@tangle-network/agent-eval 0.145.6 → 0.145.7
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 +10 -0
- package/README.md +4 -0
- package/dist/analyst/index.d.ts +10 -10
- package/dist/analyst/index.js +1 -1
- package/dist/{backend-integrity-CsVin_Wb.d.ts → backend-integrity-BffcHGdm.d.ts} +2 -2
- package/dist/{backend-integrity-CsVin_Wb.d.ts.map → backend-integrity-BffcHGdm.d.ts.map} +1 -1
- package/dist/{benchmark-Ceoan7vk.d.ts → benchmark-DbiZbPdH.d.ts} +3 -3
- package/dist/{benchmark-Ceoan7vk.d.ts.map → benchmark-DbiZbPdH.d.ts.map} +1 -1
- package/dist/{benchmark-command-DspwA7cv.js → benchmark-command-C7L0rQsD.js} +2 -2
- package/dist/{benchmark-command-DspwA7cv.js.map → benchmark-command-C7L0rQsD.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +5 -5
- package/dist/benchmarks/index.js +2 -2
- package/dist/campaign/index.d.ts +7 -7
- package/dist/campaign/index.js +3 -3
- package/dist/{campaign-jTOvqnse.js → campaign-DRcGtxEf.js} +35 -11
- package/dist/campaign-DRcGtxEf.js.map +1 -0
- package/dist/{capture-fetch-DDvpjVRU.d.ts → capture-fetch-hykwHfDI.d.ts} +2 -2
- package/dist/{capture-fetch-DDvpjVRU.d.ts.map → capture-fetch-hykwHfDI.d.ts.map} +1 -1
- package/dist/cli.js +1 -1
- package/dist/{client-DCVe0CwG.d.ts → client-D_TIV9pJ.d.ts} +4 -4
- package/dist/{client-DCVe0CwG.d.ts.map → client-D_TIV9pJ.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +11 -11
- package/dist/contract/index.js +4 -4
- package/dist/{default-registry-CgzxnJsj.d.ts → default-registry-cpHtP2Og.d.ts} +6 -6
- package/dist/{default-registry-CgzxnJsj.d.ts.map → default-registry-cpHtP2Og.d.ts.map} +1 -1
- package/dist/{define-agent-eval-DLhZ8IHi.d.ts → define-agent-eval-DBOpAM_g.d.ts} +6 -6
- package/dist/{define-agent-eval-DLhZ8IHi.d.ts.map → define-agent-eval-DBOpAM_g.d.ts.map} +1 -1
- package/dist/{define-agent-eval-CViDh2P9.js → define-agent-eval-m5lM7XGE.js} +4 -4
- package/dist/{define-agent-eval-CViDh2P9.js.map → define-agent-eval-m5lM7XGE.js.map} +1 -1
- package/dist/{engine-FRZ3RLvT.d.ts → engine-DJqRKbhs.d.ts} +7 -7
- package/dist/{engine-FRZ3RLvT.d.ts.map → engine-DJqRKbhs.d.ts.map} +1 -1
- package/dist/{eval-campaign-UB-usSQ2.js → eval-campaign-aNqpefCS.js} +2 -2
- package/dist/{eval-campaign-UB-usSQ2.js.map → eval-campaign-aNqpefCS.js.map} +1 -1
- package/dist/{exact-types-BH1twmAJ.d.ts → exact-types-CzbhVDr2.d.ts} +2 -2
- package/dist/{exact-types-BH1twmAJ.d.ts.map → exact-types-CzbhVDr2.d.ts.map} +1 -1
- package/dist/experiment/index.d.ts +3 -3
- package/dist/{feedback-trajectory-juOozjAc.d.ts → feedback-trajectory-D9vYSob_.d.ts} +3 -3
- package/dist/{feedback-trajectory-juOozjAc.d.ts.map → feedback-trajectory-D9vYSob_.d.ts.map} +1 -1
- package/dist/hosted/index.d.ts +3 -3
- package/dist/index-BKShPTwZ.d.ts +1 -0
- package/dist/{index-COQYtuRF.d.ts → index-BNPtkBPf.d.ts} +2 -2
- package/dist/{index-COQYtuRF.d.ts.map → index-BNPtkBPf.d.ts.map} +1 -1
- package/dist/{index-DF3ynqmJ.d.ts → index-D7gl974A.d.ts} +11 -11
- package/dist/{index-DF3ynqmJ.d.ts.map → index-D7gl974A.d.ts.map} +1 -1
- package/dist/{index-PgqfwhsM.d.ts → index-YrUFx3FU.d.ts} +5 -5
- package/dist/{index-PgqfwhsM.d.ts.map → index-YrUFx3FU.d.ts.map} +1 -1
- package/dist/index.d.ts +22 -22
- package/dist/index.js +9 -9
- package/dist/{insight-report-CRi-Ufrj.d.ts → insight-report-BeT8KCgI.d.ts} +3 -3
- package/dist/{insight-report-CRi-Ufrj.d.ts.map → insight-report-BeT8KCgI.d.ts.map} +1 -1
- package/dist/{llm-judge-DuYa4SEA.js → llm-judge-BO1LGjdn.js} +2 -2
- package/dist/{llm-judge-DuYa4SEA.js.map → llm-judge-BO1LGjdn.js.map} +1 -1
- package/dist/meta-eval/index.d.ts +1 -1
- package/dist/{mint-Dj9Ww_3I.js → mint-BV6tLVWl.js} +2 -2
- package/dist/{mint-Dj9Ww_3I.js.map → mint-BV6tLVWl.js.map} +1 -1
- package/dist/multishot/index.d.ts +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{pre-registration-DeRvl9sE.d.ts → pre-registration-zFSLEiFU.d.ts} +2 -2
- package/dist/{pre-registration-DeRvl9sE.d.ts.map → pre-registration-zFSLEiFU.d.ts.map} +1 -1
- package/dist/{produced-state-D6j7qy1Q.js → produced-state-C6vSgOpE.js} +2 -2
- package/dist/{produced-state-D6j7qy1Q.js.map → produced-state-C6vSgOpE.js.map} +1 -1
- package/dist/{promotion-policy-CsMZJOB-.d.ts → promotion-policy-u3wj6w3U.d.ts} +2 -2
- package/dist/{promotion-policy-CsMZJOB-.d.ts.map → promotion-policy-u3wj6w3U.d.ts.map} +1 -1
- package/dist/{provenance-8L-_4xiL.d.ts → provenance-BOtMtoiZ.d.ts} +5 -5
- package/dist/{provenance-8L-_4xiL.d.ts.map → provenance-BOtMtoiZ.d.ts.map} +1 -1
- package/dist/{registry-oJeeI4-a.d.ts → registry-B_1Frl8a.d.ts} +3 -3
- package/dist/{registry-oJeeI4-a.d.ts.map → registry-B_1Frl8a.d.ts.map} +1 -1
- package/dist/{release-confidence-BFRE5WSp.d.ts → release-confidence-4XrqlpFD.d.ts} +3 -3
- package/dist/{release-confidence-BFRE5WSp.d.ts.map → release-confidence-4XrqlpFD.d.ts.map} +1 -1
- package/dist/{release-confidence-CxDuiAev.js → release-confidence-BknrpBnO.js} +2 -2
- package/dist/{release-confidence-CxDuiAev.js.map → release-confidence-BknrpBnO.js.map} +1 -1
- package/dist/reporting.d.ts +3 -3
- package/dist/reporting.js +1 -1
- package/dist/{researcher-DaN4GST-.d.ts → researcher-Du-oniHp.d.ts} +3 -3
- package/dist/{researcher-DaN4GST-.d.ts.map → researcher-Du-oniHp.d.ts.map} +1 -1
- package/dist/{reward-hacking-DSSTuI9r.d.ts → reward-hacking-Bu8ev6PR.d.ts} +2 -2
- package/dist/{reward-hacking-DSSTuI9r.d.ts.map → reward-hacking-Bu8ev6PR.d.ts.map} +1 -1
- package/dist/{reward-hacking-DNgjilrV.js → reward-hacking-DFo2FU5J.js} +2 -2
- package/dist/{reward-hacking-DNgjilrV.js.map → reward-hacking-DFo2FU5J.js.map} +1 -1
- package/dist/rl.d.ts +5 -5
- package/dist/rl.js +4 -4
- package/dist/rollout/index.d.ts +1 -1
- package/dist/rollout/index.js +2 -2
- package/dist/{rollout-BWtw0I_6.js → rollout-ytVQ7WT8.js} +2 -2
- package/dist/{rollout-BWtw0I_6.js.map → rollout-ytVQ7WT8.js.map} +1 -1
- package/dist/{rubric-predictive-validity-BgxtKe4G.d.ts → rubric-predictive-validity-C7LnNvF2.d.ts} +2 -2
- package/dist/{rubric-predictive-validity-BgxtKe4G.d.ts.map → rubric-predictive-validity-C7LnNvF2.d.ts.map} +1 -1
- package/dist/{run-record-BvHPVS-i.js → run-record-D2lDdSAz.js} +5 -3
- package/dist/run-record-D2lDdSAz.js.map +1 -0
- package/dist/{run-record-CKiihE6f.d.ts → run-record-DVV82Gwh.d.ts} +8 -3
- package/dist/run-record-DVV82Gwh.d.ts.map +1 -0
- package/dist/{skillopt-optimization-method-CFDSzD-7.d.ts → skillopt-optimization-method-BPQwXkkY.d.ts} +5 -5
- package/dist/{skillopt-optimization-method-CFDSzD-7.d.ts.map → skillopt-optimization-method-BPQwXkkY.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-BTls2l-Z.js → skillopt-optimization-method-Ceic2bfU.js} +2 -2
- package/dist/{skillopt-optimization-method-BTls2l-Z.js.map → skillopt-optimization-method-Ceic2bfU.js.map} +1 -1
- package/dist/{statistical-heldout-C4De2tRI.d.ts → statistical-heldout-CBGPSZ5X.d.ts} +3 -3
- package/dist/{statistical-heldout-C4De2tRI.d.ts.map → statistical-heldout-CBGPSZ5X.d.ts.map} +1 -1
- package/dist/{store-tool-spans-CggeC1LB.d.ts → store-tool-spans-C9c4R7ca.d.ts} +4 -4
- package/dist/{store-tool-spans-CggeC1LB.d.ts.map → store-tool-spans-C9c4R7ca.d.ts.map} +1 -1
- package/dist/{summary-report-CaL-Hnxt.d.ts → summary-report-B__Y5ub3.d.ts} +2 -2
- package/dist/{summary-report-CaL-Hnxt.d.ts.map → summary-report-B__Y5ub3.d.ts.map} +1 -1
- package/dist/{tool-groups-BdcoEgtV.d.ts → tool-groups-CZotz-e_.d.ts} +3 -3
- package/dist/tool-groups-CZotz-e_.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +2 -2
- package/dist/traces.d.ts +6 -6
- package/dist/traces.js +1 -1
- package/dist/{types-BxLccGMf.d.ts → types-BjsNDR49.d.ts} +3 -3
- package/dist/{types-BxLccGMf.d.ts.map → types-BjsNDR49.d.ts.map} +1 -1
- package/dist/{types-BRsxjg7z.d.ts → types-C4bSVIr7.d.ts} +3 -3
- package/dist/{types-BRsxjg7z.d.ts.map → types-C4bSVIr7.d.ts.map} +1 -1
- package/dist/{types-DLQx4mKU.d.ts → types-vXyshMwx.d.ts} +2 -2
- package/dist/{types-DLQx4mKU.d.ts.map → types-vXyshMwx.d.ts.map} +1 -1
- package/dist/wire/index.d.ts +1 -1
- package/docs/eval-surface-map.md +2 -1
- package/docs/wire-protocol.md +4 -0
- package/package.json +1 -1
- package/dist/campaign-jTOvqnse.js.map +0 -1
- package/dist/index-Ba3YrbAL.d.ts +0 -1
- package/dist/run-record-BvHPVS-i.js.map +0 -1
- package/dist/run-record-CKiihE6f.d.ts.map +0 -1
- package/dist/tool-groups-BdcoEgtV.d.ts.map +0 -1
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { u as RunTokenUsage } from "./run-record-
|
|
1
|
+
import { u as RunTokenUsage } from "./run-record-DVV82Gwh.js";
|
|
2
2
|
import { it as RawProviderSink, tt as ProviderRedactor } from "./types-Cx3YUh2r.js";
|
|
3
3
|
//#region src/trace/extract-usage.d.ts
|
|
4
4
|
type ExtractedUsage = RunTokenUsage;
|
|
@@ -65,4 +65,4 @@ interface CaptureFetchOptions {
|
|
|
65
65
|
declare function captureFetchToRawSink(fetch: typeof globalThis.fetch, sink: RawProviderSink, ctx: CaptureFetchContext, opts?: CaptureFetchOptions): typeof globalThis.fetch;
|
|
66
66
|
//#endregion
|
|
67
67
|
export { ExtractedUsage as a, extractUsageFromResponse as c, ExtractUsageFromSseOptions as i, extractUsageFromSse as l, CaptureFetchOptions as n, SseUsageMode as o, captureFetchToRawSink as r, extractUsage as s, CaptureFetchContext as t };
|
|
68
|
-
//# sourceMappingURL=capture-fetch-
|
|
68
|
+
//# sourceMappingURL=capture-fetch-hykwHfDI.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"capture-fetch-
|
|
1
|
+
{"version":3,"file":"capture-fetch-hykwHfDI.d.ts","names":[],"sources":["../src/trace/extract-usage.ts","../src/trace/capture-fetch.ts"],"mappings":";;;KAgBY,iBAAiB;KACjB;UAEK;;EAEf,OAAO;;;;;;;iBAuDO,aAAa,gBAAgB;;;;;;iBA6C7B,oBACd,cACA,UAAS,6BACR;;;;;;;iBA4DmB,yBACpB,UAAU,UACV,aAAa,6BACZ,QAAQ;;;UCrKM;;;EAGf;;EAEA;;;EAGA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;;EAGX;;;EAGA;;;;;;;;EAQA,WAAW,OAAO,gBAAgB,KAAK;;EAEvC,eAAe;;iBAkDD,sBACd,cAAc,WAAW,OACzB,MAAM,iBACN,KAAK,qBACL,OAAM,6BACE,WAAW"}
|
package/dist/cli.js
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
2
|
import { o as runRolloutReleaseCli } from "./hf-dataset-XggBupCr.js";
|
|
3
|
-
import { n as runAnalystBenchmarkCommand } from "./benchmark-command-
|
|
3
|
+
import { n as runAnalystBenchmarkCommand } from "./benchmark-command-C7L0rQsD.js";
|
|
4
4
|
import { a as runRpcBatch, o as runRpcOnce, p as handleVersion, r as startServerAsync, s as buildOpenApi } from "./server-ulsOdrTI.js";
|
|
5
5
|
import { writeFileSync } from "node:fs";
|
|
6
6
|
//#region src/cli-config.ts
|
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { l as RunTerminalOutcome } from "./run-record-
|
|
2
|
-
import { _ as GateDecision, j as MutableSurface } from "./types-
|
|
3
|
-
import { o as InsightReport } from "./insight-report-
|
|
1
|
+
import { l as RunTerminalOutcome } from "./run-record-DVV82Gwh.js";
|
|
2
|
+
import { _ as GateDecision, j as MutableSurface } from "./types-BjsNDR49.js";
|
|
3
|
+
import { o as InsightReport } from "./insight-report-BeT8KCgI.js";
|
|
4
4
|
//#region src/hosted/types.d.ts
|
|
5
5
|
declare const HOSTED_WIRE_VERSION: '2026-07-24.v1';
|
|
6
6
|
type HostedWireVersion = typeof HOSTED_WIRE_VERSION;
|
|
@@ -199,4 +199,4 @@ declare function hostedClientFromEnv(overrides?: Partial<HostedTenant> & {
|
|
|
199
199
|
}): HostedClient | undefined;
|
|
200
200
|
//#endregion
|
|
201
201
|
export { UnixNanoTimestamp as _, hostedTenantFromEnv as a, EvalRunGenerationSnapshot as c, HostedIngestHeaders as d, HostedWireVersion as f, TraceSpanEvent as g, IngestTracesRequest as h, hostedClientFromEnv as i, EvalRunStatus as l, IngestResponse as m, HostedTenant as n, EvalRunCellScore as o, IngestEvalRunsRequest as p, createHostedClient as r, EvalRunEvent as s, HostedClient as t, HOSTED_WIRE_VERSION as u };
|
|
202
|
-
//# sourceMappingURL=client-
|
|
202
|
+
//# sourceMappingURL=client-D_TIV9pJ.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"client-
|
|
1
|
+
{"version":3,"file":"client-D_TIV9pJ.d.ts","names":[],"sources":["../src/hosted/types.ts","../src/hosted/client.ts"],"mappings":";;;;cAgCa;KACD,2BAA2B;;UAKtB;;EAEf;;EAEA;;EAEA,yBAAyB;;EAEzB;;;KAMU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA,YAAY,eAAe;;EAE3B,iBAAiB;;EAEjB;;EAEA;;UAGe;;EAEf;;;EAGA;;;EAGA,UAAU;;EAEV,OAAO;;EAEP;;EAEA;;EAEA;;;;;;;UAQe;;EAEf;;EAEA;;EAEA;;EAEA,QAAQ;;EAER,QAAQ;;EAER,WAAW;;EAEX,aAAa;;EAEb,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;;;;;EAMA,gBAAgB;;;;;;KASN;;;;;;UAOK;EACf;EACA;EACA;EACA;EACA,mBAAmB;EACnB,iBAAiB;EACjB,YAAY;EACZ,SAAS;IACP,cAAc;IACd;IACA,aAAa;;EAEf;IAAW;IAAgC;;;EAE3C;;EAEA;;EAEA;;EAEA;;UAKe;EACf,aAAa;EACb,QAAQ;;UAGO;EACf,aAAa;EACb,OAAO;;UAGQ;;EAEf;;EAEA,UAAU;IAAQ;IAAe;;;;;UC1JlB;;EAEf;;EAEA;;EAEA;;EAEA,mBAAmB;;EAEnB;;EAEA;;UAGe;EACf,cAAc,OAAO,cAAc,0BAA0B,QAAQ;EACrE,eAAe,QAAQ,gBAAgB,0BAA0B,QAAQ;EACzE,aAAa,OAAO,kBAAkB,0BAA0B,QAAQ;WAC/D,QAAQ;WACR,aAAa;;iBAyGR,mBAAmB,QAAQ,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsE1C,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C;iBAiBa,oBACd,YAAW,QAAQ;EAAkB,MAAM;IAC1C"}
|
package/dist/contract/index.d.ts
CHANGED
|
@@ -1,17 +1,17 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "../cost-ledger-DbQdN3nO.js";
|
|
2
|
-
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-
|
|
2
|
+
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "../run-record-DVV82Gwh.js";
|
|
3
3
|
import { p as ChatClient, v as CreateChatClientOpts, w as createChatClient } from "../types-Cx3YUh2r.js";
|
|
4
|
-
import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin, x as makeProposalFinding } from "../types-
|
|
5
|
-
import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-
|
|
6
|
-
import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-
|
|
7
|
-
import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-
|
|
8
|
-
import { C as analyzeRuns, S as SummarizeExecutionOptions, b as AnalyzeRunsOptions, w as summarizeExecution, x as ExecutionReport } from "../engine-
|
|
9
|
-
import { $ as OptimizationTokenUsage, At as llmJudge, Ct as fsCampaignStorage, D as runEval, E as RunEvalOptions, G as OptimizationMethodComparison, H as CompareOptimizationMethodsOptions, J as OptimizationMethodProvenance, K as OptimizationMethodInput, Ot as LlmJudgeDimension, Q as OptimizationPackageSource, U as ComparisonCost, W as OptimizationMethod, Y as OptimizationMethodResult, at as RunCampaignOptions, b as RunImprovementLoopResult, it as CampaignCellFailureReceipt, kt as LlmJudgeOptions, ot as runCampaign, tt as compareOptimizationMethods, wt as inMemoryCampaignStorage, x as runImprovementLoop, xt as CampaignStorage, y as RunImprovementLoopOptions } from "../provenance-
|
|
10
|
-
import { B as runReferenceEquivalenceJudge, C as ExternalTextOptimizerContext, E as ExternalTextEvaluationResponse, F as ReferenceEquivalenceJudgeInput, I as ReferenceEquivalenceJudgeOptions, L as ReferenceEquivalenceJudgeResult, N as REFERENCE_EQUIVALENCE_INPUT_LIMITS, P as REFERENCE_EQUIVALENCE_JUDGE_VERSION, R as ReferenceEquivalenceScenario, S as ExternalTextOptimizationMethodConfig, T as ExternalOptimizationExample, _ as DefaultProductionGateOptions, a as GepaAdaptiveEngineRun, b as composeGate, c as GepaOptimizationMethodConfig, d as gepaOptimizationMethod, f as OpenAICompatibleOptimizerModel, g as DefaultProductionGateCheck, h as heldOutGate, i as skillOptOptimizationMethod, j as campaignSplitDigest, l as GepaOptimizationRecipe, m as HeldOutGateOptions, n as SkillOptRunnerCommand, o as GepaEngineOptions, p as OptimizerModelBudget, r as SkillOptTrainerConfig, s as GepaEngineRun, t as SkillOptOptimizationMethodConfig, u as GepaRunnerCommand, v as DefaultProductionRewardHackingOptions, w as ExternalTextOptimizerResult, x as externalTextOptimizationMethod, y as defaultProductionGate, z as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-
|
|
11
|
-
import { a as ObjectiveSource, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-
|
|
12
|
-
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-
|
|
4
|
+
import { _ as ProposalFinding, i as AnalystFinding, v as ProposalFindingOrigin, x as makeProposalFinding } from "../types-vXyshMwx.js";
|
|
5
|
+
import { n as buildDefaultAnalystRegistry, t as DefaultAnalystRegistryOptions } from "../default-registry-cpHtP2Og.js";
|
|
6
|
+
import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-BjsNDR49.js";
|
|
7
|
+
import { a as FailureClusterInsight, c as JudgeInsight, d as Recommendation, f as ReleaseSummary, i as FailureClassTally, l as LiftInsight, m as TokenUsageInsight, n as ExecutionErrorOutcomeCell, o as InsightReport, p as ScalarDistribution, r as ExecutionInsight, s as InterRaterInsight, t as CostProvenanceSummary, u as OutcomeCorrelationInsight } from "../insight-report-BeT8KCgI.js";
|
|
8
|
+
import { C as analyzeRuns, S as SummarizeExecutionOptions, b as AnalyzeRunsOptions, w as summarizeExecution, x as ExecutionReport } from "../engine-DJqRKbhs.js";
|
|
9
|
+
import { $ as OptimizationTokenUsage, At as llmJudge, Ct as fsCampaignStorage, D as runEval, E as RunEvalOptions, G as OptimizationMethodComparison, H as CompareOptimizationMethodsOptions, J as OptimizationMethodProvenance, K as OptimizationMethodInput, Ot as LlmJudgeDimension, Q as OptimizationPackageSource, U as ComparisonCost, W as OptimizationMethod, Y as OptimizationMethodResult, at as RunCampaignOptions, b as RunImprovementLoopResult, it as CampaignCellFailureReceipt, kt as LlmJudgeOptions, ot as runCampaign, tt as compareOptimizationMethods, wt as inMemoryCampaignStorage, x as runImprovementLoop, xt as CampaignStorage, y as RunImprovementLoopOptions } from "../provenance-BOtMtoiZ.js";
|
|
10
|
+
import { B as runReferenceEquivalenceJudge, C as ExternalTextOptimizerContext, E as ExternalTextEvaluationResponse, F as ReferenceEquivalenceJudgeInput, I as ReferenceEquivalenceJudgeOptions, L as ReferenceEquivalenceJudgeResult, N as REFERENCE_EQUIVALENCE_INPUT_LIMITS, P as REFERENCE_EQUIVALENCE_JUDGE_VERSION, R as ReferenceEquivalenceScenario, S as ExternalTextOptimizationMethodConfig, T as ExternalOptimizationExample, _ as DefaultProductionGateOptions, a as GepaAdaptiveEngineRun, b as composeGate, c as GepaOptimizationMethodConfig, d as gepaOptimizationMethod, f as OpenAICompatibleOptimizerModel, g as DefaultProductionGateCheck, h as heldOutGate, i as skillOptOptimizationMethod, j as campaignSplitDigest, l as GepaOptimizationRecipe, m as HeldOutGateOptions, n as SkillOptRunnerCommand, o as GepaEngineOptions, p as OptimizerModelBudget, r as SkillOptTrainerConfig, s as GepaEngineRun, t as SkillOptOptimizationMethodConfig, u as GepaRunnerCommand, v as DefaultProductionRewardHackingOptions, w as ExternalTextOptimizerResult, x as externalTextOptimizationMethod, y as defaultProductionGate, z as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-BPQwXkkY.js";
|
|
11
|
+
import { a as ObjectiveSource, c as PromotionPolicy, d as paretoSignificanceGate, i as EvidenceVector, l as buildEvidenceVector, n as AxisVerdict, o as ParetoSignificanceGateOptions, r as BuildEvidenceVectorOptions, s as PromotionObjective, t as AxisEvidence, u as paretoPolicy } from "../promotion-policy-u3wj6w3U.js";
|
|
12
|
+
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent } from "../client-D_TIV9pJ.js";
|
|
13
13
|
import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
|
|
14
|
-
import { a as DefinedAgentEval, c as SelfImproveOptions, d as SelfImproveRunError, f as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveProgressEvent, n as AgentEvalEvaluateOptions, o as defineAgentEval, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveResult } from "../define-agent-eval-
|
|
14
|
+
import { a as DefinedAgentEval, c as SelfImproveOptions, d as SelfImproveRunError, f as selfImprove, i as DefineAgentEvalOptions, l as SelfImproveProgressEvent, n as AgentEvalEvaluateOptions, o as defineAgentEval, r as AgentEvalImproveOptions, s as SelfImproveBudget, t as AgentEvalAgent, u as SelfImproveResult } from "../define-agent-eval-DBOpAM_g.js";
|
|
15
15
|
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";
|
|
16
16
|
//#region src/contract/measured-comparison.d.ts
|
|
17
17
|
interface SealCandidateBenchmarkSuiteOptions {
|
package/dist/contract/index.js
CHANGED
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { s as ValidationError } from "../errors-Dngq5h35.js";
|
|
2
2
|
import { r as pairedBootstrap } from "../paired-tests-BHIhYVdu.js";
|
|
3
|
-
import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-
|
|
4
|
-
import {
|
|
5
|
-
import { G as runEval, K as runCampaign, f as runImprovementLoop, nt as campaignSplitDigest, t as llmJudge, x as compareOptimizationMethods, z as defaultProductionGate } from "../llm-judge-
|
|
3
|
+
import { a as summarizeExecution, i as analyzeRuns, n as SelfImproveRunError, r as selfImprove, t as defineAgentEval } from "../define-agent-eval-m5lM7XGE.js";
|
|
4
|
+
import { a as parseRunRecordSafe } from "../run-record-D2lDdSAz.js";
|
|
5
|
+
import { G as runEval, K as runCampaign, f as runImprovementLoop, nt as campaignSplitDigest, t as llmJudge, x as compareOptimizationMethods, z as defaultProductionGate } from "../llm-judge-BO1LGjdn.js";
|
|
6
6
|
import { i as isModelPriced, n as estimateCost } from "../metrics-Cl0L1KUy.js";
|
|
7
7
|
import { i as CostLedger } from "../cost-ledger-BSe92yAV.js";
|
|
8
8
|
import { d as mapConcurrentRange } from "../ledger-core-BmZt19oQ.js";
|
|
@@ -15,7 +15,7 @@ import { LLM_MODEL_ATTR_KEYS, SPAN_KIND_ATTR_KEYS } from "../trace-attributes.js
|
|
|
15
15
|
import { t as extractUsage } from "../extract-usage-BrQ8mCLX.js";
|
|
16
16
|
import { n as InMemoryOutcomeStore, t as FileSystemOutcomeStore } from "../outcome-store-ChBKlTd_.js";
|
|
17
17
|
import { i as summarizeTraceErrors, n as recordAggregateMeasurements, r as summarizeExecutionMeasurements, t as readTaskFailureLabels } from "../task-failure-attributes-CpQ4y5RD.js";
|
|
18
|
-
import { a as externalTextOptimizationMethod, c as REFERENCE_EQUIVALENCE_INPUT_LIMITS, d as runReferenceEquivalenceJudge, i as composeGate, l as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, r as heldOutGate, t as skillOptOptimizationMethod, u as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-
|
|
18
|
+
import { a as externalTextOptimizationMethod, c as REFERENCE_EQUIVALENCE_INPUT_LIMITS, d as runReferenceEquivalenceJudge, i as composeGate, l as REFERENCE_EQUIVALENCE_JUDGE_VERSION, n as gepaOptimizationMethod, r as heldOutGate, t as skillOptOptimizationMethod, u as createReferenceEquivalenceJudge } from "../skillopt-optimization-method-Ceic2bfU.js";
|
|
19
19
|
import { n as paretoPolicy, r as paretoSignificanceGate, t as buildEvidenceVector } from "../promotion-policy-xzA40Evo.js";
|
|
20
20
|
import { createHash } from "node:crypto";
|
|
21
21
|
import { agentCandidateBenchmarkSuiteSchema, agentCandidateBenchmarkTaskSchema, agentCandidateBundleSchema, agentCandidateEvaluationPolicySchema, agentCandidateExperimentSchema, agentImprovementMeasuredComparisonSchema, agentProfileImprovementExperimentSchema, agentProfileImprovementMeasuredComparisonSchema, agentProfileImprovementRunCellSchema, agentProfileImprovementRunReceiptSchema, agentProfileImprovementSuiteInputsSchema, agentProfileImprovementSuiteSchema, agentProfileImprovementTaskSchema, candidateExecutionEvidenceSchema, canonicalCandidateDigest, canonicalCandidateJson, numbersApproximatelyEqual, omitTopLevelDigest } from "@tangle-network/agent-interface";
|
|
@@ -1,8 +1,8 @@
|
|
|
1
|
-
import { i as AnalystFinding, n as AnalystContext, w as TraceAnalysisStore } from "./types-
|
|
2
|
-
import { l as ExactCapableAnalyst } from "./exact-types-
|
|
3
|
-
import { n as AnalystRegistry, r as AnalystRegistryOptions } from "./registry-
|
|
4
|
-
import { a as TraceAnalystLimits, d as RawAnalystFinding, i as TraceAnalysisEngineResult, n as TraceAnalysisEngine } from "./engine-
|
|
5
|
-
import { t as TraceToolGroupName } from "./tool-groups-
|
|
1
|
+
import { i as AnalystFinding, n as AnalystContext, w as TraceAnalysisStore } from "./types-vXyshMwx.js";
|
|
2
|
+
import { l as ExactCapableAnalyst } from "./exact-types-CzbhVDr2.js";
|
|
3
|
+
import { n as AnalystRegistry, r as AnalystRegistryOptions } from "./registry-B_1Frl8a.js";
|
|
4
|
+
import { a as TraceAnalystLimits, d as RawAnalystFinding, i as TraceAnalysisEngineResult, n as TraceAnalysisEngine } from "./engine-DJqRKbhs.js";
|
|
5
|
+
import { t as TraceToolGroupName } from "./tool-groups-CZotz-e_.js";
|
|
6
6
|
//#region src/trace-analyst/behavioral-metrics.d.ts
|
|
7
7
|
type SuboptimalCode = 'monotonic-input-growth' | 'output-length-decay' | 'single-tool-dependency' | 'no-self-verification';
|
|
8
8
|
interface SuboptimalSignal {
|
|
@@ -111,4 +111,4 @@ interface DefaultAnalystRegistryOptions {
|
|
|
111
111
|
declare function buildDefaultAnalystRegistry(options?: DefaultAnalystRegistryOptions): AnalystRegistry;
|
|
112
112
|
//#endregion
|
|
113
113
|
export { createTraceAnalyst as a, runTraceAnalyst as c, deriveEfficiencyFindings as d, TraceAnalystDefinition as i, BehavioralAnalystOptions as l, buildDefaultAnalystRegistry as n, renderPriorFindings as o, CreateTraceAnalystOptions as r, renderUpstreamFindings as s, DefaultAnalystRegistryOptions as t, behavioralAnalyst as u };
|
|
114
|
-
//# sourceMappingURL=default-registry-
|
|
114
|
+
//# sourceMappingURL=default-registry-cpHtP2Og.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"default-registry-
|
|
1
|
+
{"version":3,"file":"default-registry-cpHtP2Og.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;UC1BR;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgFI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA+QP,oBAAoB,OAAO;iBAe3B,uBAAuB,UAAU;;;UCnahC;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import { c as CostLedgerHandle, f as CostLedgerSummary, m as CostReceipt } from "./cost-ledger-DbQdN3nO.js";
|
|
2
|
-
import { _ as ProposalFinding } from "./types-
|
|
3
|
-
import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate } from "./types-
|
|
4
|
-
import { o as InsightReport } from "./insight-report-
|
|
5
|
-
import { C as RunOptimizationOptions, E as RunEvalOptions, J as OptimizationMethodProvenance, S as PremeasuredOptimizationBaseline, W as OptimizationMethod, Y as OptimizationMethodResult, b as RunImprovementLoopResult, l as LoopProvenanceRecord, xt as CampaignStorage } from "./provenance-
|
|
2
|
+
import { _ as ProposalFinding } from "./types-vXyshMwx.js";
|
|
3
|
+
import { H as SurfaceProposer, R as Scenario, S as JudgeConfig, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, p as Gate } from "./types-BjsNDR49.js";
|
|
4
|
+
import { o as InsightReport } from "./insight-report-BeT8KCgI.js";
|
|
5
|
+
import { C as RunOptimizationOptions, E as RunEvalOptions, J as OptimizationMethodProvenance, S as PremeasuredOptimizationBaseline, W as OptimizationMethod, Y as OptimizationMethodResult, b as RunImprovementLoopResult, l as LoopProvenanceRecord, xt as CampaignStorage } from "./provenance-BOtMtoiZ.js";
|
|
6
6
|
import { o as PowerPreflight } from "./pareto-BqNW3LJR.js";
|
|
7
|
-
import { n as HostedTenant } from "./client-
|
|
7
|
+
import { n as HostedTenant } from "./client-D_TIV9pJ.js";
|
|
8
8
|
//#region src/contract/self-improve.d.ts
|
|
9
9
|
interface SelfImproveBudget {
|
|
10
10
|
/** Hard spend cap across the full run. Each paid call reserves its enforced
|
|
@@ -385,4 +385,4 @@ interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
|
|
|
385
385
|
declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
|
|
386
386
|
//#endregion
|
|
387
387
|
export { DefinedAgentEval as a, SelfImproveOptions as c, SelfImproveRunError as d, selfImprove as f, DefineAgentEvalOptions as i, SelfImproveProgressEvent as l, AgentEvalEvaluateOptions as n, defineAgentEval as o, AgentEvalImproveOptions as r, SelfImproveBudget as s, AgentEvalAgent as t, SelfImproveResult as u };
|
|
388
|
-
//# sourceMappingURL=define-agent-eval-
|
|
388
|
+
//# sourceMappingURL=define-agent-eval-DBOpAM_g.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"define-agent-eval-
|
|
1
|
+
{"version":3,"file":"define-agent-eval-DBOpAM_g.d.ts","names":[],"sources":["../src/contract/self-improve.ts","../src/contract/define-agent-eval.ts"],"mappings":";;;;;;;;UA+DiB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU;;;;;;;;;;;;;;EAc9D,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;;EAIA,WAAW;;;EAIX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;EAaT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW,gBAAgB;;;;;;EAO3B,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ;;;;;EAMxB,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;EAIA,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;;EAeA,eAAe;;;EAIf,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD,WAAW;;;;;;;EAQX,mBAAmB,uBAAuB,WAAW;;UAGtC,kBAAkB,kBAAkB,UAAU;;;;EAI7D;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;EAEV;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;;EAKT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;;cAI9B,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+LhB,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KCviB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW"}
|
|
@@ -7,10 +7,10 @@ import { r as pairRunRecords } from "./paired-arms-D-XRF_fy.js";
|
|
|
7
7
|
import { o as pairedTTest, r as pairedBootstrap } from "./paired-tests-BHIhYVdu.js";
|
|
8
8
|
import { r as welchsTTest } from "./baseline-BhPRQBVn.js";
|
|
9
9
|
import { r as observedSplitScore } from "./reward-nw2xZGZG.js";
|
|
10
|
-
import {
|
|
10
|
+
import { c as validateRunRecord, i as modelHasSnapshot } from "./run-record-D2lDdSAz.js";
|
|
11
11
|
import { r as paretoChart } from "./summary-report-Blysd6Z2.js";
|
|
12
|
-
import { A as surfaceHash, G as runEval, J as resolveRunDir, f as runImprovementLoop, k as surfaceContentHash, o as emitLoopProvenance, s as loopProvenanceArgsFromResult, y as assertOptimizationResult, z as defaultProductionGate } from "./llm-judge-
|
|
13
|
-
import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "./reward-hacking-
|
|
12
|
+
import { A as surfaceHash, G as runEval, J as resolveRunDir, f as runImprovementLoop, k as surfaceContentHash, o as emitLoopProvenance, s as loopProvenanceArgsFromResult, y as assertOptimizationResult, z as defaultProductionGate } from "./llm-judge-BO1LGjdn.js";
|
|
13
|
+
import { c as campaignCellTaskScore, l as campaignCellToRunRecord, o as campaignCellExecutionEvidence, s as campaignCellJudgeDimensions } from "./reward-hacking-DFo2FU5J.js";
|
|
14
14
|
import { S as inMemoryCampaignStorage, b as createRunCostLedger, x as fsCampaignStorage } from "./external-optimizer-subprocess-BhKYK0Jv.js";
|
|
15
15
|
import { t as powerPreflight } from "./power-preflight-DEw-uC7q.js";
|
|
16
16
|
import { t as createHostedClient } from "./client-LIuo-KPv.js";
|
|
@@ -1491,4 +1491,4 @@ function requirePositiveInteger(value, field) {
|
|
|
1491
1491
|
//#endregion
|
|
1492
1492
|
export { summarizeExecution as a, analyzeRuns as i, SelfImproveRunError as n, checkCanaries as o, selfImprove as r, defineAgentEval as t };
|
|
1493
1493
|
|
|
1494
|
-
//# sourceMappingURL=define-agent-eval-
|
|
1494
|
+
//# sourceMappingURL=define-agent-eval-m5lM7XGE.js.map
|