-CyuzxKly.js";
|
|
10
|
-
import { t as runEvalCampaign } from "./eval-campaign-
|
|
10
|
+
import { t as runEvalCampaign } from "./eval-campaign-CfLQQs9B.js";
|
|
11
11
|
import { t as evaluateInterimReleaseConfidence } from "./sequential-Br0mAPHA.js";
|
|
12
12
|
import { t as rubricPredictiveValidity } from "./rubric-predictive-validity-BRR632r1.js";
|
|
13
13
|
import { n as thompsonCurriculum, r as varianceBasedCurriculum, t as observationsFromRunRecords } from "./active-curriculum-C4mk67HP.js";
|
package/dist/rollout/index.d.ts
CHANGED
|
@@ -1,3 +1,3 @@
|
|
|
1
1
|
import { A as isTrainableSplit, C as TRAINABLE_SPLITS, D as assertRolloutLine, E as assertMintedLines, O as gateGamedOutcome, S as RolloutTask, T as assertMinted, _ as RolloutPolicy, a as GatedEvidence, b as RolloutSplit, c as ROLLOUT_CAPTURES, d as ROLLOUT_SPLITS, f as RolloutArtifacts, g as RolloutOutcome, h as RolloutLine, i as ChatToolCall, j as validateRolloutLine, k as isRolloutLine, l as ROLLOUT_ROLES, m as RolloutCostBlock, n as ChatMessage, o as MintedRolloutLine, p as RolloutCapture, r as ChatRole, s as MintedRolloutOutcome, t as CHAT_ROLES, u as ROLLOUT_SCHEMA, v as RolloutProvenance, w as ToolDef, x as RolloutStep, y as RolloutRole } from "../schema-Cef2cFmb.js";
|
|
2
|
-
import { $ as ScoreOrigin, $t as toSftRows, A as ReleaseRowRef, At as GATE_CHECKS, B as readOpencodeSessionMessages, Bt as gatedEvidenceOf, C as scrubRolloutLine, Ct as HarborSubagentTrajectoryRef, D as FormatGateCounts, Dt as relabelImportedSplit, E as FORMAT_GATE_DISPOSITION, Et as fromHarborTrajectory, F as DEFAULT_OPENCODE_DB, Ft as GateCheckId, G as claudeProjectSlug, Gt as SftRow, H as ClaudeTranscriptRef, Ht as RewardRow, I as OpencodeSessionRow, It as GateCheckedOutcome, J as MintRolloutOptions, Jt as realnessLabels, K as findClaudeTranscripts, Kt as VerifiersRolloutOutput, L as findOpencodeSessionById, Lt as GateEntryPoint, M as gatedRolloutIds, Mt as GATE_POLICIES, N as measureFormatGate, Nt as GateCheck, O as GateDisposition, Ot as toHarborTrajectories, P as releaseRowRefs, Pt as GateCheckDisposition, Q as unmintableReasons, Qt as toRftItems, R as findOpencodeSessionsByDirectory, Rt as GatePolicy, S as scrubLines, St as HarborStepSource, T as EmittedEvidence, Tt as HarborTrajectory, U as ClaudeUsageTotals, Ut as RftItem, V as ClaudeTranscript, Vt as RealnessLabels, W as DEFAULT_CLAUDE_PROJECTS_DIR, Wt as SftExportOptions, X as RolloutScrubber, Xt as toRewardRows, Y as MintRolloutResult, Yt as toJsonl, Z as mintRolloutRows, Zt as toRftItem, _ as ScrubCounts, _t as HarborImageSource, a as ScrubReport, at as trainingReward, b as defaultRolloutScrubber, bt as HarborObservationResult, c as planPushCommand, ct as readRolloutJournal, d as DatasetCardInputs, dt as ATIF_SCHEMA_VERSION, en as toVerifiersRolloutOutput, et as ScorePreference, f as FORMAT_FILES, ft as FromHarborOptions, g as SCRUB_RULES, gt as HarborFinalMetrics, h as buildDatasetCard, ht as HarborContentPart, i as RolloutReleaseCliArgs, it as scoreOrigin, j as assertGateReport, jt as GATE_CHECK_IDS, k as GateReport, kt as toHarborTrajectory, l as pushDataset, lt as readRolloutLedger, m as ReleaseFormat, mt as HarborAgent, n as BuildSummary, nt as observedScore, o as buildHfDataset, ot as trainingScore, p as RELEASE_FORMATS, pt as HARBOR_IMPORT_GAP, q as readClaudeTranscript, qt as VerifiersTokenUsage, r as ROLLOUT_RELEASE_USAGE, rt as observedSplitScore, s as parseRolloutReleaseArgs, st as appendRolloutLines, t as BuildOptions, tn as toVerifiersRolloutOutputs, tt as isRealnessGated, u as runRolloutReleaseCli, ut as writeRolloutLedger, v as ScrubRule, vt as HarborMetrics, w as scrubText, wt as HarborToolCall, x as emptyScrubCounts, xt as HarborStep, y as addScrubCounts, yt as HarborObservation, z as openOpencodeDb, zt as gateErrors } from "../index-
|
|
2
|
+
import { $ as ScoreOrigin, $t as toSftRows, A as ReleaseRowRef, At as GATE_CHECKS, B as readOpencodeSessionMessages, Bt as gatedEvidenceOf, C as scrubRolloutLine, Ct as HarborSubagentTrajectoryRef, D as FormatGateCounts, Dt as relabelImportedSplit, E as FORMAT_GATE_DISPOSITION, Et as fromHarborTrajectory, F as DEFAULT_OPENCODE_DB, Ft as GateCheckId, G as claudeProjectSlug, Gt as SftRow, H as ClaudeTranscriptRef, Ht as RewardRow, I as OpencodeSessionRow, It as GateCheckedOutcome, J as MintRolloutOptions, Jt as realnessLabels, K as findClaudeTranscripts, Kt as VerifiersRolloutOutput, L as findOpencodeSessionById, Lt as GateEntryPoint, M as gatedRolloutIds, Mt as GATE_POLICIES, N as measureFormatGate, Nt as GateCheck, O as GateDisposition, Ot as toHarborTrajectories, P as releaseRowRefs, Pt as GateCheckDisposition, Q as unmintableReasons, Qt as toRftItems, R as findOpencodeSessionsByDirectory, Rt as GatePolicy, S as scrubLines, St as HarborStepSource, T as EmittedEvidence, Tt as HarborTrajectory, U as ClaudeUsageTotals, Ut as RftItem, V as ClaudeTranscript, Vt as RealnessLabels, W as DEFAULT_CLAUDE_PROJECTS_DIR, Wt as SftExportOptions, X as RolloutScrubber, Xt as toRewardRows, Y as MintRolloutResult, Yt as toJsonl, Z as mintRolloutRows, Zt as toRftItem, _ as ScrubCounts, _t as HarborImageSource, a as ScrubReport, at as trainingReward, b as defaultRolloutScrubber, bt as HarborObservationResult, c as planPushCommand, ct as readRolloutJournal, d as DatasetCardInputs, dt as ATIF_SCHEMA_VERSION, en as toVerifiersRolloutOutput, et as ScorePreference, f as FORMAT_FILES, ft as FromHarborOptions, g as SCRUB_RULES, gt as HarborFinalMetrics, h as buildDatasetCard, ht as HarborContentPart, i as RolloutReleaseCliArgs, it as scoreOrigin, j as assertGateReport, jt as GATE_CHECK_IDS, k as GateReport, kt as toHarborTrajectory, l as pushDataset, lt as readRolloutLedger, m as ReleaseFormat, mt as HarborAgent, n as BuildSummary, nt as observedScore, o as buildHfDataset, ot as trainingScore, p as RELEASE_FORMATS, pt as HARBOR_IMPORT_GAP, q as readClaudeTranscript, qt as VerifiersTokenUsage, r as ROLLOUT_RELEASE_USAGE, rt as observedSplitScore, s as parseRolloutReleaseArgs, st as appendRolloutLines, t as BuildOptions, tn as toVerifiersRolloutOutputs, tt as isRealnessGated, u as runRolloutReleaseCli, ut as writeRolloutLedger, v as ScrubRule, vt as HarborMetrics, w as scrubText, wt as HarborToolCall, x as emptyScrubCounts, xt as HarborStep, y as addScrubCounts, yt as HarborObservation, z as openOpencodeDb, zt as gateErrors } from "../index-B6-B0zTB.js";
|
|
3
3
|
export { ATIF_SCHEMA_VERSION, type BuildOptions, type BuildSummary, CHAT_ROLES, type ChatMessage, type ChatRole, type ChatToolCall, type ClaudeTranscript, type ClaudeTranscriptRef, type ClaudeUsageTotals, DEFAULT_CLAUDE_PROJECTS_DIR, DEFAULT_OPENCODE_DB, type DatasetCardInputs, type EmittedEvidence, FORMAT_FILES, FORMAT_GATE_DISPOSITION, type FormatGateCounts, type FromHarborOptions, GATE_CHECKS, GATE_CHECK_IDS, GATE_POLICIES, type GateCheck, type GateCheckDisposition, type GateCheckId, type GateCheckedOutcome, type GateDisposition, type GateEntryPoint, type GatePolicy, type GateReport, type GatedEvidence, HARBOR_IMPORT_GAP, type HarborAgent, type HarborContentPart, type HarborFinalMetrics, type HarborImageSource, type HarborMetrics, type HarborObservation, type HarborObservationResult, type HarborStep, type HarborStepSource, type HarborSubagentTrajectoryRef, type HarborToolCall, type HarborTrajectory, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type MintedRolloutOutcome, type OpencodeSessionRow, RELEASE_FORMATS, ROLLOUT_CAPTURES, ROLLOUT_RELEASE_USAGE, ROLLOUT_ROLES, ROLLOUT_SCHEMA, ROLLOUT_SPLITS, type RealnessLabels, type ReleaseFormat, type ReleaseRowRef, type RewardRow, type RftItem, type RolloutArtifacts, type RolloutCapture, type RolloutCostBlock, type RolloutLine, type RolloutOutcome, type RolloutPolicy, type RolloutProvenance, type RolloutReleaseCliArgs, type RolloutRole, type RolloutScrubber, type RolloutSplit, type RolloutStep, type RolloutTask, SCRUB_RULES, type ScoreOrigin, type ScorePreference, type ScrubCounts, type ScrubReport, type ScrubRule, type SftExportOptions, type SftRow, TRAINABLE_SPLITS, type ToolDef, type VerifiersRolloutOutput, type VerifiersTokenUsage, addScrubCounts, appendRolloutLines, assertGateReport, assertMinted, assertMintedLines, assertRolloutLine, buildDatasetCard, buildHfDataset, claudeProjectSlug, defaultRolloutScrubber, emptyScrubCounts, findClaudeTranscripts, findOpencodeSessionById, findOpencodeSessionsByDirectory, fromHarborTrajectory, gateErrors, gateGamedOutcome, gatedEvidenceOf, gatedRolloutIds, isRealnessGated, isRolloutLine, isTrainableSplit, measureFormatGate, mintRolloutRows, observedScore, observedSplitScore, openOpencodeDb, parseRolloutReleaseArgs, planPushCommand, pushDataset, readClaudeTranscript, readOpencodeSessionMessages, readRolloutJournal, readRolloutLedger, realnessLabels, relabelImportedSplit, releaseRowRefs, runRolloutReleaseCli, scoreOrigin, scrubLines, scrubRolloutLine, scrubText, toHarborTrajectories, toHarborTrajectory, toJsonl, toRewardRows, toRftItem, toRftItems, toSftRows, toVerifiersRolloutOutput, toVerifiersRolloutOutputs, trainingReward, trainingScore, unmintableReasons, validateRolloutLine, writeRolloutLedger };
|
package/dist/{rubric-predictive-validity-C4r4Y-q8.d.ts → rubric-predictive-validity-9qAwzkZm.d.ts}
RENAMED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { a as RunRecord } from "./run-record-
|
|
1
|
+
import { a as RunRecord } from "./run-record-DdSa93_W.js";
|
|
2
2
|
import { o as OutcomeStore } from "./outcome-store-BYHIuO0e.js";
|
|
3
3
|
//#region src/meta-eval/rubric-predictive-validity.d.ts
|
|
4
4
|
interface RubricPredictiveValidityInput {
|
|
@@ -71,4 +71,4 @@ interface RubricPredictiveValidityReport {
|
|
|
71
71
|
declare function rubricPredictiveValidity(input: RubricPredictiveValidityInput): Promise<RubricPredictiveValidityReport>;
|
|
72
72
|
//#endregion
|
|
73
73
|
export { rubricPredictiveValidity as a, RubricRanking as i, RubricPredictiveValidityInput as n, RubricPredictiveValidityReport as r, RubricOutcomePair as t };
|
|
74
|
-
//# sourceMappingURL=rubric-predictive-validity-
|
|
74
|
+
//# sourceMappingURL=rubric-predictive-validity-9qAwzkZm.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"rubric-predictive-validity-
|
|
1
|
+
{"version":3,"file":"rubric-predictive-validity-9qAwzkZm.d.ts","names":[],"sources":["../src/meta-eval/rubric-predictive-validity.ts"],"mappings":";;;UAoCiB;;;;;EAKf,MAAM;EACN,UAAU;;;;;;EAMV;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;;;;EAKA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;IAAQ;IAAa;;;;;;;;EAOrB;;UAGe;EACf;;EAEA;EACA;EACA;EACA;EACA,SAAS;;UAGM;EACf,OAAO;;EAEP,QAAQ;EACR;EACA;;EAEA;;iBAGoB,yBACpB,OAAO,gCACN,QAAQ"}
|
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { o as FailureClass } from "./schema-BtVldJ3T.js";
|
|
2
|
-
import { a as RunRecord, c as RunTaskFailure, s as RunSplitTag, u as RunTokenUsage } from "./run-record-
|
|
2
|
+
import { a as RunRecord, c as RunTaskFailure, s as RunSplitTag, u as RunTokenUsage } from "./run-record-DdSa93_W.js";
|
|
3
3
|
import { s as TraceStore } from "./store-CT9YIIve.js";
|
|
4
4
|
import { i as TraceEmitter } from "./emitter-DGQGoLyj.js";
|
|
5
|
-
import { $ as ControlRunResult, Q as ControlEvalResult, a as FeedbackLabel, et as ControlRuntimeConfig, x as ProposedSideEffect } from "./feedback-trajectory-
|
|
5
|
+
import { $ as ControlRunResult, Q as ControlEvalResult, a as FeedbackLabel, et as ControlRuntimeConfig, x as ProposedSideEffect } from "./feedback-trajectory-GgoS0-MK.js";
|
|
6
6
|
//#region src/action-policy.d.ts
|
|
7
7
|
interface ActionExecutionPolicy {
|
|
8
8
|
allowedTypes?: string[];
|
|
@@ -222,4 +222,4 @@ declare function controlRunToRunRecord<TState, TAction, TActionResult, TEval ext
|
|
|
222
222
|
declare function scoreFromEvals(evals: readonly ControlEvalResult[]): number | undefined;
|
|
223
223
|
//#endregion
|
|
224
224
|
export { ActionExecutionPolicy as A, ReviewMemoryStore as C, inMemoryReviewStore as D, createLlmReviewer as E, evaluateActionPolicy as M, jsonlReviewStore as O, ReviewMemoryEntry as S, VerifyFn as T, ProposeReviewReport as _, ProposeReviewControlAction as a, ReviewFn as b, ProposeReviewControlState as c, LlmJsonCall as d, LlmReviewerConfig as f, ProposeReviewConfig as g, ProposeOutput as h, scoreFromEvals as i, ActionPolicyDecision as j, runProposeReview as k, controlFailureClassFromVerification as l, ProposeInput as m, RunEvidenceMetadata as n, ProposeReviewControlConfig as o, ProposeFn as p, controlRunToRunRecord as r, ProposeReviewControlResult as s, ControlRunToRunRecordOptions as t, runProposeReviewAsControlLoop as u, ProposeReviewShot as v, Verification as w, ReviewInput as x, Review as y };
|
|
225
|
-
//# sourceMappingURL=run-evidence-
|
|
225
|
+
//# sourceMappingURL=run-evidence-C6G41MSI.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"run-evidence-
|
|
1
|
+
{"version":3,"file":"run-evidence-C6G41MSI.d.ts","names":[],"sources":["../src/action-policy.ts","../src/propose-review.ts","../src/propose-review-control.ts","../src/run-evidence.ts"],"mappings":";;;;;;UAEiB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA,QAAQ;;iBAGM,qBACd,QAAQ,oBACR,SAAQ,uBACR;EAAW;IACV;;;UCmBc;EACf;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe,0BAA0B;EACzC;EACA;EACA;IACE;IACA;IACA;;;UAIa,aAAa;EAC5B;EACA;EACA,OAAO;EACP,aAAa;EACb,aAAa;EACb,UAAU;;UAGK,cAAc,OAAO;EACpC,OAAO;EACP,eAAe;;UAGA,YAAY,OAAO;EAClC;EACA;EACA,OAAO;EACP,cAAc;EACd,cAAc;EACd,QAAQ;;KAGE,UAAU,OAAO,sBAC3B,OAAO,aAAa,WACjB,QAAQ,cAAc,OAAO;KAEtB,SAAS,UAAU,OAAO,UAAU,QAAQ;KAE5C,SAAS,OAAO,sBAC1B,OAAO,YAAY,OAAO,aACvB,QAAQ;UAEI;EACf,QAAQ,QAAQ;EAChB,OAAO,OAAO,oBAAoB;;UAGnB,oBAAoB,OAAO;EAC1C;EACA,cAAc;EACd,SAAS,UAAU,OAAO;EAC1B,QAAQ,SAAS;EACjB,QAAQ,SAAS,OAAO;;EAExB;;EAEA;;;;;;EAMA;EACA;;EAEA,SAAS;;EAET,QAAQ;EACR;EACA;EACA;;;;;EAKA;;UAGe,kBAAkB,OAAO;EACxC;EACA,OAAO;EACP,cAAc;EACd,cAAc;EACd,QAAQ;EACR;EACA;EACA;;UAGe,oBAAoB,OAAO;EAC1C;EACA;EACA,OAAO,kBAAkB,OAAO;EAChC,YAAY;EACZ,mBAAmB;EACnB,eAAe;EACf;EACA;;iBAKc,oBAAoB,UAAS,sBAA2B;iBAYxD,iBAAiB,eAAe;iBA8B1B,iBAAiB,OAAO,mBAC5C,QAAQ,oBAAoB,OAAO,WAClC,QAAQ,oBAAoB,OAAO;KAiN1B,eAAe;EAAO;EAAgB;MAAmB;UAEpD,kBAAkB,OAAO;EACxC,UAAU;EACV,eAAe,OAAO;EACtB,sBAAsB,SAAS;;EAE/B;;iBAQc,kBAAkB,OAAO,mBACvC,KAAK,kBAAkB,OAAO,WAC7B,SAAS,OAAO;;;UC9ZF,0BAA0B,OAAO;EAChD;EACA,OAAO;EACP,aAAa;EACb,cAAc;EACd,eAAe;EACf,QAAQ;EACR;EACA;EACA;;UAGe;EACf;EACA;;UAGe,2BAA2B,OAAO;EACjD,OAAO;EACP,cAAc;EACd,eAAe;EACf,QAAQ;EACR;EACA;;UAGe,2BAA2B,OAAO;EACjD;EACA,cAAc;EACd,SAAS,UAAU,OAAO;EAC1B,QAAQ,SAAS;EACjB,QAAQ,SAAS,OAAO;EACxB;EACA;EACA,SAAS;EACT,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA,gCAAgC,cAAc,iBAAiB;EAC/D,gBAAgB,qBACd,0BAA0B,OAAO,UACjC,4BACA,2BAA2B,OAAO;;iBAOhB,8BAA8B,OAAO,mBACzD,QAAQ,2BAA2B,OAAO,WACzC,QACD,iBACE,0BAA0B,OAAO,UACjC,4BACA,2BAA2B,OAAO;iBAwKtB,oCACd,cAAc,eACb;;;UChPc;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,UAAU;EACV,YAAY;EACZ,gBAAgB;EAChB;EACA,gBAAgB;EAChB,MAAM;;KAGI,+BAA+B,sBACzC;EACE;EACA;;;;;;;;;;iBAWY,sBACd,QACA,SACA,eACA,cAAc,oBAAoB,mBAElC,KAAK,iBAAiB,QAAQ,SAAS,eAAe,QACtD,SAAS,+BACR;iBAwDa,eAAe,gBAAgB"}
|
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { l as ValidationError } from "./errors-
|
|
2
|
-
import { r as AgentProfileCell } from "./agent-profile-cell-
|
|
3
|
-
import { p as CostProvenance } from "./cost-ledger-
|
|
1
|
+
import { l as ValidationError } from "./errors-CKPfb2aH.js";
|
|
2
|
+
import { r as AgentProfileCell } from "./agent-profile-cell-BOP-iA9Q.js";
|
|
3
|
+
import { p as CostProvenance } from "./cost-ledger-Bv_e8XHY.js";
|
|
4
4
|
import { o as FailureClass } from "./schema-BtVldJ3T.js";
|
|
5
5
|
//#region src/run-record.d.ts
|
|
6
6
|
/** Search/dev/holdout split tag. 'search' is the paper-grade alias for the
|
|
@@ -236,4 +236,4 @@ declare function roundTripRunRecord(record: RunRecord): RunRecord;
|
|
|
236
236
|
declare function modelHasSnapshot(model: string): boolean;
|
|
237
237
|
//#endregion
|
|
238
238
|
export { RunRecord as a, RunTaskFailure as c, isRunRecord as d, modelHasSnapshot as f, validateRunRecord as g, runTaskScore as h, RunOutcome as i, RunTerminalOutcome as l, roundTripRunRecord as m, RunCostProvenance as n, RunRecordValidationError as o, parseRunRecordSafe as p, RunJudgeMetadata as r, RunSplitTag as s, JudgeScoresRecord as t, RunTokenUsage as u };
|
|
239
|
-
//# sourceMappingURL=run-record-
|
|
239
|
+
//# sourceMappingURL=run-record-DdSa93_W.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"run-record-
|
|
1
|
+
{"version":3,"file":"run-record-DdSa93_W.d.ts","names":[],"sources":["../src/run-record.ts"],"mappings":";;;;;;;KAsCY;;;;;;;KAQA;UAEK;EACf;;EAEA;;EAEA;;EAEA;;EAEA;;;KAIU,oBAAoB;UAEf;EACf;EACA;;;EAGA;;;EAGA;;;;;;;;;;;;;;;;;;;;;UAsBe;;EAEf,UAAU,eAAe;;EAEzB,YAAY;;;EAGZ;;;;EAIA;;;EAGA;;UAGe;;;EAGf;;;EAGA;;;;EAIA,KAAK;;;;;;EAML,cAAc;;;;;;;EAOd;IAAa;IAAe;IAAgB;;;;;;;;;;;;;;;;;;;UAmB7B;;EAEf;;;EAGA;;;EAGA;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,gBAAgB;;EAEhB,YAAY;;EAEZ,iBAAiB;;;EAGjB;;EAEA,gBAAgB;;EAEhB,SAAS;;;;;EAKT,eAAe;;;EAGf;;EAEA,UAAU;;;;;EAKV;;;;;;;;EAQA,eAAe;;;;;;;;;KAUL;EACN;EAA0B;;EAC1B;EAAyB;;EAEzB,cAAc,QAAQ;EACtB;;;;;;;;;;iBAWU,aAAa,QAAQ;cAmCxB,iCAAiC;WACnC;EACT,YAAY,iBAAiB;;;;;;;iBAWf,kBAAkB,iBAAiB;;iBAiOnC,YAAY,iBAAiB,SAAS;;iBAUtC,mBACd;EACG;EAAU,OAAO;;EAAgB;EAAW,OAAO;;;iBAUxC,mBAAmB,QAAQ,YAAY;;;;;;;iBAuFvC,iBAAiB"}
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import { o as NotFoundError } from "./errors-D-LKuDhb.js";
|
|
2
2
|
import { i as CostLedger } from "./cost-ledger-DMFxsLKr.js";
|
|
3
|
-
import { c as callLlmJson, f as maximumChargeForLlmRequest, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-
|
|
3
|
+
import { c as callLlmJson, f as maximumChargeForLlmRequest, l as costReceiptFromLlm, u as costReceiptFromLlmError } from "./llm-client-Dv5BiKLE.js";
|
|
4
4
|
import { a as computeFindingId } from "./usage-receipt-EVI8B8Xu.js";
|
|
5
5
|
import { a as clamp01, i as aggregateRunScore } from "./proposal-findings-2GIUo1et.js";
|
|
6
6
|
import { f as Mutex } from "./ledger-core-DXZIqu17.js";
|
|
@@ -764,4 +764,4 @@ function createSemanticConceptJudge(options = {}) {
|
|
|
764
764
|
//#endregion
|
|
765
765
|
export { RunCritic as a, buildSkillUsageReport as c, defaultIsMaterial as d, diffFindings as f, defineTraceAnalyst as h, runSemanticConceptJudge as i, emitSkillUsageFindings as l, defineCustomAnalyst as m, SEMANTIC_CONCEPT_JUDGE_VERSION as n, SKILL_USAGE_ANALYST as o, LockedJsonlAppender as p, createSemanticConceptJudge as r, SkillUsageAnalyst as s, DEFAULT_COMPLEXITY_WEIGHTS as t, FindingsStore as u };
|
|
766
766
|
|
|
767
|
-
//# sourceMappingURL=semantic-concept-judge-
|
|
767
|
+
//# sourceMappingURL=semantic-concept-judge-DwF6n05O.js.map
|