@tangle-network/agent-eval 0.144.1 → 0.144.3
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 +24 -0
- package/dist/analyst/index.d.ts +9 -9
- package/dist/analyst/index.js +3 -3
- package/dist/{analyze-runs-BScZvqMV.js → analyze-runs-DWIvOAGk.js} +2 -2
- package/dist/{analyze-runs-BScZvqMV.js.map → analyze-runs-DWIvOAGk.js.map} +1 -1
- package/dist/{benchmark-DxaZfy0w.d.ts → benchmark-CP6kWfj8.d.ts} +3 -3
- package/dist/{benchmark-DxaZfy0w.d.ts.map → benchmark-CP6kWfj8.d.ts.map} +1 -1
- package/dist/{benchmark-command-95Zqd1e8.js → benchmark-command-CQPKRUr-.js} +6 -6
- package/dist/{benchmark-command-95Zqd1e8.js.map → benchmark-command-CQPKRUr-.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +1 -1
- package/dist/benchmarks/index.js +1 -1
- package/dist/{benchmarks-CRT4ScEA.js → benchmarks-CkG1bWFa.js} +4 -4
- package/dist/{benchmarks-CRT4ScEA.js.map → benchmarks-CkG1bWFa.js.map} +1 -1
- package/dist/campaign/index.d.ts +4 -4
- package/dist/campaign/index.js +3 -3
- package/dist/{campaign-CefP3fPS.js → campaign-DjGFyPxH.js} +73 -39
- package/dist/campaign-DjGFyPxH.js.map +1 -0
- package/dist/cli.js +1 -1
- package/dist/{client-Cgl6KasJ.d.ts → client-Bbht4xxl.d.ts} +4 -4
- package/dist/{client-Cgl6KasJ.d.ts.map → client-Bbht4xxl.d.ts.map} +1 -1
- package/dist/{completion-verifier-D15NHYSk.d.ts → completion-verifier-EJERfFwF.d.ts} +3 -3
- package/dist/{completion-verifier-D15NHYSk.d.ts.map → completion-verifier-EJERfFwF.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +8 -8
- package/dist/contract/index.js +9 -7
- package/dist/contract/index.js.map +1 -1
- package/dist/control.d.ts +2 -2
- package/dist/control.js +1 -1
- package/dist/{default-registry-D3uqKbo6.d.ts → default-registry-iXfu2trt.d.ts} +5 -5
- package/dist/{default-registry-D3uqKbo6.d.ts.map → default-registry-iXfu2trt.d.ts.map} +1 -1
- package/dist/{dspy-rlm-engine-BR7TtB0o.js → dspy-rlm-engine-IRCG8kdi.js} +2 -2
- package/dist/{dspy-rlm-engine-BR7TtB0o.js.map → dspy-rlm-engine-IRCG8kdi.js.map} +1 -1
- package/dist/{eval-campaign-lI7wB9DH.js → eval-campaign-lZcDIwQM.js} +2 -2
- package/dist/{eval-campaign-lI7wB9DH.js.map → eval-campaign-lZcDIwQM.js.map} +1 -1
- package/dist/{exact-types-B0lJV3tu.d.ts → exact-types-BygCBR4L.d.ts} +2 -2
- package/dist/{exact-types-B0lJV3tu.d.ts.map → exact-types-BygCBR4L.d.ts.map} +1 -1
- package/dist/{feedback-trajectory-BCHqzLh3.d.ts → feedback-trajectory-CSIkRLQX.d.ts} +3 -3
- package/dist/{feedback-trajectory-BCHqzLh3.d.ts.map → feedback-trajectory-CSIkRLQX.d.ts.map} +1 -1
- package/dist/hosted/index.d.ts +3 -3
- package/dist/{index-BrpxwELi.d.ts → index-BuHs_OnD.d.ts} +17 -12
- package/dist/{index-BrpxwELi.d.ts.map → index-BuHs_OnD.d.ts.map} +1 -1
- package/dist/{index-D2enbqA0.d.ts → index-CNOCxBLh.d.ts} +2 -2
- package/dist/{index-D2enbqA0.d.ts.map → index-CNOCxBLh.d.ts.map} +1 -1
- package/dist/{index-DFLVtPZ9.d.ts → index-DGIzNtRv.d.ts} +2 -2
- package/dist/{index-DFLVtPZ9.d.ts.map → index-DGIzNtRv.d.ts.map} +1 -1
- package/dist/{index-DRI1DU5z.d.ts → index-D_qTihaQ.d.ts} +5 -5
- package/dist/{index-DRI1DU5z.d.ts.map → index-D_qTihaQ.d.ts.map} +1 -1
- package/dist/index.d.ts +20 -20
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +15 -15
- package/dist/{insight-report-Bh_8ksel.d.ts → insight-report-D5m1z0_n.d.ts} +3 -3
- package/dist/{insight-report-Bh_8ksel.d.ts.map → insight-report-D5m1z0_n.d.ts.map} +1 -1
- package/dist/ledger-core/index.js +1 -1
- package/dist/{ledger-core-Dxz0Rkwa.js → ledger-core-DXZIqu17.js} +3 -3
- package/dist/ledger-core-DXZIqu17.js.map +1 -0
- package/dist/meta-eval/index.d.ts +1 -1
- package/dist/{mint-Ctwk079K.js → mint-DD-0oQTA.js} +2 -2
- package/dist/{mint-Ctwk079K.js.map → mint-DD-0oQTA.js.map} +1 -1
- package/dist/multishot/index.d.ts +1 -1
- package/dist/openapi.json +1 -1
- package/dist/{propose-review-control-DLXz4FCX.js → propose-review-control-BciUCZoh.js} +2 -2
- package/dist/{propose-review-control-DLXz4FCX.js.map → propose-review-control-BciUCZoh.js.map} +1 -1
- package/dist/{release-report-B_bQOHM-.d.ts → release-report-BZRWdq_t.d.ts} +3 -3
- package/dist/{release-report-B_bQOHM-.d.ts.map → release-report-BZRWdq_t.d.ts.map} +1 -1
- package/dist/{release-report-B5XPBvAU.js → release-report-Sl0xfkFv.js} +2 -2
- package/dist/{release-report-B5XPBvAU.js.map → release-report-Sl0xfkFv.js.map} +1 -1
- package/dist/{replay-DjUfTrHD.js → replay-CqOsGjzU.js} +2 -9
- package/dist/replay-CqOsGjzU.js.map +1 -0
- package/dist/{replay-BuJM6kLh.d.ts → replay-DQ-55DC_.d.ts} +4 -4
- package/dist/{replay-BuJM6kLh.d.ts.map → replay-DQ-55DC_.d.ts.map} +1 -1
- package/dist/reporting.d.ts +3 -3
- package/dist/reporting.js +1 -1
- package/dist/{researcher-C6lzl-rP.d.ts → researcher-BSiCoM1s.d.ts} +3 -3
- package/dist/{researcher-C6lzl-rP.d.ts.map → researcher-BSiCoM1s.d.ts.map} +1 -1
- package/dist/{reward-hacking-DjTi9HLb.js → reward-hacking-CyuzxKly.js} +2 -2
- package/dist/{reward-hacking-DjTi9HLb.js.map → reward-hacking-CyuzxKly.js.map} +1 -1
- package/dist/{reward-hacking-CEVVmy3h.d.ts → reward-hacking-DFgkEY4p.d.ts} +2 -2
- package/dist/{reward-hacking-CEVVmy3h.d.ts.map → reward-hacking-DFgkEY4p.d.ts.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-8nj3mYvx.js → rollout-C2fD1cf4.js} +2 -2
- package/dist/{rollout-8nj3mYvx.js.map → rollout-C2fD1cf4.js.map} +1 -1
- package/dist/{rubric-predictive-validity-C2CthIfY.d.ts → rubric-predictive-validity-C4r4Y-q8.d.ts} +2 -2
- package/dist/{rubric-predictive-validity-C2CthIfY.d.ts.map → rubric-predictive-validity-C4r4Y-q8.d.ts.map} +1 -1
- package/dist/{run-evidence-8Ou28QSa.d.ts → run-evidence-H1vRpIdT.d.ts} +3 -3
- package/dist/{run-evidence-8Ou28QSa.d.ts.map → run-evidence-H1vRpIdT.d.ts.map} +1 -1
- package/dist/{run-record-vRgqWmJw.js → run-record-CWN8-VsV.js} +40 -14
- package/dist/{run-record-vRgqWmJw.js.map → run-record-CWN8-VsV.js.map} +1 -1
- package/dist/{run-record-Tb3TTtUn.d.ts → run-record-ooo9FWns.d.ts} +5 -9
- package/dist/{run-record-Tb3TTtUn.d.ts.map → run-record-ooo9FWns.d.ts.map} +1 -1
- package/dist/{semantic-concept-judge-l8Lb3Kqz.js → semantic-concept-judge-Do5aM9wP.js} +2 -2
- package/dist/{semantic-concept-judge-l8Lb3Kqz.js.map → semantic-concept-judge-Do5aM9wP.js.map} +1 -1
- package/dist/{single-run-lock-B-cFnnNn.js → single-run-lock-D5iN0Xzb.js} +2 -2
- package/dist/{single-run-lock-B-cFnnNn.js.map → single-run-lock-D5iN0Xzb.js.map} +1 -1
- package/dist/{skill-usage-3vUWdZ9n.d.ts → skill-usage-DtpLou9L.d.ts} +6 -6
- package/dist/{skill-usage-3vUWdZ9n.d.ts.map → skill-usage-DtpLou9L.d.ts.map} +1 -1
- package/dist/{skillopt-optimization-method-DA5mJmFE.js → skillopt-optimization-method-C4FX42dy.js} +29 -8
- package/dist/skillopt-optimization-method-C4FX42dy.js.map +1 -0
- package/dist/{skillopt-optimization-method-DNniakI5.d.ts → skillopt-optimization-method-CvSJGdm3.d.ts} +7 -7
- package/dist/{skillopt-optimization-method-DNniakI5.d.ts.map → skillopt-optimization-method-CvSJGdm3.d.ts.map} +1 -1
- package/dist/{summary-report-o3eJ3gxG.d.ts → summary-report-BOM6dfP7.d.ts} +2 -2
- package/dist/{summary-report-o3eJ3gxG.d.ts.map → summary-report-BOM6dfP7.d.ts.map} +1 -1
- package/dist/{tool-groups-Bqy4A3QB.d.ts → tool-groups-CMmsgTzj.d.ts} +7 -7
- package/dist/{tool-groups-Bqy4A3QB.d.ts.map → tool-groups-CMmsgTzj.d.ts.map} +1 -1
- package/dist/traces.d.ts +4 -4
- package/dist/traces.js +1 -1
- package/dist/{types-KEqL1pZc.d.ts → types-DcJxgsLy.d.ts} +8 -6
- package/dist/{types-KEqL1pZc.d.ts.map → types-DcJxgsLy.d.ts.map} +1 -1
- package/dist/{types-D3jh6F98.d.ts → types-y8jrxXWd.d.ts} +2 -2
- package/dist/{types-D3jh6F98.d.ts.map → types-y8jrxXWd.d.ts.map} +1 -1
- package/dist/wire/index.d.ts +1 -1
- package/docs/eval-surface-map.md +4 -0
- package/package.json +1 -1
- package/dist/campaign-CefP3fPS.js.map +0 -1
- package/dist/ledger-core-Dxz0Rkwa.js.map +0 -1
- package/dist/replay-DjUfTrHD.js.map +0 -1
- package/dist/skillopt-optimization-method-DA5mJmFE.js.map +0 -1
|
@@ -1,14 +1,14 @@
|
|
|
1
1
|
import { o as Severity } from "./multi-layer-verifier-BHY1gWAc.js";
|
|
2
2
|
import { b as CustomTokenPricing, c as CostLedgerHandle } from "./cost-ledger-FuQvHxPm.js";
|
|
3
3
|
import { C as TraceEvent, _ as Span, f as Run, n as BudgetLedgerEntry, t as Artifact } from "./schema-BtVldJ3T.js";
|
|
4
|
-
import "./index-
|
|
4
|
+
import "./index-DGIzNtRv.js";
|
|
5
5
|
import { q as LlmClientOptions } from "./types-BjMFz88h.js";
|
|
6
6
|
import { s as TraceStore } from "./store-CT9YIIve.js";
|
|
7
|
-
import { i as AnalystFinding, n as AnalystContext, r as AnalystCost, t as Analyst, x as TraceAnalysisStore } from "./types-
|
|
8
|
-
import { i as TraceAnalystDefinition } from "./default-registry-
|
|
9
|
-
import { l as ExactCapableAnalyst } from "./exact-types-
|
|
7
|
+
import { i as AnalystFinding, n as AnalystContext, r as AnalystCost, t as Analyst, x as TraceAnalysisStore } from "./types-y8jrxXWd.js";
|
|
8
|
+
import { i as TraceAnalystDefinition } from "./default-registry-iXfu2trt.js";
|
|
9
|
+
import { l as ExactCapableAnalyst } from "./exact-types-BygCBR4L.js";
|
|
10
10
|
import { f as ExternalOptimizerModelExecutionObservation, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, r as ExternalOptimizerCallbackLimits } from "./external-optimizer-contracts-CdmX2K2S.js";
|
|
11
|
-
import { i as TraceAnalysisEngine } from "./tool-groups-
|
|
11
|
+
import { i as TraceAnalysisEngine } from "./tool-groups-CMmsgTzj.js";
|
|
12
12
|
import { b as SupervisorRunTree, y as SupervisorRunSources } from "./types-yLK8gXE9.js";
|
|
13
13
|
import { z } from "zod";
|
|
14
14
|
//#region src/run-score.d.ts
|
|
@@ -574,4 +574,4 @@ declare class SkillUsageAnalyst implements ExactCapableAnalyst<SkillUsageReport>
|
|
|
574
574
|
declare const SKILL_USAGE_ANALYST: SkillUsageAnalyst;
|
|
575
575
|
//#endregion
|
|
576
576
|
export { RunScore as $, parseFindingSubject as A, ConceptSpec as B, FINDING_SUBJECT_KINDS as C, FindingSubjectStringSchema as D, FindingSubjectKind as E, DefineExactCustomAnalystOptions as F, SemanticConceptJudgeOptions as G, DEFAULT_COMPLEXITY_WEIGHTS as H, defineCustomAnalyst as I, runSemanticConceptJudge as J, SemanticConceptJudgeResult as K, defineTraceAnalyst as L, DspyRlmTraceEngineOptions as M, createDspyRlmTraceEngine as N, KIND_EXPECTED_SUBJECTS as O, DefineCustomAnalystOptions as P, DEFAULT_RUN_SCORE_WEIGHTS as Q, ConceptComplexity as R, FINDING_SUBJECT_GRAMMAR_PROMPT as S, FindingSubject as T, SEMANTIC_CONCEPT_JUDGE_VERSION as U, ConceptWeightStrategy as V, SemanticConceptJudgeInput as W, RunCriticOptions as X, RunCritic as Y, RunTrace as Z, FindingsDiff as _, SkillUsageScanConfig as a, defaultIsMaterial as b, DEFAULT_TRACE_ANALYST_KINDS as c, IMPROVEMENT_KIND_SPEC as d, RunScoreWeights as et, FAILURE_MODE_KIND_SPEC as f, DiffPolicy as g, emitControlIntegrityFindings as h, SkillUsageReport as i, renderFindingSubject as j, findingSubjectGrammarPromptFor as k, KNOWLEDGE_POISONING_KIND_SPEC as l, ControlIntegrityAnalyst as m, SkillUsageAnalyst as n, clamp01 as nt, buildSkillUsageReport as o, CONTROL_INTEGRITY_ANALYST as p, createSemanticConceptJudge as q, SkillUsageRecord as r, emitSkillUsageFindings as s, SKILL_USAGE_ANALYST as t, aggregateRunScore as tt, KNOWLEDGE_GAP_KIND_SPEC as u, FindingsStore as v, FINDING_SUBJECT_SYNTAX as w, diffFindings as x, PersistedFinding as y, ConceptFinding as z };
|
|
577
|
-
//# sourceMappingURL=skill-usage-
|
|
577
|
+
//# sourceMappingURL=skill-usage-DtpLou9L.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"skill-usage-
|
|
1
|
+
{"version":3,"file":"skill-usage-DtpLou9L.d.ts","names":[],"sources":["../src/run-score.ts","../src/run-critic.ts","../src/semantic-concept-judge.ts","../src/analyst/define.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts","../src/analyst/finding-subject.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/control-integrity.ts","../src/analyst/kinds/failure-mode.ts","../src/analyst/kinds/improvement.ts","../src/analyst/kinds/knowledge-gap.ts","../src/analyst/kinds/knowledge-poisoning.ts","../src/analyst/kinds/index.ts","../src/analyst/kinds/skill-usage.ts"],"mappings":";;;;;;;;;;;;;;UAAiB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;cAGW,2BAA2B;iBAcxB,kBAAkB,OAAO,UAAU,UAAS,QAAQ;iBAiBpD,QAAQ;;;UCxDP;EACf,KAAK;EACL,OAAO;EACP,QAAQ;EACR,WAAW;EACX,QAAQ;;UAGO;EACf,UAAU,QAAQ;EAClB,gBAAgB;;cAYL;mBACM;mBACA;EAEjB,YAAY,UAAS;EAKf,MAAM,OAAO,YAAY,gBAAgB,QAAQ;EAYvD,WAAW,OAAO,WAAW;EAmH7B,KAAK,OAAO;UAIJ;;;;;;;;;;;;;;;;;;;;;;;;;KC/GE;UAEK;EACf;;EAEA;;EAEA;;EAEA,aAAa;;UAGE;EACf;EACA;;EAEA;EACA;EACA,UAAU;;UAGK;;EAEf;;EAEA;;EAEA,aAAa;IAAQ;IAAc;;;EAEnC,kBAAkB;;EAElB;EACA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA;;EAEA;EACA;;;;;;;;KASU;cAEC,4BAA4B,OAAO;UAM/B;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,MAAM;EACN,aAAa;EACb;EACA,WAAW;EACX,SAAS;;;;;;EAMT,iBAAiB;;EAEjB,oBAAoB,QAAQ,OAAO;;cAKxB;;;;;;iBAkGS,wBACpB,OAAO,2BACP,UAAS,8BACR,QAAQ;;;;;iBAsJK,2BACd,UAAS,+BACP,OAAO,8BAA8B,QAAQ;;;;;;;;;;iBCxYjC,mBAAmB,YAAY,yBAAyB;UAkBvD;EACf;EACA;EACA;EACA,MAAM;EACN,SAAS,QAAQ;;UAGF,wCAAwC;;EAEvD,iBAAiB,SAAS;;;iBAIZ,oBACd,SAAS,kCACR,oBAAoB;iBACP,oBACd,SAAS,6BACR,QAAQ;;;KCjCC,0BAA0B;;;UCoBrB;;EAEf,MAAM;;EAEN;;EAEA,kBAAkB,aAAa;EAC/B;;EAEA,UAAU;;EAEV;;EAEA;;;;;;EAMA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,kBAAkB,QAAQ;;EAE1B;;;;;;;;;;EAUA;;EAEA,SAAS;;EAET;;;iBAIc,yBAAyB,SAAS,4BAA4B;;;;;;;;;;KCvClE;EAEN;EAAwB;EAAc;;EACtC;EAAyB;;EACzB;EAAuB;;EACvB;EAAyB;;EAGzB;EAAuB;;EACvB;EAAe;;EACf;EAAkB;EAAc;;EAChC;EAAkB;;EAClB;EAAa;EAAgB;;EAC7B;EAAc;;EACd;EAAkB;;EAClB;EAAkB;;EAClB;EAAwB;;EACxB;EAAuB;;EACvB;EAAc;;EACd;EAAa;EAAgB;;EAC7B;EAAgB;;EAChB;EAAqB;;EACrB;EAAuB;;EAEvB;EAA4B;;EAC5B;EAA2B;;EAE3B;EAAiB;;KAEX,qBAAqB;cAEpB,uBAAuB,cAAc;;;;;;;;;;;;;;;;;iBA2ClC,oBAAoB,iCAAiC;;;;;;;iBAoHrD,qBAAqB,GAAG;;;;;;;;;;cA8D3B,wBAAwB,SAAS,OAAO;cA0DxC;;;;;;;;;cAYA,wBAAwB,eAAe,cAAc;;iBAqElD,+BAA+B;;;;;;;;;;;;cAmBlC,4BAA0B,EAAA;;;;;;;;UC1atB,yBAAyB;EACxC;;cAGW;WAGiB;mBAFX;EAEjB,YAA4B;EAItB,OAAO,eAAe,UAAU,mBAAmB;;EAQzD,WAAW;;EAkBX,QAAQ,gBAAgB;;UAOT;;EAEf,UAAU;;EAEV,aAAa;;EAEb,WAAW;;;;;EAKX,SAAS;IAAQ,UAAU;IAAkB,SAAS;;;UAGvC;;;;;;;;EAQf,cAAc,UAAU,gBAAgB,SAAS;;;;;;iBAOnC,kBAAkB,GAAG,gBAAgB,GAAG;;;;;iBAWxC,aACd,UAAU,oBACV,SAAS,oBACT,SAAQ,aACP;;;;iBC7Fa,6BACd,OAAO,uBAAuB,mBAC9B,qBACC;;cA4BU,mCACA,oBAAoB,uBAAuB;WAE7C;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;;EAGI,QACJ,OAAO,uBAAuB,mBAC9B,KAAK,iBACJ,QAAQ;;cAUA,2BAAyB;;;cCHzB,wBAAwB;;;cC3BxB,uBAAuB;;;cCGvB,yBAAyB;;;cCPzB,+BAA+B;;;;;;;;;;;;cCP/B,sCAAsC;;;KCbvC;;UAGK;EACf;EACA,MAAM;;EAEN;EACA;;EAEA;;EAEA;;;EAGA;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;UAGe;EACf;EACA,SAAS;;UAGM;;EAEf;;EAEA;IAAc;IAAc,MAAM;;;EAElC;;;EAGA,kBAAkB;;EAElB;;;iBA2Ec,sBAAsB,QAAQ,uBAAuB;;iBAiHrD,uBACd,QAAQ,kBACR,qBACC;cA2HU,6BAA6B,oBAAoB;WACnD;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;aACA;;EAGI,QAAQ,OAAO,kBAAkB,KAAK,iBAAiB,QAAQ;;cAS1D,qBAAmB"}
|
package/dist/{skillopt-optimization-method-DA5mJmFE.js → skillopt-optimization-method-C4FX42dy.js}
RENAMED
|
@@ -2,11 +2,11 @@ import { c as ValidationError, i as JudgeError, t as AgentEvalError } from "./er
|
|
|
2
2
|
import { i as CostLedger, t as CostAccountingIncompleteError } from "./cost-ledger-DMFxsLKr.js";
|
|
3
3
|
import { f as maximumChargeForLlmRequest, l as costReceiptFromLlm, m as stripFencedJson, u as costReceiptFromLlmError } from "./llm-client-D3EoChAU.js";
|
|
4
4
|
import { a as clamp01, o as combineAbortSignals, t as assertProposalFindings } from "./proposal-findings-2GIUo1et.js";
|
|
5
|
-
import { A as removeCredentialEnvironment, D as isCandidateText, E as assertNoCredentialValues, M as resolveExternalOptimizerProcessLimits, N as safePathComponent, O as isExternalTextCandidate, T as assertJsonValue, d as closeExternalOptimizerResources, f as runWithCleanup, j as resolveExternalOptimizerCallbackLimits, k as isRecord, l as startExternalOptimizerCallback, n as createRunCostLedger, p as startExternalOptimizerModelProxy, r as fsCampaignStorage, t as acquireSingleRunLock, u as runExternalOptimizerProcess, v as canonicalJson, w as assertExternalOptimizerModelBudget, y as contentHash } from "./single-run-lock-
|
|
6
|
-
import { p as mapConcurrent } from "./ledger-core-
|
|
5
|
+
import { A as removeCredentialEnvironment, D as isCandidateText, E as assertNoCredentialValues, M as resolveExternalOptimizerProcessLimits, N as safePathComponent, O as isExternalTextCandidate, T as assertJsonValue, d as closeExternalOptimizerResources, f as runWithCleanup, j as resolveExternalOptimizerCallbackLimits, k as isRecord, l as startExternalOptimizerCallback, n as createRunCostLedger, p as startExternalOptimizerModelProxy, r as fsCampaignStorage, t as acquireSingleRunLock, u as runExternalOptimizerProcess, v as canonicalJson, w as assertExternalOptimizerModelBudget, y as contentHash } from "./single-run-lock-D5iN0Xzb.js";
|
|
6
|
+
import { p as mapConcurrent } from "./ledger-core-DXZIqu17.js";
|
|
7
7
|
import { E as pairedBootstrap, H as weightedComposite, M as pairedRiskDifferenceScore, N as pairedSignTest, O as pairedDeltaTieFraction, T as pairedBinaryScale, d as confidenceInterval, j as pairedRiskDifferenceExact } from "./statistics-ByxzSiOM.js";
|
|
8
8
|
import { t as DEFAULT_REDACTION_RULES } from "./redact-7Aq1ukl-.js";
|
|
9
|
-
import { a as campaignCellCostProvenance, o as campaignCellExecutionEvidence, t as detectRewardHacking, u as projectCampaignCellQuality } from "./reward-hacking-
|
|
9
|
+
import { a as campaignCellCostProvenance, o as campaignCellExecutionEvidence, t as detectRewardHacking, u as projectCampaignCellQuality } from "./reward-hacking-CyuzxKly.js";
|
|
10
10
|
import { z } from "zod";
|
|
11
11
|
import { writeFileSync } from "node:fs";
|
|
12
12
|
import { basename, isAbsolute, join } from "node:path";
|
|
@@ -1900,10 +1900,11 @@ async function executeCell(args) {
|
|
|
1900
1900
|
if (cached.status === "hit") {
|
|
1901
1901
|
const receiptProblem = cachedCellReceiptProblem(cached.cell, args.costLedger, stableCostTags);
|
|
1902
1902
|
if (receiptProblem === void 0) {
|
|
1903
|
-
|
|
1903
|
+
const cell = withCurrentAgentModelEvidence(cached.cell, args.costLedger, stableCostTags);
|
|
1904
|
+
enforceDispatchUsage(cell, args.opts.expectUsage ?? "warn");
|
|
1904
1905
|
return {
|
|
1905
1906
|
cell: {
|
|
1906
|
-
...
|
|
1907
|
+
...cell,
|
|
1907
1908
|
cached: true
|
|
1908
1909
|
},
|
|
1909
1910
|
artifactsByPath: {}
|
|
@@ -2047,7 +2048,7 @@ async function executeCell(args) {
|
|
|
2047
2048
|
...agentCost.cachedTokens > 0 ? { cached: agentCost.cachedTokens } : {},
|
|
2048
2049
|
...agentCost.cacheWriteTokens !== void 0 && agentCost.cacheWriteTokens > 0 ? { cacheWrite: agentCost.cacheWriteTokens } : {}
|
|
2049
2050
|
};
|
|
2050
|
-
const
|
|
2051
|
+
const agentModelEvidence = modelEvidenceFromReceipts(agentReceipts);
|
|
2051
2052
|
const dispatchResult = {
|
|
2052
2053
|
cellId: args.slot.cellId,
|
|
2053
2054
|
artifact,
|
|
@@ -2104,7 +2105,7 @@ async function executeCell(args) {
|
|
|
2104
2105
|
costProvenance: agentCost.costProvenance,
|
|
2105
2106
|
costCallIds,
|
|
2106
2107
|
tokenUsage,
|
|
2107
|
-
...
|
|
2108
|
+
...agentModelEvidence,
|
|
2108
2109
|
durationMs: Date.now() - startMs,
|
|
2109
2110
|
seed: args.slot.cellSeed,
|
|
2110
2111
|
cached: false,
|
|
@@ -2420,6 +2421,25 @@ function assertScheduleCachesReusable(args) {
|
|
|
2420
2421
|
}
|
|
2421
2422
|
if (blocked.length > 0) throw invalidCachedCellsError(blocked);
|
|
2422
2423
|
}
|
|
2424
|
+
function modelEvidenceFromReceipts(receipts) {
|
|
2425
|
+
const models = [...new Set(receipts.map((receipt) => receipt.model))];
|
|
2426
|
+
if (models.length === 0) return {};
|
|
2427
|
+
return {
|
|
2428
|
+
resolvedModels: models,
|
|
2429
|
+
...models.length === 1 ? { resolvedModel: models[0] } : {}
|
|
2430
|
+
};
|
|
2431
|
+
}
|
|
2432
|
+
function withCurrentAgentModelEvidence(cached, costLedger, stableCostTags) {
|
|
2433
|
+
const cachedCallIds = new Set(cached.costCallIds ?? []);
|
|
2434
|
+
const receipts = costLedger.list({
|
|
2435
|
+
channel: "agent",
|
|
2436
|
+
tags: stableCostTags
|
|
2437
|
+
}).filter((receipt) => cachedCallIds.has(receipt.callId));
|
|
2438
|
+
const cell = { ...cached };
|
|
2439
|
+
delete cell.resolvedModel;
|
|
2440
|
+
delete cell.resolvedModels;
|
|
2441
|
+
return Object.assign(cell, modelEvidenceFromReceipts(receipts));
|
|
2442
|
+
}
|
|
2423
2443
|
function cachedCellReceiptProblem(cached, costLedger, stableCostTags) {
|
|
2424
2444
|
const reportsPaidActivity = cached.costProvenance.kind !== "observed" || cached.costUsd > 0 || cached.tokenUsage.input > 0 || cached.tokenUsage.output > 0;
|
|
2425
2445
|
if (cached.costCallIds === void 0) {
|
|
@@ -6487,6 +6507,7 @@ function campaignMeasurementDigest(campaign) {
|
|
|
6487
6507
|
costProvenance: cell.costProvenance,
|
|
6488
6508
|
costCallIds: [...cell.costCallIds ?? []].sort(),
|
|
6489
6509
|
tokenUsage: cell.tokenUsage,
|
|
6510
|
+
resolvedModels: [...cell.resolvedModels ?? []].sort(),
|
|
6490
6511
|
resolvedModel: cell.resolvedModel ?? null,
|
|
6491
6512
|
durationMs: cell.durationMs,
|
|
6492
6513
|
seed: cell.seed,
|
|
@@ -7143,4 +7164,4 @@ function skillOptOptimizationMethod(config) {
|
|
|
7143
7164
|
//#endregion
|
|
7144
7165
|
export { assertCampaignSplitIdentity as $, assertOptimizationResult as A, summarizeAgentReceiptIntegrity as At, surfaceContentHash as B, detectScale as C, decidePairedPromotion as Ct, runCanaries as D, BackendIntegrityError as Dt, pairHoldout as E, pairedDeltaTest as Et, assertCodeSurfaceIdentity as F, DEFAULT_MUTATION_PRIMITIVES as G, campaignBreakdown as H, assertComponentSurface as I, planCampaignRun as J, buildReflectionPrompt as K, codeSurfaceIdentityMaterial as L, compareOptimizationMethods as M, JudgeParseError as Mt, costFromLedgerSummary as N, composeGate as O, assertRealAgentReceipts as Ot, optimizationTokenUsageFromSummary as P, assertCampaignDesign as Q, componentSurfaceIdentityMaterial as R, defaultProductionGate as S, recoverTruncatedJson as St, heldoutSignificance as T, minimumPairsForPairedDeltaTest as Tt, campaignMeanComposite as U, surfaceHash as V, compareRankKeys as W, resolveRunDir as X, runCampaign as Y, tangleTracesRoot as Z, buildEvidenceVector as _, crowdingDistance as _t, emitLoopProvenance as a, REFERENCE_EQUIVALENCE_JUDGE_VERSION as at, powerPreflight as b, paretoFrontierWithCrowding as bt, provenanceRecordPath as c, DEFAULT_RED_TEAM_CORPUS as ct, runImprovementLoop as d, scoreRedTeamOutput as dt, campaignScenarioIdentity as et, runOptimization as f, toolNamesForRun as ft, gepaOptimizationMethod as g, llmJudge as gt, labelTrustRank as h, hashScenarios as ht, canonicalDigest as i, REFERENCE_EQUIVALENCE_INPUT_LIMITS as it, combineComparisonCosts as j, summarizeBackendIntegrity as jt, externalTextOptimizationMethod as k, assertRealBackend as kt, provenanceSpansPath as l, redTeamDataset as lt, isProposedCandidate as m, HoldoutLockedError as mt, buildLoopProvenanceRecord as n, campaignSplitDigestFromIdentities as nt, loopProvenanceArgsFromResult as o, createReferenceEquivalenceJudge as ot, runEval as p, Dataset as pt, parseReflectionResponse as q, campaignMeasurementDigest as r, openAutoPr as rt, loopProvenanceSpans as s, runReferenceEquivalenceJudge as st, skillOptOptimizationMethod as t, campaignSplitDigest as tt, verifyLoopProvenanceRecord as u, redTeamReport as ut, paretoPolicy as v, dominates as vt, dimensionRegressions as w, pairedDecisionShape as wt, heldOutGate as x, scalarScore as xt, paretoSignificanceGate as y, paretoFrontier as yt, renderSurfaceDiff as z };
|
|
7145
7166
|
|
|
7146
|
-
//# sourceMappingURL=skillopt-optimization-method-
|
|
7167
|
+
//# sourceMappingURL=skillopt-optimization-method-C4FX42dy.js.map
|