@tangle-network/agent-eval 0.171.0 → 0.172.1
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 +85 -0
- package/README.md +3 -0
- package/dist/adapters/http.d.ts +2 -2
- package/dist/analyst/index.d.ts +5 -5
- package/dist/analyst/index.js +2 -2
- package/dist/{experiment-tracker-Dm8yQMqb.d.ts → attestation-CJBGmMVh.d.ts} +78 -2
- package/dist/attestation-CJBGmMVh.d.ts.map +1 -0
- package/dist/{experiment-tracker-BKEumQug.js → attestation-XSUpbc4o.js} +96 -2
- package/dist/attestation-XSUpbc4o.js.map +1 -0
- package/dist/{benchmark-command-D8k3Gf0J.js → benchmark-command--qeZUHbu.js} +2 -2
- package/dist/{benchmark-command-D8k3Gf0J.js.map → benchmark-command--qeZUHbu.js.map} +1 -1
- package/dist/benchmarks/index.d.ts +4 -4
- package/dist/benchmarks/index.js +2 -2
- package/dist/bounded-process-VIi0KSL2.js +212 -0
- package/dist/bounded-process-VIi0KSL2.js.map +1 -0
- package/dist/builder-eval/index.js +39 -98
- package/dist/builder-eval/index.js.map +1 -1
- package/dist/campaign/index.d.ts +8 -7
- package/dist/campaign/index.js +4 -4
- package/dist/{campaign-B72njjHj.js → campaign-Dp35pBbS.js} +4 -4
- package/dist/{campaign-B72njjHj.js.map → campaign-Dp35pBbS.js.map} +1 -1
- package/dist/canonical-CFpojCN5.d.ts +31 -0
- package/dist/canonical-CFpojCN5.d.ts.map +1 -0
- package/dist/{chat-client-DEtybj5i.js → chat-client-DI79OPye.js} +2 -2
- package/dist/{chat-client-DEtybj5i.js.map → chat-client-DI79OPye.js.map} +1 -1
- package/dist/cli.js +1 -1
- package/dist/{client-CDtcZ3p9.d.ts → client-Df7wdslk.d.ts} +2 -2
- package/dist/{client-CDtcZ3p9.d.ts.map → client-Df7wdslk.d.ts.map} +1 -1
- package/dist/contract/index.d.ts +9 -9
- package/dist/contract/index.js +4 -4
- package/dist/{default-registry-B0s2zU-s.d.ts → default-registry-XxedTLwu.d.ts} +3 -3
- package/dist/{default-registry-B0s2zU-s.d.ts.map → default-registry-XxedTLwu.d.ts.map} +1 -1
- package/dist/{define-agent-eval-Cjy2yhqP.d.ts → define-agent-eval-0wW7gFhr.d.ts} +4 -4
- package/dist/{define-agent-eval-Cjy2yhqP.d.ts.map → define-agent-eval-0wW7gFhr.d.ts.map} +1 -1
- package/dist/{define-agent-eval-Dy8QgxAI.js → define-agent-eval-jS8xj_Q_.js} +2 -2
- package/dist/{define-agent-eval-Dy8QgxAI.js.map → define-agent-eval-jS8xj_Q_.js.map} +1 -1
- package/dist/descriptive-B2iPaT9J.d.ts +89 -0
- package/dist/descriptive-B2iPaT9J.d.ts.map +1 -0
- package/dist/{engine-CAmTUk52.d.ts → engine-BfRay1qD.d.ts} +2 -2
- package/dist/{engine-CAmTUk52.d.ts.map → engine-BfRay1qD.d.ts.map} +1 -1
- package/dist/experiment/index.d.ts +3 -54
- package/dist/experiment/index.d.ts.map +1 -1
- package/dist/experiment/index.js +1 -95
- package/dist/experiment/index.js.map +1 -1
- package/dist/{heldout-gate-Dh2b62w8.d.ts → heldout-gate-JgNRDZwZ.d.ts} +3 -3
- package/dist/{heldout-gate-Dh2b62w8.d.ts.map → heldout-gate-JgNRDZwZ.d.ts.map} +1 -1
- package/dist/hosted/index.d.ts +1 -1
- package/dist/{index-lfaSeKSD.d.ts → index-D-UdhAmg.d.ts} +3 -31
- package/dist/index-D-UdhAmg.d.ts.map +1 -0
- package/dist/{index-DT73JraI.d.ts → index-DDAPhUJJ.d.ts} +4 -4
- package/dist/{index-DT73JraI.d.ts.map → index-DDAPhUJJ.d.ts.map} +1 -1
- package/dist/{index-fNXZMCzX.d.ts → index-DnglhM0A.d.ts} +9 -9
- package/dist/{index-fNXZMCzX.d.ts.map → index-DnglhM0A.d.ts.map} +1 -1
- package/dist/{index-8VIogTyS.d.ts → index-_vPrVMRX.d.ts} +6 -6
- package/dist/{index-8VIogTyS.d.ts.map → index-_vPrVMRX.d.ts.map} +1 -1
- package/dist/index.d.ts +154 -15
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +7 -6
- package/dist/index.js.map +1 -1
- package/dist/{integrity-CyWSSoQS.js → integrity-BWywb34E.js} +34 -11
- package/dist/{integrity-CyWSSoQS.js.map → integrity-BWywb34E.js.map} +1 -1
- package/dist/ledger-core/index.d.ts +2 -1
- package/dist/{llm-judge-BtJ2Sfk_.js → llm-judge-aQHIk5_-.js} +16 -10
- package/dist/llm-judge-aQHIk5_-.js.map +1 -0
- package/dist/{matrix-DiHmUobV.d.ts → matrix-Ch8JO1pG.d.ts} +2 -2
- package/dist/{matrix-DiHmUobV.d.ts.map → matrix-Ch8JO1pG.d.ts.map} +1 -1
- package/dist/meta-eval/index.d.ts +162 -2
- package/dist/meta-eval/index.d.ts.map +1 -1
- package/dist/meta-eval/index.js +287 -2
- package/dist/meta-eval/index.js.map +1 -1
- package/dist/multishot/golden/index.d.ts +1 -1
- package/dist/multishot/index.d.ts +2 -2
- package/dist/openapi.json +1 -1
- package/dist/{produced-state-Cm6DU_Ao.js → produced-state-CxmbFxFd.js} +2 -2
- package/dist/{produced-state-Cm6DU_Ao.js.map → produced-state-CxmbFxFd.js.map} +1 -1
- package/dist/{promotion-policy-WSXtBgBb.d.ts → promotion-policy-BBBcz5_3.d.ts} +2 -2
- package/dist/{promotion-policy-WSXtBgBb.d.ts.map → promotion-policy-BBBcz5_3.d.ts.map} +1 -1
- package/dist/{provenance-CafMdZKM.d.ts → provenance-Dp-vvyrU.d.ts} +13 -5
- package/dist/provenance-Dp-vvyrU.d.ts.map +1 -0
- package/dist/rl.d.ts +1 -1
- package/dist/sandbox-harness-BlSOu4LX.d.ts.map +1 -1
- package/dist/{skillopt-optimization-method-DzlF2RM7.js → skillopt-optimization-method-LHi02MzH.js} +2 -2
- package/dist/{skillopt-optimization-method-DzlF2RM7.js.map → skillopt-optimization-method-LHi02MzH.js.map} +1 -1
- package/dist/{statistical-heldout-UhiexnjU.d.ts → statistical-heldout-Yldkntvy.d.ts} +2 -2
- package/dist/{statistical-heldout-UhiexnjU.d.ts.map → statistical-heldout-Yldkntvy.d.ts.map} +1 -1
- package/dist/{store-tool-spans-D_qMl2__.d.ts → store-tool-spans-BvdUbeOB.d.ts} +3 -3
- package/dist/{store-tool-spans-D_qMl2__.d.ts.map → store-tool-spans-BvdUbeOB.d.ts.map} +1 -1
- package/dist/supervisor-run/index.d.ts +35 -2
- package/dist/supervisor-run/index.d.ts.map +1 -1
- package/dist/supervisor-run/index.js +83 -38
- package/dist/supervisor-run/index.js.map +1 -1
- package/dist/{tool-groups-RGYfVWpc.d.ts → tool-groups-DjwlMBvW.d.ts} +2 -2
- package/dist/tool-groups-DjwlMBvW.d.ts.map +1 -0
- package/dist/trace-repair/index.d.ts +1 -1
- package/dist/traces.d.ts +2 -2
- package/dist/{types-CMyW4GnH.d.ts → types-CCZ34qmV.d.ts} +2 -2
- package/dist/{types-CMyW4GnH.d.ts.map → types-CCZ34qmV.d.ts.map} +1 -1
- package/dist/{types-DeIUdzNd.d.ts → types-CoPUTiXb.d.ts} +23 -92
- package/dist/types-CoPUTiXb.d.ts.map +1 -0
- package/dist/{types-JHMOqZI4.d.ts → types-nokrtr7M.d.ts} +11 -1
- package/dist/types-nokrtr7M.d.ts.map +1 -0
- package/docs/eval-surface-map.md +36 -0
- package/docs/insight-report.md +19 -0
- package/docs/plants.md +123 -0
- package/docs/public-api.md +45 -20
- package/package.json +1 -1
- package/dist/experiment-tracker-BKEumQug.js.map +0 -1
- package/dist/experiment-tracker-Dm8yQMqb.d.ts.map +0 -1
- package/dist/index-lfaSeKSD.d.ts.map +0 -1
- package/dist/llm-judge-BtJ2Sfk_.js.map +0 -1
- package/dist/provenance-CafMdZKM.d.ts.map +0 -1
- package/dist/tool-groups-RGYfVWpc.d.ts.map +0 -1
- package/dist/types-DeIUdzNd.d.ts.map +0 -1
- package/dist/types-JHMOqZI4.d.ts.map +0 -1
|
@@ -1,8 +1,8 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { a as RunRecord } from "./run-record-DQjRcYwA.js";
|
|
3
|
-
import { A as CanaryOptions, F as OptimizationMethod,
|
|
3
|
+
import { A as CanaryOptions, F as OptimizationMethod, U as OptimizationPackageSource, b as RedTeamCase, er as ServedModelPolicy, yt as SearchRunIdentity } from "./provenance-Dp-vvyrU.js";
|
|
4
4
|
import { C as LlmCallMetadata, p as ChatClient } from "./types-Bfk0uxRj.js";
|
|
5
|
-
import { R as Scenario, S as JudgeConfig, i as CampaignCostMeter, j as MutableSurface, o as CampaignScenarioIdentity, p as Gate } from "./types-
|
|
5
|
+
import { R as Scenario, S as JudgeConfig, i as CampaignCostMeter, j as MutableSurface, o as CampaignScenarioIdentity, p as Gate } from "./types-nokrtr7M.js";
|
|
6
6
|
import { _ as ExternalTextCandidate, c as ExternalOptimizerModelBudget, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, m as ExternalOptimizerResumeMode, r as ExternalOptimizerCallbackLimits, v as ExternalTextEvaluationRequest } from "./external-optimizer-contracts-szBJ_1vh.js";
|
|
7
7
|
import { t as DetectRewardHackingInput } from "./reward-hacking-CgPRUesA.js";
|
|
8
8
|
//#region src/campaign/coverage.d.ts
|
|
@@ -450,4 +450,4 @@ interface HeldOutGateOptions<TScenario extends Scenario = Scenario> {
|
|
|
450
450
|
declare function heldOutGate<TArtifact, TScenario extends Scenario>(options: HeldOutGateOptions<TScenario>): Gate<TArtifact, TScenario>;
|
|
451
451
|
//#endregion
|
|
452
452
|
export { ReferenceEquivalenceJudgeInput as A, campaignSplitDigestFromIdentities as B, ExternalTextOptimizerContext as C, decodeExternalTextCandidate as D, ExternalTextEvaluationResponse as E, runReferenceEquivalenceJudge as F, assertCampaignDesign as I, assertCampaignSplitIdentity as L, ReferenceEquivalenceJudgeResult as M, ReferenceEquivalenceScenario as N, REFERENCE_EQUIVALENCE_INPUT_LIMITS as O, createReferenceEquivalenceJudge as P, campaignScenarioIdentity as R, ExternalTextOptimizationMethodConfig as S, ExternalOptimizationExample as T, GepaRunnerCommand as _, DefaultProductionRewardHackingOptions as a, OptimizerModelBudget as b, SkillOptOptimizationMethodConfig as c, skillOptOptimizationMethod as d, GepaAdaptiveEngineRun as f, GepaOptimizationRecipe as g, GepaOptimizationMethodConfig as h, DefaultProductionGateOptions as i, ReferenceEquivalenceJudgeOptions as j, REFERENCE_EQUIVALENCE_JUDGE_VERSION as k, SkillOptRunnerCommand as l, GepaEngineRun as m, heldOutGate as n, defaultProductionGate as o, GepaEngineOptions as p, DefaultProductionGateCheck as r, composeGate as s, HeldOutGateOptions as t, SkillOptTrainerConfig as u, gepaOptimizationMethod as v, ExternalTextOptimizerResult as w, externalTextOptimizationMethod as x, OpenAICompatibleOptimizerModel as y, campaignSplitDigest as z };
|
|
453
|
-
//# sourceMappingURL=heldout-gate-
|
|
453
|
+
//# sourceMappingURL=heldout-gate-JgNRDZwZ.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"heldout-gate-
|
|
1
|
+
{"version":3,"file":"heldout-gate-JgNRDZwZ.d.ts","names":[],"sources":["../src/campaign/coverage.ts","../src/reference-equivalence-judge.ts","../src/campaign/external-text-evaluation.ts","../src/campaign/external-text-optimization-contract.ts","../src/campaign/external-text-optimization.ts","../src/campaign/optimizer-model.ts","../src/campaign/gepa-optimization-method.ts","../src/campaign/skillopt-optimization-method.ts","../src/campaign/gates/compose.ts","../src/campaign/gates/default-production-gate.ts","../src/campaign/gates/heldout-gate.ts"],"mappings":";;;;;;;;;iBAWgB,qBAAqB,kBAAkB,UACrD,oBAAoB,aACpB;;iBA2Bc,yBAAyB,kBAAkB,UACzD,UAAU,YACT,2BAA2B,KAAK;;iBAUnB,kCACd,oBAAoB,4BACpB;;iBAgBc,oBAAoB,kBAAkB,UACpD,oBAAoB,aACpB;;iBAOc,4BACd,oBAAoB,4BACpB,cACA;;;cC3EW;cAEA;WACX;WACA;WACA;;UAGe,qCAAqC;EACpD;EACA;;UAGe;EACf;EACA;EACA;;UAGe;;EAEf,MAAM;;EAEN;;EAEA,SAAS;;EAET,aAAa;;UAGE,wCAAwC;EACvD;EACA;EACA;EACA;;;iBA0Bc,gCACd,SAAS,mCACR,oBAAoB;;iBA2CD,6BACpB,OAAO,gCACP,SAAS,mCACR,QAAQ;;;UCrGM;EACf;EACA;;UAGe;EACf;EACA;IACE;IACA,YAAY;IACZ;IACA;;;IAGA;IACA;MACE;MACA;;;;iBAwHU,4BAA4B,WAAW,wBAAwB;;;UC3I9D;WACN;WACA;WACA;WACA;WACA;WACA,eAAe;WACf,mBAAmB;WACnB,uBAAuB;WACvB;WACA;;WAEA;WACA;WACA;WACA,QAAQ;;WAER,MAAM;WACN,WACP,SAFa,kCAGV,QAAQ;;UAGE;EACf,eAAe;EACf;EACA;IACM;;IACA;;IACA;IAAkB;;;;;;;;;UAqBT,qCACf,kBAAkB,UAClB;EAEA;EACA,QAAQ,KAAK;EACb;EACA;EACA;EACA;;EAEA;;EAEA;;EAEA,SAAS;EACT;EACA;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,MAAM,SAAS,iCAAiC,QAAQ;;;;;;;;;;iBCN1C,+BAA+B,kBAAkB,UAAU,WACzE,QAAQ,qCAAqC,WAAW,aACvD,mBAAmB,WAAW;;;KCvErB,uBAAuB;;UAGlB;EACf;EACA,QAAQ;;EAER,MAAM;;EAEN;;;;;;;EAOA,oBAAoB;;;;;;;EAOpB;;;;;UCyCe;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;;;;EAKA,eAAe;;;UAIA,sBAAsB;;EAErC;;;KAIU,wBAAwB;;;;;;;KAQxB;EAEN;EACA,KAAK;;EAGL;EACA,eAAe;;EAGf;EACA,eAAe;;EAEf;;EAEA;EACA;EACA;EACA;EACA;EACA;EACA;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,eAAe;EACf;;EAGA;EACA,kBAAkB;EAClB,cAAc;EACd;;;KAIM,oBAAoB;UAEf,6BAA6B,kBAAkB,UAAU;;EAExE;;EAEA,QAAQ;;EAER;;EAEA;;EAEA;;;;;EAKA;;EAEA;;EAEA;;EAEA,2BAA2B,QAAQ;;EAEnC;;;;;;EAMA,YAAY;;;;;;EAMZ,oBAAoB,UAAU;;EAE9B,oBAAoB,UAAU,WAAW,UAAU;;EAEnD,SAAS;;;;;EAKT;EACA,SAAS;;;;;;;;;;EAUT;IAAiB,UAAU;IAAmB;;;iBA6ChC,uBAAuB,kBAAkB,UAAU,WACjE,QAAQ,6BAA6B,WAAW,aAC/C,mBAAmB,WAAW;;;UC7LhB;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;EAKA,YAAY;;KAGF,wBAAwB;UAEnB,iCAAiC,kBAAkB,UAAU;EAC5E;;EAEA;EACA;;EAEA;EACA,SAAS;;;;;EAKT,WAAW;;EAEX;;EAEA;EACA;;EAEA;;EAEA,2BAA2B,QAAQ;EACnC;EACA,oBAAoB,UAAU;EAC9B,oBAAoB,UAAU,WAAW,UAAU;EACnD,SAAS;EACT,SAAS;;;iBAIK,2BAA2B,kBAAkB,UAAU,WACrE,QAAQ,iCAAiC,WAAW,aACnD,mBAAmB,WAAW;;;;;;iBC1GjB,YAAY,qBAAqB,kBAAkB,WAAW,aACzE,OAAO,MAAM,KAAK,WAAW,cAC/B,KAAK,WAAW;;;KCaP;KAOA,wCAAwC,KAClD;EAGA,SAAS,YAAY;;UAGN;;;EAGf,kBAAkB;;;;;EAKlB;;EAEA;;EAEA;;EAEA;;;;EAIA;;;;EAIA;;;;;EAKA;;;;EAIA;;;EAGA;;;;EAIA,iBAAiB;;;EAGjB,aAAa;;EAEb,gBAAgB;;EAEhB,SAAS;;;;;EAKT,iBAAiB;;;;;iBAMH,sBAAsB,WAAW,kBAAkB,UACjE,SAAS,+BACR,KAAK,WAAW;;;UC/EF,mBAAmB,kBAAkB,WAAW;EAC/D,WAAW;;;EAGX;;EAEA;;;;EAIA;;EAEA;;EAEA;;;;;;;;iBASc,YAAY,WAAW,kBAAkB,UACvD,SAAS,mBAAmB,aAC3B,KAAK,WAAW"}
|
package/dist/hosted/index.d.ts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { o as InsightReport } from "../insight-report-08F022xN.js";
|
|
2
|
-
import { _ as UnixNanoTimestamp, a as hostedTenantFromEnv, c as EvalRunGenerationSnapshot, d as HostedIngestHeaders, f as HostedWireVersion, g as TraceSpanEvent, h as IngestTracesRequest, i as hostedClientFromEnv, l as EvalRunStatus, m as IngestResponse, n as HostedTenant, o as EvalRunCellScore, p as IngestEvalRunsRequest, r as createHostedClient, s as EvalRunEvent, t as HostedClient, u as HOSTED_WIRE_VERSION } from "../client-
|
|
2
|
+
import { _ as UnixNanoTimestamp, a as hostedTenantFromEnv, c as EvalRunGenerationSnapshot, d as HostedIngestHeaders, f as HostedWireVersion, g as TraceSpanEvent, h as IngestTracesRequest, i as hostedClientFromEnv, l as EvalRunStatus, m as IngestResponse, n as HostedTenant, o as EvalRunCellScore, p as IngestEvalRunsRequest, r as createHostedClient, s as EvalRunEvent, t as HostedClient, u as HOSTED_WIRE_VERSION } from "../client-Df7wdslk.js";
|
|
3
3
|
import { z } from "zod";
|
|
4
4
|
//#region src/hosted/schemas.d.ts
|
|
5
5
|
declare const UnixNanoTimestampSchema: z.ZodType<UnixNanoTimestamp>;
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import {
|
|
1
|
+
import { r as LedgerHash } from "./canonical-CFpojCN5.js";
|
|
2
2
|
//#region src/ledger-core/atomic-file-lock.d.ts
|
|
3
3
|
/** Crash-safe filesystem lock shared by campaign persistence and run exclusion. */
|
|
4
4
|
interface AtomicFileLockOwner {
|
|
@@ -39,34 +39,6 @@ declare function probeAtomicFileLock(options: AtomicFileLockOptions): AtomicFile
|
|
|
39
39
|
*/
|
|
40
40
|
declare function tryAcquireAtomicFileLock(options: AtomicFileLockOptions): AtomicFileLockAcquisition;
|
|
41
41
|
//#endregion
|
|
42
|
-
//#region src/ledger-core/canonical.d.ts
|
|
43
|
-
type LedgerHash = `sha256:${string}`;
|
|
44
|
-
/** Shape of every ledger digest: `sha256:` followed by 64 lowercase hex chars. */
|
|
45
|
-
declare const LEDGER_HASH_PATTERN: RegExp;
|
|
46
|
-
/** A value has no faithful canonical-JSON form, so it cannot be stored or hashed. */
|
|
47
|
-
declare class LedgerCanonicalizationError extends ValidationError {}
|
|
48
|
-
/** Canonical JSON encoding (RFC 8785): object keys sorted by UTF-16 code unit,
|
|
49
|
-
* recursively. This is the byte form that is hashed and stored, so equal values
|
|
50
|
-
* always encode identically. */
|
|
51
|
-
declare function canonicalString(value: unknown): string;
|
|
52
|
-
declare function hashCanonical(value: unknown): LedgerHash;
|
|
53
|
-
/**
|
|
54
|
-
* The JSON document form of `value` — the value as a JSON file carries it.
|
|
55
|
-
*
|
|
56
|
-
* A JSON document has no `undefined`, so an undefined-valued key is absent
|
|
57
|
-
* from the document and absent from its digest. Use this, and say so, when a
|
|
58
|
-
* record is digested BEFORE being written and re-digested AFTER being read
|
|
59
|
-
* back: without it the two digests disagree for a record carrying an
|
|
60
|
-
* undefined-valued key, and the reader cannot verify the writer.
|
|
61
|
-
*
|
|
62
|
-
* Only that one rule is applied. Every other non-canonical value — `NaN`, a
|
|
63
|
-
* class instance, an `undefined` array item, a cycle — reaches
|
|
64
|
-
* {@link canonicalString} unchanged and is refused there, so this form never
|
|
65
|
-
* turns an ambiguous value into a valid-looking digest the way
|
|
66
|
-
* `JSON.stringify` turns `NaN` into `null`.
|
|
67
|
-
*/
|
|
68
|
-
declare function jsonDocument(value: unknown): unknown;
|
|
69
|
-
//#endregion
|
|
70
42
|
//#region src/ledger-core/journal-file.d.ts
|
|
71
43
|
/** Filesystem durability and cross-process exclusion for append-only journal files. */
|
|
72
44
|
/** How a journal binding reports file-layer faults in its own error taxonomy. */
|
|
@@ -262,5 +234,5 @@ declare class FileLedgerJournal<Header extends object, Event extends LedgerEvent
|
|
|
262
234
|
private project;
|
|
263
235
|
}
|
|
264
236
|
//#endregion
|
|
265
|
-
export {
|
|
266
|
-
//# sourceMappingURL=index-
|
|
237
|
+
export { AtomicFileLockUnavailable as A, withLedgerFileLock as C, AtomicFileLockError as D, AtomicFileLockAcquisition as E, tryAcquireAtomicFileLock as M, AtomicFileLockOptions as O, tryWithLedgerFileLock as S, AtomicFileLock as T, trustedHeadPathFor as _, LedgerChainFields as a, LedgerFileContext as b, LedgerJournalCodec as c, LedgerReplayResult as d, LedgerAnchoredEntry as f, readTrustedHeadFile as g, LedgerTrustedHeadSubject as h, LedgerAppendResult as i, probeAtomicFileLock as j, AtomicFileLockOwner as k, LedgerLineContext as l, LedgerTrustedHeadRemoval as m, FileLedgerJournalOptions as n, LedgerEntryOf as o, LedgerTrustedHead as p, LedgerAppendOptions as r, LedgerEventBase as s, FileLedgerJournal as t, LedgerProjector as u, verifyEntriesAgainstTrustedHead as v, writeLedgerFileAtomically as w, appendLedgerLine as x, FileLockResult as y };
|
|
238
|
+
//# sourceMappingURL=index-D-UdhAmg.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index-D-UdhAmg.d.ts","names":[],"sources":["../src/ledger-core/atomic-file-lock.ts","../src/ledger-core/journal-file.ts","../src/ledger-core/trusted-head.ts","../src/ledger-core/journal.ts"],"mappings":";;;UAiBiB;WACN;WACA;WACA;;UAGM;WACN,OAAO;EAChB;;KAGU;WACG;WAA0B;WAAyB,QAAQ;;WAC3D;WAA0B;;KAE7B;WACG;WAAyB,MAAM;IAC1C;UAEa;WACN;WACA;;cAGE,4BAA4B;WACrB;;;iBASJ,oBACd,SAAS,wBACR;;;;;;;iBAea,yBACd,SAAS,wBACR;;;;;UCpDc;;EAEf;EACA,eAAe,iBAAiB;IAAY;MAAoB;;;;iBAKlD,iBAAiB,cAAc,cAAc,SAAS;;;;iBA6CtD,0BACd,cACA,kBACA,SAAS;iBAqCK,mBAAmB,GAAG,cAAc,SAAS,mBAAmB,WAAW,IAAI;KAQnF,eAAe;EAAO;EAAgB,OAAO;;EAAQ;;iBAEjD,sBAAsB,GACpC,cACA,SAAS,mBACT,WAAW,IACV,eAAe;;;;UCnED;EACf;EACA,WAAW;;;UAII;EACf;EACA,WAAW;;;;iBAKG,mBAAmB;;;;;iBAWnB,oBACd,cACA,SAAS,oBACR;;;;;;;KA2ES;EACN;EAAe,MAAM;EAAmB;;EACxC;EAAe;EAAY;EAAoB;;EAC/C;EAAgB;;;UAsCL;;EAEf;;;EAGA;;;;;iBAMc,gCACd,kBAAkB,uBAClB,MAAM,mBACN,SAAS,mBACT,QAAQ;;;;UCjLO;EACf;;;;;UAMe,kBAAkB,cAAc;EAC/C;EACA,cAAc;EACd,OAAO;EACP,WAAW;;KAGD,cAAc,uBAAuB,cAAc,mBAAmB,SAChF,kBAAkB;UAEH;EACf;EACA;;;;;UAMe,gBAAgB,OAAO;EACtC,MAAM,OAAO,OAAO;EACpB,OAAO,SAAS,UAAU;;UAGX,mBACf,uBACA,cAAc,iBACd,oBACQ;;EAER,QAAQ;EACR,cAAc,kBAAkB;;;;EAIhC,WAAW,cAAc,SAAS,oBAAoB,cAAc,QAAQ;;EAE5E,iBAAiB,OAAO,cAAc,QAAQ,QAAQ;EACtD,mBAAmB,gBAAgB,cAAc,QAAQ,QAAQ;;UAGlD,mBAAmB,OAAO;EACzC,SAAS;EACT,YAAY;;UAGG,mBAAmB,OAAO;EACzC,OAAO;;EAEP;EACA,YAAY;;UAGG;;;;;;;;;;;;;;EAcf;;UAGe;;;;;;;EAOf;;;;cAiBW,kBAAkB,uBAAuB,cAAc,iBAAiB;WAC1E;;WAEA;mBACQ;mBACA;mBACA;EAEjB,YACE,cACA,OAAO,mBAAmB,QAAQ,OAAO,aACzC,UAAS;EASL,UAAU,QAAQ,mBAAmB,cAAc,QAAQ,QAAQ;EAMnE,OACJ,OAAO,OACP,UAAS,sBACR,QAAQ,mBAAmB,cAAc,QAAQ,QAAQ;;EA8CtD,eAAe,QAAQ;;;;EAWvB,kBAAkB,QAAQ;;;;;;;;EA4B1B,oBAAoB,QAAQ;;;;;;;;;;;;;UAoB1B;UAiBA;UAUA;UAgBA;UAgDA"}
|
|
@@ -3,10 +3,10 @@ import { i as AnalystFinding, n as AnalystContext } from "./types-DMoNFDWi.js";
|
|
|
3
3
|
import { p as ChatClient } from "./types-Bfk0uxRj.js";
|
|
4
4
|
import { f as ExternalOptimizerModelExecutionObservation, h as ExternalOptimizerRunnerCommand, l as ExternalOptimizerModelCall, r as ExternalOptimizerCallbackLimits } from "./external-optimizer-contracts-szBJ_1vh.js";
|
|
5
5
|
import { o as Severity } from "./multi-layer-verifier-BUaQ4C17.js";
|
|
6
|
-
import {
|
|
7
|
-
import { i as TraceAnalystDefinition } from "./default-registry-B0s2zU-s.js";
|
|
6
|
+
import { i as TraceAnalystDefinition } from "./default-registry-XxedTLwu.js";
|
|
8
7
|
import { l as ExactCapableAnalyst } from "./exact-types-BEecmnWm.js";
|
|
9
|
-
import { t as TraceAnalysisEngine } from "./engine-
|
|
8
|
+
import { t as TraceAnalysisEngine } from "./engine-BfRay1qD.js";
|
|
9
|
+
import { S as SupervisorRunTree, x as SupervisorRunSources } from "./types-CoPUTiXb.js";
|
|
10
10
|
//#region src/semantic-concept-judge.d.ts
|
|
11
11
|
/**
|
|
12
12
|
* Implementation complexity class for weighted scoring.
|
|
@@ -440,4 +440,4 @@ declare const KNOWLEDGE_POISONING_KIND_SPEC: TraceAnalystDefinition;
|
|
|
440
440
|
declare const DEFAULT_TRACE_ANALYST_KINDS: readonly TraceAnalystDefinition[];
|
|
441
441
|
//#endregion
|
|
442
442
|
export { SemanticConceptJudgeResult as A, DspyRlmTraceEngineOptions as C, SEMANTIC_CONCEPT_JUDGE_VERSION as D, ConceptSpec as E, SemanticConceptJudgeInput as O, renderFindingSubject as S, ConceptFinding as T, FindingSubject as _, FAILURE_MODE_KIND_SPEC as a, findingSubjectGrammarPromptFor as b, emitControlIntegrityFindings as c, FindingsStore as d, PersistedFinding as f, FINDING_SUBJECT_SYNTAX as g, FINDING_SUBJECT_KINDS as h, IMPROVEMENT_KIND_SPEC as i, runSemanticConceptJudge as j, SemanticConceptJudgeOptions as k, DiffPolicy as l, diffFindings as m, KNOWLEDGE_POISONING_KIND_SPEC as n, CONTROL_INTEGRITY_ANALYST as o, defaultIsMaterial as p, KNOWLEDGE_GAP_KIND_SPEC as r, ControlIntegrityAnalyst as s, DEFAULT_TRACE_ANALYST_KINDS as t, FindingsDiff as u, FindingSubjectKind as v, createDspyRlmTraceEngine as w, parseFindingSubject as x, KIND_EXPECTED_SUBJECTS as y };
|
|
443
|
-
//# sourceMappingURL=index-
|
|
443
|
+
//# sourceMappingURL=index-DDAPhUJJ.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-DDAPhUJJ.d.ts","names":[],"sources":["../src/semantic-concept-judge.ts","../src/analyst/trace-tool-callback.ts","../src/analyst/dspy-rlm-engine.ts","../src/analyst/finding-subject.ts","../src/analyst/findings-store.ts","../src/analyst/kinds/control-integrity.ts","../src/analyst/kinds/failure-mode.ts","../src/analyst/kinds/improvement.ts","../src/analyst/kinds/knowledge-gap.ts","../src/analyst/kinds/knowledge-poisoning.ts","../src/analyst/kinds/index.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KAuDY;UAEK;EACf;;EAEA;;EAEA;;EAEA,aAAa;;UAGE;EACf;EACA;;EAEA;EACA;EACA,UAAU;;UAGK;;EAEf;;EAEA;;EAEA,aAAa;IAAQ;IAAc;;;EAEnC,kBAAkB;;EAElB;EACA;;UAGe;EACf;EACA;;EAEA;EACA;EACA;EACA,UAAU;EACV;EACA;EACA;;EAEA;EACA;;;;;;;;KASU;UAQK;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA,MAAM;;EAEN,UAAU;EACV,aAAa;EACb;EACA,WAAW;EACX,SAAS;;;;;;EAMT,iBAAiB;;EAEjB,oBAAoB,QAAQ,OAAO;;cAKxB;;;;;;iBAkGS,wBACpB,OAAO,2BACP,SAAS,8BACR,QAAQ;;;KCzOC,0BAA0B;;;UCerB;;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;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;KCzClE;EAEN;EAAwB;EAAc;;EACtC;EAAyB;;EACzB;EAAuB;;EACvB;EAAyB;;EAGzB;EAAuB;;EACvB;EAAe;;EACf;EAAkB;EAAc;;EAChC;EAAkB;;EAClB;EAAa;EAAgB;;EAC7B;EAAc;;EACd;EAAkB;;EAClB;EAAkB;;EAClB;EAAwB;;EACxB;EAAuB;;EACvB;EAAc;;EACd;EAAa;EAAgB;;EAC7B;EAAgB;;EAChB;EAAqB;;EACrB;EAAuB;;EAEvB;EAA4B;;EAC5B;EAA2B;;EAE3B;EAAiB;;KAEX,qBAAqB;cAEpB,uBAAuB,cAAc;;;;;;;;;;;;;;;;;iBAgGlC,oBAAoB,iCAAiC;;;;;;;iBAoHrD,qBAAqB,GAAG;;;;;;;;;;cA8D3B,wBAAwB,SAAS,OAAO;;;;;;;;;cAoExC,wBAAwB,eAAe,cAAc;;iBAqElD,+BAA+B;;;;;;;;UCxc9B,yBAAyB;EACxC;;cAGW;WAGiB;mBAFX;EAEjB,YAA4B;EAItB,OAAO,eAAe,UAAU,mBAAmB;;EAQzD,WAAW;;EAkBX,QAAQ,gBAAgB;;UAOT;;EAEf,UAAU;;EAEV,aAAa;;EAEb,WAAW;;;;;EAKX,SAAS;IAAQ,UAAU;IAAkB,SAAS;;;UAGvC;;;;;;;;EAQf,cAAc,UAAU,gBAAgB,SAAS;;;;;;iBAOnC,kBAAkB,GAAG,gBAAgB,GAAG;;;;;iBAWxC,aACd,UAAU,oBACV,SAAS,oBACT,SAAQ,aACP;;;;iBC7Fa,6BACd,OAAO,uBAAuB,mBAC9B,qBACC;;cA4BU,mCACA,oBAAoB,uBAAuB;WAE7C;WACA;WAEA;WACA;IAAS;IAAgC;;WACzC;WACA;aACP;aACA;;EAGI,QACJ,OAAO,uBAAuB,mBAC9B,KAAK,iBACJ,QAAQ;;cAUA,2BAAyB;;;cCHzB,wBAAwB;;;cC3BxB,uBAAuB;;;cCGvB,yBAAyB;;;cCPzB,+BAA+B;;;;;;;;;;;;cCP/B,sCAAsC"}
|
|
@@ -1,15 +1,15 @@
|
|
|
1
1
|
import { c as CostLedgerHandle } from "./cost-ledger-DbQdN3nO.js";
|
|
2
2
|
import { a as RunRecord, n as RunCostProvenance, s as RunSplitTag } from "./run-record-DQjRcYwA.js";
|
|
3
3
|
import "./types-DMoNFDWi.js";
|
|
4
|
-
import "./provenance-
|
|
5
|
-
import { _ as GateDecision } from "./types-
|
|
6
|
-
import "./heldout-gate-
|
|
7
|
-
import "./promotion-policy-
|
|
4
|
+
import "./provenance-Dp-vvyrU.js";
|
|
5
|
+
import { _ as GateDecision } from "./types-nokrtr7M.js";
|
|
6
|
+
import "./heldout-gate-JgNRDZwZ.js";
|
|
7
|
+
import "./promotion-policy-BBBcz5_3.js";
|
|
8
8
|
import { o as InsightReport } from "./insight-report-08F022xN.js";
|
|
9
|
-
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, o as EvalRunCellScore, s as EvalRunEvent } from "./client-
|
|
10
|
-
import "./define-agent-eval-
|
|
11
|
-
import "./default-registry-
|
|
12
|
-
import { _ as AnalyzeRunsOptions } from "./engine-
|
|
9
|
+
import { c as EvalRunGenerationSnapshot, g as TraceSpanEvent, o as EvalRunCellScore, s as EvalRunEvent } from "./client-Df7wdslk.js";
|
|
10
|
+
import "./define-agent-eval-0wW7gFhr.js";
|
|
11
|
+
import "./default-registry-XxedTLwu.js";
|
|
12
|
+
import { _ as AnalyzeRunsOptions } from "./engine-BfRay1qD.js";
|
|
13
13
|
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";
|
|
14
14
|
//#region src/contract/measured-comparison.d.ts
|
|
15
15
|
interface SealCandidateBenchmarkSuiteOptions {
|
|
@@ -701,4 +701,4 @@ interface FromOtelSpansOptions {
|
|
|
701
701
|
declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
|
|
702
702
|
//#endregion
|
|
703
703
|
export { FromRunRecordDirOptions as $, observeCodeAgentSession as A, verifyCandidateExperimentComparison as At, parseAgentTrace as B, CodeAgentSessionActionKind as C, evaluatePairedMeasurements as Ct, CodeAgentSessionObservation as D, sealCandidateBenchmarkTask as Dt, CodeAgentSessionExecutionReceipt as E, sealCandidateBenchmarkSuite as Et, AgentTraceIndex as F, EvalRunDiff as G, EvalCellScoreDelta as H, AgentTraceRange as I, diffRuns as J, diffGenerations as K, AgentTraceRecord as L, AgentTraceContributorType as M, AgentTraceConversation as N, CodeAgentSessionSource as O, sealCandidateExperiment as Ot, AgentTraceFile as P, evalReportingSuite as Q, AuthoringProvenance as R, CodeAgentSessionAction as S, SealCandidateBenchmarkSuiteOptions as St, CodeAgentSessionActionSurface as T, runCandidateExperiment as Tt, EvalDimensionDelta as U, partitionRunsByAuthoringModel as V, EvalGenerationDiff as W, EvalReportingSuiteOptions as X, EvalReportingSuiteInput as Y, EvalReportingSuiteResult as Z, fromOpenCodeSession as _, EvaluatePairedMeasurementsOptions as _t, FromFeedbackTableOptions as a, CompareAgentProfileImprovementExperimentOptions as at, parseCodeAgentJsonlFile as b, PairedMeasurementEvaluation as bt, CodeAgentJsonlLine as c, measuredComparisonFromAgentProfileImprovementExperiment as ct, CodeAgentSessionIntakeResult as d, sealAgentProfileImprovementSuite as dt, FromRunRecordDirResult as et, CodeAgentSessionMetrics as f, sealAgentProfileImprovementTask as ft, fromKimiCodeSession as g, CompareCandidateExperimentOptions as gt, fromCodexSession as h, CandidateExperimentRun as ht, FeedbackTableRow as i, AgentProfileImprovementExperimentRun as it, AgentTraceContributor as j, CodeAgentSessionTerminalStatus as k, verifyCandidateExperiment as kt, CodeAgentSessionDiagnostic as l, runAgentProfileImprovementExperiment as lt, fromClaudeCodeSession as m, CandidateExperimentExecutionInput as mt, fromOtelSpans as n, fromRunRecordDir as nt, FromFeedbackTableResult as o, RunAgentProfileImprovementExperimentOptions as ot, ParsedCodeAgentJsonl as p, verifyAgentProfileImprovementExperimentComparison as pt, diffRunBaselineToWinner as q, FeedbackTableMeta as r, AgentProfileImprovementExperimentExecutionInput as rt, fromFeedbackTable as s, SealAgentProfileImprovementSuiteOptions as st, FromOtelSpansOptions as t, RunRecordRejection as tt, CodeAgentSessionIntakeOptions as u, sealAgentProfileImprovementExperiment as ut, fromPiSession as v, PairedMeasurement as vt, CodeAgentSessionActionStatus as w, measuredComparisonFromCandidateExperiment as wt, streamCodeAgentJsonlFile as x, RunCandidateExperimentOptions as xt, parseCodeAgentJsonl as y, PairedMeasurementAdapter as yt, PartitionByAuthoringModelResult as z };
|
|
704
|
-
//# sourceMappingURL=index-
|
|
704
|
+
//# sourceMappingURL=index-DnglhM0A.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-DnglhM0A.d.ts","names":[],"sources":["../src/contract/measured-comparison.ts","../src/contract/profile-measured-comparison.ts","../src/contract/intake/run-record-dir.ts","../src/contract/eval-reporting-suite.ts","../src/contract/diff.ts","../src/contract/intake/agent-trace.ts","../src/contract/intake/code-agent-observation.ts","../src/contract/intake/code-agent-session.ts","../src/contract/intake/feedback-table.ts","../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;;;;UAiCiB;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,eAAe,KAAK,OAAO;EAC3B,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA,kBAAkB;;EAElB,kBAAkB;;;KAIR,8BAA8B,KACxC;EAGA,iBAAiB;EACjB,WAAW;EACX;;;iBAIc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;;;;;;;iBAkEK,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;;iBA8Ya,0CACd,SAAS,oCACR;;iBAuEa,oCACd,iBACC;;;UCjrBc;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;;EAEA,aAAa;EACb,SAAS;;UAGM;EACf,cAAc;EACd;IACE;IACA,MAAM;;;UAIO;EACf,YAAY;EACZ,cAAc;EACd;IACE;IACA,MAAM;;EAER,aAAa;EACb;EACA,YAAY;EACZ;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;;;;;;iBAsBmB,qCACpB,SAAS,8CACR,QAAQ;;iBA0CK,wDACd,SAAS,kDACR;;iBAuEa,kDACd,iBACC;;;;UC7Oc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;;;;;;;iBAkBoB,iBACpB,cACA,UAAS,0BACR,QAAQ;;;;;KC7CC,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;;;;;;iBAUoB,mBACpB,OAAO,yBACP,UAAS,4BACR,QAAQ;;;;;;UC5DM;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;EAGA,YAAY,eAAe,eAAe;;;;UAK3B;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;EAET,SAAS;;EAET,OAAO;;EAEP;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;UAMe;EACf;EACA;EACA;EACA;EACA,oBAAoB;EACpB,mBAAmB;EACnB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA,cAAc;;;;EAId,aAAa;;;;;;;;;iBAoDC,gBACd,QAAQ,2BACR,OAAO,4BACN;;;;;;iBAqEa,SAAS,QAAQ,cAAc,OAAO,eAAe;;;;;;;iBAsCrD,wBAAwB,KAAK,eAAe;;;KC1OhD;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;KCjLS;KAEA;KAEA;KAEA;KASA;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,MAAM;EACN,SAAS;EACT;EACA,QAAQ;EACR;EACA;EACA,UAAU;;UAGK;EACf,QAAQ;EACR;EACA;EACA;IACE,QAAQ;IACR;;EAEF,SAAS;;UAGM;EACf,QAAQ;EACR;EACA;EACA,YAAY;;;;;;;iBAeE,wBACd,SAAS,iCACR;;;UCjCc;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,QAAQ;EACR;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;;UAGe;EACf,MAAM;EACN,aAAa;EACb,SAAS;EACT,cAAc;;UAGC;EACf;EACA;EACA;EACA;EACA;EACA;EACA,WAAW;EACX;EACA;EACA;EACA;EACA;EACA;;;EAGA,iBAAiB;;;EAGjB,YAAY;;;;KAKF;EACN;EAAe;EAAoB;;EACnC;EAAmB;;iBAcT,oBAAoB,gBAAgB;;;;;;;;;;iBAuB7B,yBAAyB,eAAe,eAAe;;;;;iBA6BxD,wBAAwB,eAAe,QAAQ;iBAUrD,iBACd,SAAS,gCACR;iBAIa,sBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,oBACd,SAAS,gCACR;iBAIa,cACd,SAAS,gCACR;;;UCrNc;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;;;UCvBlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}
|
|
@@ -2,16 +2,16 @@ import { c as ValidationError, t as AgentEvalError } from "./errors-DEE6u6ot.js"
|
|
|
2
2
|
import { T as RunPaidCallInput, p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
|
|
3
3
|
import { a as RunRecord, s as RunSplitTag } from "./run-record-DQjRcYwA.js";
|
|
4
4
|
import { f as AnalystUsageReceipt, i as AnalystFinding, l as AnalystRunResult, w as TraceAnalysisStore } from "./types-DMoNFDWi.js";
|
|
5
|
-
import { Fn as CampaignRunPlan,
|
|
6
|
-
import { A as LabeledScenarioWrite, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, p as Gate, u as ComponentSurface } from "./types-JHMOqZI4.js";
|
|
7
|
-
import "./heldout-gate-Dh2b62w8.js";
|
|
5
|
+
import { Fn as CampaignRunPlan, Gn as CampaignStorage, Ln as PlanCampaignRunOptions } from "./provenance-Dp-vvyrU.js";
|
|
8
6
|
import { _ as PairedArmsComparison } from "./pre-registration-DHz6P_6f.js";
|
|
9
7
|
import { d as PairedBootstrapResult } from "./paired-promotion-decision-CGzg0cI_.js";
|
|
10
|
-
import "./
|
|
8
|
+
import { A as LabeledScenarioWrite, D as LabeledScenarioSampleArgs, E as LabeledScenarioRecord, O as LabeledScenarioSource, R as Scenario, S as JudgeConfig, T as LabelTrust, a as CampaignResult, d as DispatchContext, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, p as Gate, u as ComponentSurface } from "./types-nokrtr7M.js";
|
|
9
|
+
import "./heldout-gate-JgNRDZwZ.js";
|
|
10
|
+
import "./promotion-policy-BBBcz5_3.js";
|
|
11
11
|
import { d as CompletionVerdict, f as CorrectnessChecker, h as ProducedState, n as BackendIntegrityReport, s as RuntimeEventLike, u as CompletionRequirement } from "./backend-integrity-e79K3UPD.js";
|
|
12
12
|
import { _ as AnalystIssueExpectation, a as AnalystBenchmarkLabelState, t as AnalystBenchmarkCase } from "./benchmark-h-h4bfqj.js";
|
|
13
13
|
import { t as AgentProfile$1 } from "./agent-profile-B9_GGsG8.js";
|
|
14
|
-
import "./statistical-heldout-
|
|
14
|
+
import "./statistical-heldout-Yldkntvy.js";
|
|
15
15
|
import { AgentProfile } from "@tangle-network/agent-interface";
|
|
16
16
|
//#region src/campaign/search-ledger-errors.d.ts
|
|
17
17
|
/** Base error for invalid search-ledger input or operations. */
|
|
@@ -1132,4 +1132,4 @@ declare function verifyCodeSurface(surface: CodeSurface, worktreeDir?: string):
|
|
|
1132
1132
|
declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
|
|
1133
1133
|
//#endregion
|
|
1134
1134
|
export { makePlaybackDispatch as $, CrossSurfaceEvidenceBreakdown as $t, scoreDiscrimination as A, LoadEvalFixtureScenariosOptions as At, ProfileMatrixSegmentResult as B, CrossSurfaceAttemptCompleteness as Bt, acquireSingleRunLock as C, SearchLedgerIntegrityError as Cn, neutralizationGate as Ct, compareRankKeys as D, EvalFixtureRunPlan as Dt, campaignMeanComposite as E, EvalFixtureLoadOptions as Et, FinalizeProfileMatrixOptions as F, planEvalFixtureRun as Ft, PlaybackContext as G, CrossSurfaceCandidateEvidence as Gt, createProfileMatrixPlan as H, CrossSurfaceBootstrapPolicy as Ht, FinalizedProfileMatrixResult as I, analyzeCrossSurfaceInteractions as It, ScoreboardRenderOptions as J, CrossSurfaceComponent as Jt, PlaybackDriver as K, CrossSurfaceCandidateOutcome as Kt, ProfileMatrixCoverage as L, AnalyzeCrossSurfaceInteractionsInput as Lt, resolveRunDir as M, discoverEvalFixtures as Mt, tangleTracesRoot as N, loadEvalFixture as Nt, DiscriminationScore as O, EvalFixtureScenario as Ot, CreateProfileMatrixPlanOptions as P, loadEvalFixtureScenarios as Pt, UserStoryVerdict as Q, CrossSurfaceEligibility as Qt, ProfileMatrixPlan as R, CrossSurfaceAdditionDecision as Rt, SingleRunLockOptions as S, SearchLedgerError as Sn, NeutralizationGateOptions as St, campaignBreakdown as T, EvalFixtureFile as Tt, finalizeProfileMatrix as U, CrossSurfaceCandidate as Ut, RunProfileMatrixSegmentOptions as V, CrossSurfaceBestSingleSelection as Vt, runProfileMatrixSegment as W, CrossSurfaceCandidateComparison as Wt, ScoreboardSummary as X, CrossSurfaceCompositionStep as Xt, ScoreboardRow as Y, CrossSurfaceComponentEvidence as Yt, UserStory as Z, CrossSurfaceDistribution as Zt, componentSurfaceIdentityMaterial as _, TraceAnalystArtifact as _n, RolloutCall as _t, WorktreeAdapterError as a, CrossSurfaceInteractionTask as an, ProfileMatrixError as at, surfaceHash as b, traceAnalystQualityJudge as bn, classifyUngroundedLiterals as bt, verifyCodeSurface as c, CrossSurfacePairEvidence as cn, RunProfileMatrixResult as ct, PhoenixEvaluationResultLike as d, CrossSurfaceRankedSingle as dn, neutralizeText as dt, CrossSurfaceIneligibilityReason as en, renderScoreboardMarkdown as et, PhoenixEvaluatorLike as f, CrossSurfaceRelativeCost as fn, FsLabeledScenarioStore as ft, codeSurfaceIdentityMaterial as g, BuildTraceAnalystSurfaceDispatchOptions as gn, RolloutArgumentDiffOptions as gt, assertCodeSurfaceIdentity as h, CrossSurfaceTaskRow as hn, RolloutArgumentDiff as ht, WorktreeAdapter as i, CrossSurfaceInteractionReport as in, ProfileDispatchFn as it, selectDiscriminative as j, PlanEvalFixtureRunOptions as jt, ScenarioSignal as k, EvalFixtureValidationMode as kt, AutoevalsScoreLike as l, CrossSurfacePairIncompatibilityReason as ln, ScenarioRollup as lt, phoenixEvaluatorJudge as m, CrossSurfaceSelections as mn, LabeledScenarioStoreError as mt, GitWorktreeAdapterOptions as n, CrossSurfaceInteractionEffect as nn, scoreboardSummary as nt, gitWorktreeAdapter as o, CrossSurfaceNaiveStackSelection as on, ProfileSummary as ot, autoevalsScorerJudge as p, CrossSurfaceSelectionPolicy as pn, FsLabeledScenarioStoreOptions as pt, PlaybackStep as q, CrossSurfaceCandidateSummary as qt, Worktree as r, CrossSurfaceInteractionPath as rn, userStoryScoreboard as rt, resolveWorktreePath as s, CrossSurfacePairCompatibility as sn, RunProfileMatrixOptions as st, CodeSurfaceVerification as t, CrossSurfaceInteractionAwareSelection as tn, scoreUserStory as tt, AutoevalsScorerLike as u, CrossSurfacePairwiseEntry as un, runProfileMatrix as ut, renderSurfaceDiff as v, TraceAnalystScenario as vn, ScoredRollout as vt, CampaignBreakdown as w, EvalFixture as wt, SingleRunLock as x, SearchLedgerConflictError as xn, rolloutArgumentDiff as xt, surfaceContentHash as y, buildTraceAnalystSurfaceDispatch as yn, UngroundedLiteralReport as yt, ProfileMatrixRow as z, CrossSurfaceAdditionRejectionReason as zt };
|
|
1135
|
-
//# sourceMappingURL=index-
|
|
1135
|
+
//# sourceMappingURL=index-_vPrVMRX.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index-
|
|
1
|
+
{"version":3,"file":"index-_vPrVMRX.d.ts","names":[],"sources":["../src/campaign/search-ledger-errors.ts","../src/campaign/analyst-surface.ts","../src/campaign/cross-surface-types.ts","../src/campaign/cross-surface-interaction.ts","../src/campaign/fixtures.ts","../src/campaign/gates/neutralization-gate.ts","../src/campaign/grounded-reflection.ts","../src/campaign/labeled-store/fs-adapter.ts","../src/campaign/neutralize.ts","../src/campaign/presets/run-profile-matrix.ts","../src/campaign/presets/playback.ts","../src/campaign/presets/segmented-profile-matrix.ts","../src/campaign/run-dir.ts","../src/campaign/scenario-selection.ts","../src/campaign/score-utils.ts","../src/campaign/single-run-lock.ts","../src/campaign/surface-identity.ts","../src/campaign/upstream-evaluators.ts","../src/campaign/worktree/index.ts"],"mappings":";;;;;;;;;;;;;;;;;cAGa,0BAA0B;;cAG1B,mCAAmC;;cAGnC,kCAAkC;;;UCE9B,6BAA6B;EAC5C;EACA,YAAY;EACZ,YAAY;EACZ,yBAAyB;EACzB,kBAAkB;;UAGH;EACf,mBAAmB;EACnB,QAAQ;EACR,MAAM;;UAGS;EACf,QAAQ;IACN;IACA,YAAY;IACZ;IACA,QAAQ;MACN,QAAQ;;iBAGE,iCACd,SAAS,2CAET,SAAS,gBACT,UAAU,sBACV,SAAS,oBACN,QAAQ;iBAgBG,4BAA4B,YAC1C,sBACA;;;;KCtDU;;UAGK;EACf;EACA;;EAEA;;;UAIe;EACf;EACA;EACA;EACA;;;UAIe;EACf;;EAEA;;EAEA;;;;;;;UAQe;EACf;EACA;;EAEA;EACA,cAAc;EACd;EACA;;;;;;EAMA,MAAM;EACN,mBAAmB;;EAEnB;;UAGe;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;;EAEA,kCAAkC;;EAElC;;UAGe,qCACf,aAAa,sBAAsB;EAEnC,qBAAqB;EACrB,qBAAqB;EACrB,eAAe;EACf;;EAEA;;EAEA;;EAEA;;EAEA;EACA,WAAW;EACX,WAAW;;UAGI;EACf;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;;UAGe;EACf,aAAa;EACb;EACA;EACA;EACA;;KAGU;UAWK;EACf;EACA,SAAS;;UAGM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf,WAAW;EACX,SAAS;EACT,OAAO;EACP,OAAO,eAAe;EACtB,QAAQ;EACR,QAAQ;;EAER,sBAAsB;;EAEtB,aAAa;;UAGE;EACf;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA,QAAQ;EACR,cAAc,eAAe;;UAGd;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;;EAEA;EACA;;UAGe;EACf,SAAS;EACT;EACA;EACA;EACA;EACA,eAAe;EACf,gBAAgB;;KAGN;UAYK;EACf;EACA,SAAS;EACT;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,4BAA4B,iCAAiC;EAC7D,wBAAwB,eAAe;EACvC,QAAQ;EACR,QAAQ;EACR,aAAa;EACb,eAAe;;UAGA;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA,SAAS;;UAGM;;EAEf;EACA;;KAGU;UAaK;EACf;EACA;EACA;EACA;EACA;EACA,uBAAuB;EACvB;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,YAAY;EACZ;;;UAIe;EACf;EACA;EACA;EACA;EACA,OAAO;;UAGQ;;EAEf;;EAEA;EACA;;EAEA;EACA;;EAEA,gBAAgB;;EAEhB,OAAO;;UAGQ;EACf,YAAY;EACZ,YAAY;EACZ,kBAAkB;;UAGH,8BACf,aAAa,sBAAsB;EAEnC;EACA;EACA;EACA;;EAEA,MAAM;EACN,iBAAiB;EACjB,iBAAiB;EACjB,YAAY;EACZ,UAAU;EACV,YAAY;;;;;;;;;iBCnRE,gCAAgC,aAAa,qBAC3D,OAAO,qCAAqC,QAC3C,8BAA8B;;;KC7CrB;UAEK;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;EACA;EACA,OAAO;EACP;;UAGe,4BAA4B;EAC3C;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;;EAEf,aAAa;;EAEb;;UAGe,wCAAwC;EACvD;;UAGe,0BAA0B,6BACjC,KACN,uBAAuB,qBAAqB;EAG9C;EACA,aAAa;EACb;EACA;EACA,UAAU;;KAGA,qBAAqB;EAC/B,UAAU,MAAM,KAAK;;;iBAIP,qBAAqB;;;;;iBA4BrB,gBACd,kBACA,cACA,UAAS,yBACR;;iBA4Ca,yBACd,kBACA,UAAS,kCACR;;;;;iBAuBa,mBAAmB,qBACjC,SAAS,0BAA0B,aAClC;;;UChJc,0BAA0B,kBAAkB,WAAW;EACtE,WAAW;;;;;EAKX;;;;;;iBAuBc,mBAAmB,WAAW,kBAAkB,UAC9D,SAAS,0BAA0B,aAClC,KAAK,WAAW;;;;;;;;;;;;;;;;;;;;;;;;;;;;;UC9BF;WACN;WACA,MAAM,SAAS;;;UAIT;;WAEN;;WAEA;WACA,gBAAgB;;UAGV;;WAEN;;WAEA;;UAGM;;WAEN;;WAEA,eAAe;;WAEf,eAAe;;;;;;;;iBASV,oBACd,mBAAmB,iBACnB,OAAM,6BACL;UAsCc;;WAEN;;WAEA;;;;;;;;;iBAUK,2BACd,cACA,MAAM,KAAK,0DACV;;;UChFc;;EAEf;;;EAGA;;EAEA;;;cAIW,kCAAkC;WAE3B;EADlB,YACkB,cAChB;;;;;;cAiBS,kCAAkC;mBAIhB;mBAHZ;mBACA;EAEjB,YAA6B,SAAS;EAKhC,QAAQ,OAAO,uBAAuB;EAYtC,OAAO,MAAM,4BAA4B,QAAQ;EA8CjD,QAAQ;IACZ;IACA;IACA,UAAU;IACV,SAAS,OAAO;;UAkCV;UAiCA;UAmBA;UAkBA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBCvNM,eAAe;;;;;;cC8ClB,2BAA2B;EACtC,YAAY;;;;;KAQF,kBAAkB,kBAAkB,UAAU,cACxD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;UAEI,wBAAwB,kBAAkB,UAAU;;EAEnE,UAAU;;EAEV,WAAW;;EAEX,UAAU,kBAAkB,WAAW;;EAEvC,SAAS,YAAY,WAAW;;;EAGhC;;;EAGA;;;EAGA;;;;EAIA;;EAEA,WAAW;;EAEX;;EAEA;;;;;;;;EAQA;;;EAGA;;EAEA;;;EAGA;;EAEA;;EAEA,eAAe;EACf,gBAAgB;;;EAGhB,UAAU;;EAEV,YAAY;;;EAGZ,aAAa,UAAU;;;;EAIvB;;;;;;;EAOA,cACE,UAAU,WACV,UAAU;IACL;IAAgB;;;;;;;EAMvB,aAAa;IAAS,SAAS;IAAc,UAAU;IAAW;;;EAElE;;EAEA;;UAGe;EACf;EACA;EACA;;EAEA;;EAEA;;EAEA;EACA,gBAAgB;;;EAGhB,WAAW;;UAGI;EACf;EACA;;UAGe,uBAAuB,WAAW,kBAAkB;EACnE;EACA;;;;EAIA,SAAS;EACT,WAAW,eAAe;EAC1B,YAAY,eAAe;;EAE3B,YAAY,eAAe;;EAE3B,WAAW;;EAEX,WAAW,eAAe,eAAe,WAAW;;;;;iBAoOhC,iBAAiB,kBAAkB,UAAU,WACjE,MAAM,wBAAwB,WAAW,aACxC,QAAQ,uBAAuB,WAAW;;;;;UC9Z5B;;EAEf;;EAEA,UAAU;;;;;;;UAQK,kBAAkB;;EAEjC;;EAEA,OAAO;;EAEP,cAAc;;;UAIC,wBAAwB;EACvC,SAAS;;;;;;;;;;;UAYM,eAAe,eAAe,YAAY;EACzD,IAAI,OAAO,QAAQ,KAAK,kBAAkB,iBAAiB;;;;;;;iBAQ7C,qBAAqB,eAAe,WAClD,QAAQ,eAAe,UACtB,kBAAkB,QAAQ;;UAQZ,yBAAyB;EACxC;;;;;;;;iBASoB,eACpB,OAAO,WACP,OAAO,eACP,kBAAkB,qBACjB,QAAQ;;UAUM;EACf;EACA;EACA;EACA;EACA;EACA;;;;;;;iBAQc,oBAAoB,mBAAmB,qBAAqB;;UAkB3D;;EAEf;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;iBAIc,kBAAkB,eAAe,kBAAkB;UAwBlD;;EAEf;;EAEA,OAAO;;EAEP;;;;;;;;;iBAkBc,yBACd,eAAe,iBACf,OAAM;;;UCpKS;EACf;EACA;EACA;EACA;EACA;;UAGe,kBAAkB,kBAAkB,UAAU;WACpD;WACA;WACA;WACA;WACA,mBAAmB;WACnB,oBAAoB;WACpB,iBAAiB,YAAY,WAAW;WACxC;WACA;WACA,UAAU,YAAY,wBAAwB,WAAW;WACzD;;WAEA;WACA,WAAW,YAAY,wBAAwB,WAAW;WAC1D;WACA;WACA,aAAa,UAAU;WACvB,cACP,UAAU,WACV,UAAU;IACL;IAAgB;;WACd,eAAe;;UAGT,+BAA+B,kBAAkB,UAAU,mBAClE,KACN,wBAAwB,WAAW;;EAgBrC;;UAGe;EACf;EACA;EACA;EACA;EACA;;UAGe,+BAA+B,kBAAkB,UAAU;EAC1E,MAAM,kBAAkB,WAAW;;EAEnC;;EAEA,yBAAyB;EACzB,UAAU,kBAAkB,WAAW;EACvC;EACA,UAAU;EACV;EACA;EACA;EACA,eAAe,wBAAwB,WAAW;EAClD,gBAAgB,wBAAwB,WAAW;EACnD,YAAY;;UAGG,2BAA2B,WAAW,kBAAkB;EACvE;EACA;EACA,QAAQ,uBAAuB,WAAW;EAC1C,UAAU;;UAGK,6BAA6B,kBAAkB,UAAU;EACxE,MAAM,kBAAkB,WAAW;EACnC;EACA,UAAU;EACV;EACA;EACA,YAAY;;UAGG,6BAA6B,WAAW,kBAAkB,kBACjE,uBAAuB,WAAW;EAC1C,UAAU;;iBAGI,wBAAwB,kBAAkB,UAAU,WAClE,MAAM,+BAA+B,WAAW,aAC/C,kBAAkB,WAAW;iBA0HV,wBAAwB,kBAAkB,UAAU,WACxE,MAAM,+BAA+B,WAAW,aAC/C,QAAQ,2BAA2B,WAAW;iBAkD3B,sBAAsB,kBAAkB,UAAU,WACtE,MAAM,6BAA6B,WAAW,aAC7C,QAAQ,6BAA6B,WAAW;;;;;;;;iBCrTnC;;;;;iBAQA,cAAc,gBAAgB;;;;;;;;;;;;;;;;;UCD7B;EACf;;EAEA;;UAGe;EACf;;EAEA;;EAEA;EACA;;EAEA;;;;;;;;;;iBA4Cc,oBACd,SAAS,kBACT;EAAS;IACR;;;;;;;;;;;iBA0Ba,qBACd,SAAS,kBACT,WACA;EAAS;;;;;;;iBC3FK,sBAAsB,WAAW,kBAAkB,UACjE,UAAU,eAAe,WAAW;;;;iBA0CtB,gBAAgB,sBAAsB;UAWrC;;EAEf,YAAY;;;;;EAKZ,WAAW;IAAQ;IAAoB;IAAmB;IAAgB;;;;;iBAK5D,kBAAkB,WAAW,kBAAkB,UAC7D,UAAU,eAAe,WAAW,aACnC;;;;;;;;;;;;;;;;;;UC5Dc;;WAEN;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf;;;;;;;iBAwBc,qBAAqB,MAAM,uBAAuB;;;;iBCnDlD,0BAA0B,2BAA2B,WAAW;;;;;;;;;;iBAyEhE,iCAAiC,SAAS;;;;iBAoC1C,4BAA4B,SAAS;;iBAgBrC,mBAAmB,SAAS;;iBAW5B,YAAY,SAAS;;iBA4BrB,kBACd,eAAe,gBACf,iBAAiB;;;UCvKF;EACf;EACA;EACA;;UAGe,qBAAqB,gBAAgB;EACpD;EACA;EACA;EACA,SACE,QAAQ,SACR,SAAS,4BACR,QAAQ;;UAGI;EACf;EACA;EACA,WAAW;;KAGD,oBAAoB,eAAe,4BAC7C,OAAO,QACP,SAAS,8BACN,qBAAqB,QAAQ;UAEjB;WACN,QAAQ;WACR;;KAGN,sBAAsB,WAAW,KACpC,iBAAiB;EAGjB;;UAGQ,qBAAqB,kBAAkB;EAC/C;EACA;EACA;EACA,aAAa,UAAU;;EAEvB,eAAe;;iBAGD,sBACd,gBAAgB,yBAChB,WACA,kBAAkB,WAAW,UAE7B,WAAW,qBAAqB,UAChC,SAAS,qBAAqB;EAC5B,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;EAC/D,WAAW,sBAAsB;IAElC,YAAY,WAAW;iBA8CV,qBACd,eAAe,yBACf,WACA,kBAAkB,WAAW,UAE7B,QAAQ,oBAAoB,SAC5B,SAAS,qBAAqB;EAC5B;EACA,SAAS;IAAS,UAAU;IAAW,UAAU;MAAc;;EAEzD;EAAc;;EACd;EAAa,UAAU,sBAAsB;KAEpD,YAAY,WAAW;;;KCxFrB,qBAAqB;KACrB,iBAAiB,SAAS;KAC1B,aAAa,gBAAgB,aAAa,MAAM,mBAAmB;UAcvD;;WAEN;;WAEA;;WAEA;;WAEA;;WAEA;;UAGM;;EAEf,OAAO;IAAQ;IAAiB;MAAkB,QAAQ;;;EAG1D,SAAS,UAAU,UAAU,kBAAkB,QAAQ;;EAEvD,QAAQ,UAAU,WAAW;;;cAIlB,6BAA6B;WAG7B;EAFX,YACE,iBACS;;UAOI;;EAEf;;EAEA;;EAEA;;;;EAIA,MAAM;;UAobS;;EAEf;;EAEA;;EAEA;;;EAGA,YAAY;;;;;iBAoJE,mBAAmB,MAAM,4BAA4B;;;;iBA0FrD,kBACd,SAAS,aACT,uBACC;;;iBAOa,oBAAoB,SAAS,aAAa"}
|
package/dist/index.d.ts
CHANGED
|
@@ -4,37 +4,37 @@ import { C as verifyAgentProfileCell, h as agentProfileCellKey, i as AgentProfil
|
|
|
4
4
|
import { C as TraceEvent, E as isToolSpan, S as ToolSpan, T as isLlmSpan, _ as Span, a as FAILURE_CLASSES, c as JudgeSpan, f as Run, h as RunStatus, l as LlmSpan, n as BudgetLedgerEntry, o as FailureClass, r as BudgetSpec, s as GenericSpan, t as Artifact, w as isJudgeSpan } from "./schema-DID1Cqct.js";
|
|
5
5
|
import { _ as validateRunRecord, a as RunRecord, c as RunTaskFailure, d as UNKNOWN_MODEL, f as isRunRecord, g as runTaskScore, h as roundTripRunRecord, i as RunOutcome, l as RunTerminalOutcome, m as parseRunRecordSafe, n as RunCostProvenance, o as RunRecordValidationError, p as modelHasSnapshot, r as RunJudgeMetadata, s as RunSplitTag, t as JudgeScoresRecord, u as RunTokenUsage } from "./run-record-DQjRcYwA.js";
|
|
6
6
|
import { A as SearchSpanResult, B as ViewSpansResult, D as DatasetOverview, E as DEFAULT_TRACE_ANALYST_BUDGETS, F as TraceAnalystFilters, H as ViewTraceResult, I as TraceAnalystSpan, L as TraceAnalystSpanKind, M as SpanMatchRecord, N as TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, O as ErrorCluster, P as TraceAnalystByteBudgets, R as TraceAnalystSpanStatus, V as ViewTraceOversized, _ as ProposalFinding, b as makeFinding, c as AnalystRunInputs, d as AnalystSeverity, f as AnalystUsageReceipt, i as AnalystFinding, j as SearchTraceResult, k as QueryTracesPage, l as AnalystRunResult, n as AnalystContext, p as EvidenceRef, s as AnalystRunEvent, t as Analyst, u as AnalystRunSummary, w as TraceAnalysisStore, x as makeProposalFinding, y as computeFindingId, z as TraceAnalystTraceSummary } from "./types-DMoNFDWi.js";
|
|
7
|
-
import { A as REDACTION_VERSION, B as exportRunAsOtlp, C as scoreTraceInsightReadiness, D as AnalyzeTracesResult, F as OtlpFlatLine, G as ExtractedUsage, L as OtlpExport, M as RedactionRule, N as redactString, O as analyzeTraces, U as captureFetchToRawSink, X as createBoundedTraceAnalysisStore, Y as extractUsageFromSse, _ as buildTraceInsightPrompt, b as domainEvidencePattern, g as buildTraceInsightContext, k as DEFAULT_REDACTION_RULES, m as TraceInsightSuite, n as toolSpansToTraceAnalysisStore, o as otlpTextToTraceAnalysisStore, p as TraceInsightReadiness, q as extractUsage, r as OtlpFileTraceStore, w as tokenizeDomainWords, x as inferDomainKeywords, y as describeTraceInsightScope, z as OtlpSpan } from "./store-tool-spans-
|
|
8
|
-
import { $n as
|
|
7
|
+
import { A as REDACTION_VERSION, B as exportRunAsOtlp, C as scoreTraceInsightReadiness, D as AnalyzeTracesResult, F as OtlpFlatLine, G as ExtractedUsage, L as OtlpExport, M as RedactionRule, N as redactString, O as analyzeTraces, U as captureFetchToRawSink, X as createBoundedTraceAnalysisStore, Y as extractUsageFromSse, _ as buildTraceInsightPrompt, b as domainEvidencePattern, g as buildTraceInsightContext, k as DEFAULT_REDACTION_RULES, m as TraceInsightSuite, n as toolSpansToTraceAnalysisStore, o as otlpTextToTraceAnalysisStore, p as TraceInsightReadiness, q as extractUsage, r as OtlpFileTraceStore, w as tokenizeDomainWords, x as inferDomainKeywords, y as describeTraceInsightScope, z as OtlpSpan } from "./store-tool-spans-BvdUbeOB.js";
|
|
8
|
+
import { $n as ServedModelCheck, A as CanaryOptions, C as RedTeamReport, D as CanaryAlert, E as scoreRedTeamOutput, M as runCanaries, Mn as CampaignCellRetryPolicy, Nn as RunCampaignOptions, O as CanaryEvaluation, Pn as runCampaign, Qn as ServedCrossFamilyError, S as RedTeamFinding, T as redTeamReport, Xn as AssertServedModelOptions, Yn as AssertCrossFamilyServedOptions, Zn as ModelSubstitutionError, ar as checkServedModel, at as transientDispatchFailure, b as RedTeamCase, cr as CrossFamilyError, dr as judgeFamily, er as ServedModelPolicy, ir as assertServedModels, j as CanaryReport, jn as CampaignCellFailureReceipt, k as CanaryKind, lr as JudgeFamily, nr as assertCrossFamilyServed, nt as llmJudge, or as servedModelAcceptable, rr as assertServedModel, sr as AssertCrossFamilyOptions, tr as ServedModelVerdict, tt as LlmJudgeOptions, ur as assertCrossFamily, w as redTeamDataset, x as RedTeamCategory, y as DEFAULT_RED_TEAM_CORPUS } from "./provenance-Dp-vvyrU.js";
|
|
9
9
|
import { a as NoopRawProviderSink, c as RawProviderEvent, l as RawProviderSink, r as InMemoryRawProviderSink, t as FileSystemRawProviderSink } from "./raw-provider-sink-BU29Sh8h.js";
|
|
10
10
|
import { A as LlmToolCall, C as LlmCallMetadata, D as LlmMessage, E as LlmChargeBounds, F as costReceiptFromLlmError, I as isTransientLlmError, L as maximumChargeForLlmRequest, M as LlmToolDefinition, N as LlmUsage, O as LlmResponseError, P as costReceiptFromLlm, R as stripFencedJson, S as LlmCallError, T as LlmCallResult, _ as CreateChatClientOpts, c as PersonaConfig, d as Scenario, f as ChatCallOpts, h as ChatResponse, i as JudgeFn, j as LlmToolChoice, k as LlmTokenLogprob, l as ProductClientConfig, m as ChatRequest, n as CompletionCriterion, o as JudgeRubric, p as ChatClient, r as DriverState, s as JudgeScore, t as CheckResult, u as RouteMap, w as LlmCallRequest, x as createChatClient } from "./types-Bfk0uxRj.js";
|
|
11
|
-
import { S as JudgeConfig, a as CampaignResult, w as JudgeScore$1 } from "./types-JHMOqZI4.js";
|
|
12
|
-
import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario, t as Dataset } from "./dataset-DQqhOCPt.js";
|
|
13
|
-
import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, s as TraceStore, t as EventFilter } from "./store-Cq9oOrI1.js";
|
|
14
|
-
import { C as defineEquivalenceCheck, S as buildEquivalenceRecord, _ as StrategyChecker, a as CheckerIdentity, b as VerificationStrategyProfile, c as EquivalenceCheckDefinition, d as EquivalenceCheckerInput, f as EquivalenceCheckerResult, g as EquivalenceRecord, h as EquivalenceProtocolError, i as equivalenceVerdict, l as EquivalenceCheckSpec, m as EquivalenceObligationStatus, n as VerdictCertification, o as CheckerOutcome, p as EquivalenceObligation, r as certificationEvidenceDigest, s as EquivalenceArm, t as DefaultVerdict, u as EquivalenceChecker, v as VERIFICATION_STRATEGIES, w as runEquivalenceCheck, x as VerificationStrategySource, y as VERIFICATION_STRATEGY_SOURCES } from "./verdict-E4eRNf7-.js";
|
|
15
|
-
import { a as MultiLayerVerifier, c as VerifyOptions, i as LayerStatus, l as gradeSemanticStatus, n as Layer, o as Severity, r as LayerResult, s as VerificationReport, t as Finding } from "./multi-layer-verifier-BUaQ4C17.js";
|
|
16
11
|
import { a as ContinuousCalibrationResult, c as calibrateJudge, d as verbosityBias, i as ContinuousAgreementOptions, l as calibrateJudgeContinuous, n as CandidateScore, o as GoldenItem, r as ContinuousAgreement, s as VerbosityBiasResult, t as CalibrationResult, u as continuousAgreement } from "./judge-calibration-C5CbMYce.js";
|
|
17
12
|
import { a as CorpusAgreementPerDimension, c as corpusInterRaterAgreement, i as CorpusAgreementOptions, l as corpusInterRaterAgreementFromJudgeScores, n as SeriesConvergenceResult, o as CorpusAgreementReport, r as analyzeSeries, s as CorpusScoreRecord, t as SeriesConvergenceOptions, u as interRaterReliability } from "./series-convergence-D9WgpXGi.js";
|
|
18
|
-
import {
|
|
13
|
+
import { a as partialCredit, c as spearmanR, d as weightedMean, i as confidenceInterval, l as summarizeNumberSeries, n as WeightedCompositeInput, o as pearsonR, r as WeightedCompositeResult, s as ranks, t as SeriesDistribution, u as weightedComposite } from "./descriptive-B2iPaT9J.js";
|
|
19
14
|
import { n as bonferroni, r as holm, t as benjaminiHochberg } from "./multiplicity-DIWHvysC.js";
|
|
20
15
|
import { A as ProportionInterval, B as wilson, C as EProcess, D as eProcess, E as EProcessStep, F as pairedBinaryScale, I as pairedRiskDifference, L as pairedRiskDifferenceExact, M as ScoreRiskDifferenceResult, N as isBinaryOutcomeVector, O as ExactRiskDifferenceResult, P as mcnemar, R as pairedRiskDifferenceScore, S as pairRunRecords, T as EProcessState, _ as PairedArmsComparison, b as comparePairedArms, d as MatchedPair, f as MatchedRunRecordPair, g as PairedArmRow, h as PairRunRecordsResult, j as RiskDifferenceResult, k as McNemarResult, m as PairArmsResult, o as hashJson, p as PairArmsOptions, s as manifestContentDigest, u as ComparePairedArmsOptions, v as PairedCorrectness, w as EProcessOptions, x as pairArms, y as PairedMetricDelta, z as passAtK } from "./pre-registration-DHz6P_6f.js";
|
|
21
16
|
import { _ as pairedSignTest, a as PairedPromotionDecision, c as BOOTSTRAP_GATE_MIN_N, d as PairedBootstrapResult, f as PairedSignTestResult, g as pairedDeltaTieFraction, h as pairedBootstrap, i as PairedMcNemarEvidence, l as DECISION_PAIRED_DELTA_STATISTIC, m as SignTestAlternative, n as PairedDecisionShape, o as PairedPromotionDecisionOptions, p as PairedTTestResult, r as PairedDecisionStatistic, s as decidePairedPromotion, t as PairedDecisionMethod, u as PairedBootstrapOptions, v as pairedTTest } from "./paired-promotion-decision-CGzg0cI_.js";
|
|
22
|
-
import {
|
|
17
|
+
import { C as requiredSampleSize, S as requiredPairedSampleSize, a as attest, b as mcnemarRequiredN, c as ExperimentRep, f as ImprovementThresholds, g as improvementVerdict, i as AttestedReport, l as ExperimentStats, m as computeExperimentStats, n as AttestationProvenance, o as verifyAttestation, p as ImprovementVerdictResult, r as AttestationVerification, t as ATTESTATION_ALGORITHM, v as mulberry32, x as pairedMde, y as mcnemarPower } from "./attestation-CJBGmMVh.js";
|
|
23
18
|
import { C as RankTestMethod, D as WilcoxonSignedRankResult, E as WILCOXON_EXACT_MAX_N, O as mannWhitneyU, S as MannWhitneyResult, T as RankTestOptions, _ as bootstrapCi, a as ReleaseConfidenceIssue, b as MANN_WHITNEY_EXACT_MAX_STATES, f as evaluateReleaseConfidence, g as Verdict, i as ReleaseConfidenceInput, k as wilcoxonSignedRank, m as BootstrapResult, o as ReleaseConfidenceMetrics, p as BootstrapOptions, s as ReleaseConfidenceScorecard, t as ActionableSideInfo, u as ReleaseTraceEvidence, w as RankTestMethodRequest, x as MANN_WHITNEY_EXACT_MAX_WORK, y as DEFAULT_PERMUTATIONS } from "./release-confidence-D6lQw_o7.js";
|
|
19
|
+
import { S as JudgeConfig, a as CampaignResult, w as JudgeScore$1 } from "./types-nokrtr7M.js";
|
|
20
|
+
import { i as DatasetSplit, n as DatasetManifest, r as DatasetScenario, t as Dataset } from "./dataset-DQqhOCPt.js";
|
|
21
|
+
import { a as RunFilter, i as InMemoryTraceStore, n as FileSystemTraceStore, o as SpanFilter, s as TraceStore, t as EventFilter } from "./store-Cq9oOrI1.js";
|
|
22
|
+
import { C as defineEquivalenceCheck, S as buildEquivalenceRecord, _ as StrategyChecker, a as CheckerIdentity, b as VerificationStrategyProfile, c as EquivalenceCheckDefinition, d as EquivalenceCheckerInput, f as EquivalenceCheckerResult, g as EquivalenceRecord, h as EquivalenceProtocolError, i as equivalenceVerdict, l as EquivalenceCheckSpec, m as EquivalenceObligationStatus, n as VerdictCertification, o as CheckerOutcome, p as EquivalenceObligation, r as certificationEvidenceDigest, s as EquivalenceArm, t as DefaultVerdict, u as EquivalenceChecker, v as VERIFICATION_STRATEGIES, w as runEquivalenceCheck, x as VerificationStrategySource, y as VERIFICATION_STRATEGY_SOURCES } from "./verdict-E4eRNf7-.js";
|
|
23
|
+
import { a as MultiLayerVerifier, c as VerifyOptions, i as LayerStatus, l as gradeSemanticStatus, n as Layer, o as Severity, r as LayerResult, s as VerificationReport, t as Finding } from "./multi-layer-verifier-BUaQ4C17.js";
|
|
24
24
|
import { a as Objective, c as paretoFrontier, o as ParetoResult, s as dominates } from "./power-preflight-Ptse_Kq7.js";
|
|
25
25
|
import { C as HeldOutGateConfig, S as HeldOutGate, T as SplitCoverage, _ as paretoChart, a as ParetoPoint, b as GateDecision, d as ResearchReportOptions, g as gainHistogram, h as SummaryTableRow, m as SummaryTableOptions, n as GainDistributionFigureSpec, p as SummaryTable, r as GainDistributionOptions, s as ResearchReport, t as GainDistributionBin, w as HeldOutGateRejectionCode, x as GateEvidence, y as summaryTable } from "./summary-report-DRstQNBX.js";
|
|
26
26
|
import { a as FailureContext, i as FailureClassification, o as FailureRule, r as failureClusterView, s as classifyFailure } from "./failure-cluster-6YSvsKlp.js";
|
|
27
27
|
import { o as InsightReport } from "./insight-report-08F022xN.js";
|
|
28
|
-
import { a as DefinedAgentEval, c as SelfImproveOptions, f as selfImprove, i as DefineAgentEvalOptions, o as defineAgentEval, u as SelfImproveResult } from "./define-agent-eval-
|
|
29
|
-
import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-
|
|
28
|
+
import { a as DefinedAgentEval, c as SelfImproveOptions, f as selfImprove, i as DefineAgentEvalOptions, o as defineAgentEval, u as SelfImproveResult } from "./define-agent-eval-0wW7gFhr.js";
|
|
29
|
+
import { a as createTraceAnalyst, i as TraceAnalystDefinition, n as buildDefaultAnalystRegistry, r as CreateTraceAnalystOptions, t as DefaultAnalystRegistryOptions } from "./default-registry-XxedTLwu.js";
|
|
30
30
|
import { i as ExactAnalystRunEvent, l as ExactCapableAnalyst, o as ExactAnalystRunResult } from "./exact-types-BEecmnWm.js";
|
|
31
31
|
import { _ as SatisfiedBy, a as ArtifactEventLike, b as createLlmCorrectnessChecker, c as ToolCallEventLike, d as CompletionVerdict, f as CorrectnessChecker, g as RequirementCheck, h as ProducedState, i as summarizeBackendIntegrity, l as extractProducedState, m as ProducedProposal, n as BackendIntegrityReport, o as ProposalEventLike, p as LlmCorrectnessCheckerOpts, r as assertRealBackend, s as RuntimeEventLike, t as BackendIntegrityError, u as CompletionRequirement, v as TaskGold, x as verifyCompletion, y as completionVerdict } from "./backend-integrity-e79K3UPD.js";
|
|
32
32
|
import { c as RegistryRunOpts, i as BudgetPolicy, n as AnalystRegistry, r as AnalystRegistryOptions, s as ExactRegistryRunOpts } from "./registry-xEb_xfns.js";
|
|
33
|
-
import { _ as AnalyzeRunsOptions, b as analyzeRuns, c as RawAnalystFinding, r as TraceAnalysisEngineResult, t as TraceAnalysisEngine } from "./engine-
|
|
34
|
-
import { A as SemanticConceptJudgeResult, C as DspyRlmTraceEngineOptions, D as SEMANTIC_CONCEPT_JUDGE_VERSION, E as ConceptSpec, O as SemanticConceptJudgeInput, T as ConceptFinding, _ as FindingSubject, a as FAILURE_MODE_KIND_SPEC, d as FindingsStore, f as PersistedFinding, j as runSemanticConceptJudge, k as SemanticConceptJudgeOptions, l as DiffPolicy, m as diffFindings, t as DEFAULT_TRACE_ANALYST_KINDS, u as FindingsDiff, v as FindingSubjectKind, w as createDspyRlmTraceEngine } from "./index-
|
|
33
|
+
import { _ as AnalyzeRunsOptions, b as analyzeRuns, c as RawAnalystFinding, r as TraceAnalysisEngineResult, t as TraceAnalysisEngine } from "./engine-BfRay1qD.js";
|
|
34
|
+
import { A as SemanticConceptJudgeResult, C as DspyRlmTraceEngineOptions, D as SEMANTIC_CONCEPT_JUDGE_VERSION, E as ConceptSpec, O as SemanticConceptJudgeInput, T as ConceptFinding, _ as FindingSubject, a as FAILURE_MODE_KIND_SPEC, d as FindingsStore, f as PersistedFinding, j as runSemanticConceptJudge, k as SemanticConceptJudgeOptions, l as DiffPolicy, m as diffFindings, t as DEFAULT_TRACE_ANALYST_KINDS, u as FindingsDiff, v as FindingSubjectKind, w as createDspyRlmTraceEngine } from "./index-DDAPhUJJ.js";
|
|
35
35
|
import { a as ProfileAxisSpec, c as expandProfileAxes, i as HarnessType, l as harnessAxisOf, n as CODING_HARNESSES, o as agentProfileHash, r as HARNESS_NATIVE_MODEL, s as agentProfileId, t as AgentProfile } from "./agent-profile-B9_GGsG8.js";
|
|
36
36
|
import { a as MintedRolloutLine } from "./schema-BzWDXhOR.js";
|
|
37
|
-
import { i as BenchmarkEvaluation } from "./types-
|
|
37
|
+
import { i as BenchmarkEvaluation } from "./types-CCZ34qmV.js";
|
|
38
38
|
import { i as TraceEmitter, r as SpanHandle } from "./emitter-Bvnu0VzL.js";
|
|
39
39
|
import { n as SandboxDriver } from "./sandbox-harness-BlSOu4LX.js";
|
|
40
40
|
import { a as PairedEvalueStep, c as SequentialDecision, i as PairedEvalueSequence, l as evaluateInterimReleaseConfidence, n as InterimReleaseConfidenceInput, r as PairedEvalueOptions, t as InterimReleaseConfidence, u as pairedEvalueSequence } from "./sequential-BhsrMupG.js";
|
|
@@ -516,6 +516,145 @@ interface ActiveLearningOptions {
|
|
|
516
516
|
}
|
|
517
517
|
declare function proposeSynthesisTargets(dataset: Dataset, traceStore: TraceStore, options?: ActiveLearningOptions): Promise<SynthesisTarget[]>;
|
|
518
518
|
//#endregion
|
|
519
|
+
//#region src/bounded-process.d.ts
|
|
520
|
+
/**
|
|
521
|
+
* `runBoundedProcess` — run one command under a wall-clock bound and return
|
|
522
|
+
* what it produced, leaving no descendant behind.
|
|
523
|
+
*
|
|
524
|
+
* The child leads its own process group (`detached` on every non-win32
|
|
525
|
+
* platform), so a deadline or an abort kills the group with
|
|
526
|
+
* `process.kill(-pid, 'SIGKILL')` rather than only the shell that was
|
|
527
|
+
* spawned. That distinction is the whole point of this module: `sh -c
|
|
528
|
+
* 'solver & wait'` puts the real work in a grandchild, and a kill aimed at
|
|
529
|
+
* the shell alone leaves that grandchild running, holding the stdout and
|
|
530
|
+
* stderr pipes open. `close` then never fires, so the deadline produces no
|
|
531
|
+
* result the caller can observe and the descendant outlives the run.
|
|
532
|
+
*
|
|
533
|
+
* Two further guarantees the caller depends on:
|
|
534
|
+
*
|
|
535
|
+
* - A killed run never reports success. A SIGKILLed child can still close
|
|
536
|
+
* with exit code 0, so a run killed by the deadline or by an abort has
|
|
537
|
+
* its exit code forced non-zero.
|
|
538
|
+
* - A run always resolves. If a grandchild escaped the group and still
|
|
539
|
+
* holds the pipes, the result is delivered {@link KILL_DRAIN_MS} after
|
|
540
|
+
* the kill instead of waiting on a `close` that will not come.
|
|
541
|
+
*
|
|
542
|
+
* The captured output is bounded. Once the cap is reached the streams are
|
|
543
|
+
* still drained — the child needs its pipes read to reach `close` — but the
|
|
544
|
+
* bytes are discarded and `outputTruncated` is set.
|
|
545
|
+
*/
|
|
546
|
+
interface BoundedProcessInput {
|
|
547
|
+
/**
|
|
548
|
+
* The command line, interpreted by a shell unless `shell` is `false`. When
|
|
549
|
+
* {@link BoundedProcessInput.args} is present this is a program path
|
|
550
|
+
* instead, and no shell reads it.
|
|
551
|
+
*/
|
|
552
|
+
command: string;
|
|
553
|
+
/**
|
|
554
|
+
* Argument vector for `command`, delivered to the program as separate
|
|
555
|
+
* argv entries with no shell between them.
|
|
556
|
+
*
|
|
557
|
+
* This is the form to use whenever an argument carries text the caller did
|
|
558
|
+
* not author — a script body, a path, a pattern. Without it, `interpreter
|
|
559
|
+
* -flag <script>` can only be expressed by quoting the script into a
|
|
560
|
+
* shell string, and a quoting bug there is a command injection rather than
|
|
561
|
+
* a wrong answer. Each entry arrives verbatim, so there is nothing to
|
|
562
|
+
* quote: `{ command: 'bash', args: ['-n', '-c', body] }` parses `body`
|
|
563
|
+
* whatever shell metacharacters it holds.
|
|
564
|
+
*
|
|
565
|
+
* Two limits are the platform's, not this module's. A NUL byte in an entry
|
|
566
|
+
* cannot be passed to a process at all, and is reported as a `runnerError`
|
|
567
|
+
* rather than thrown. A lone surrogate is encoded to UTF-8 as U+FFFD, and a
|
|
568
|
+
* non-string entry from an untypechecked caller is coerced with `String()`,
|
|
569
|
+
* so an entry is verbatim exactly when it is a string a UTF-8 argv can hold.
|
|
570
|
+
*
|
|
571
|
+
* A shell cannot interpret an argument vector, so `args` and a truthy
|
|
572
|
+
* `shell` are contradictory. Passing both spawns nothing and reports
|
|
573
|
+
* `runnerError`.
|
|
574
|
+
*/
|
|
575
|
+
args?: string[];
|
|
576
|
+
/**
|
|
577
|
+
* Shell that interprets `command`. `true` (the default) uses the platform
|
|
578
|
+
* default — `/bin/sh` on POSIX, `cmd.exe` on Windows. A string names the
|
|
579
|
+
* shell binary, e.g. `'bash'`, which is required for a bash-only construct
|
|
580
|
+
* such as `[[ ]]` or `$BASH_VERSION`. `false` executes `command` directly
|
|
581
|
+
* as a program path with no arguments. Ignored — and refused when truthy —
|
|
582
|
+
* once {@link BoundedProcessInput.args} is present.
|
|
583
|
+
*/
|
|
584
|
+
shell?: boolean | string;
|
|
585
|
+
/** Working directory. The parent's cwd is inherited when omitted. */
|
|
586
|
+
cwd?: string;
|
|
587
|
+
/**
|
|
588
|
+
* Environment for the child. How it combines with the parent environment
|
|
589
|
+
* is decided by {@link BoundedProcessInput.envMode}.
|
|
590
|
+
*/
|
|
591
|
+
env?: NodeJS.ProcessEnv;
|
|
592
|
+
/**
|
|
593
|
+
* `'merge'` (the default) layers `env` on top of `process.env`, so the
|
|
594
|
+
* child sees every variable this process has plus the named ones.
|
|
595
|
+
* `'replace'` passes exactly `env` and nothing else, which is what a
|
|
596
|
+
* grader needs when the claim under test is about the variables the
|
|
597
|
+
* command could read. `PATH` is not re-added under `'replace'`, and a
|
|
598
|
+
* command that must find a binary on `PATH` names `PATH` itself.
|
|
599
|
+
*
|
|
600
|
+
* With no `PATH` in the child environment the two forms do NOT search the
|
|
601
|
+
* same directories, so a caller under `'replace'` that names no `PATH` gets
|
|
602
|
+
* a resolution that depends on which form it picked. The shell form execs
|
|
603
|
+
* the shell, which applies its own compiled-in default; the argv form
|
|
604
|
+
* reaches `execvp`, which falls back to `confstr(_CS_PATH)`. Measured on
|
|
605
|
+
* Linux/glibc: `/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin`
|
|
606
|
+
* against `/bin:/usr/bin`, so a binary in `/usr/local/bin` resolves on the
|
|
607
|
+
* shell form and reports ENOENT on the argv form. Name `PATH` and both
|
|
608
|
+
* forms search exactly what was named.
|
|
609
|
+
*/
|
|
610
|
+
envMode?: 'merge' | 'replace';
|
|
611
|
+
/** Wall-clock bound in ms. Defaults to 10 minutes. */
|
|
612
|
+
timeoutMs?: number;
|
|
613
|
+
/**
|
|
614
|
+
* Cap on captured stdout+stderr. Reaching it truncates the captured text
|
|
615
|
+
* and sets `outputTruncated`; it never kills the process. Defaults to
|
|
616
|
+
* 16 MiB.
|
|
617
|
+
*/
|
|
618
|
+
maxOutputBytes?: number;
|
|
619
|
+
/**
|
|
620
|
+
* Kills the process group when it aborts, the same way the deadline does.
|
|
621
|
+
* A signal that is already aborted when `runBoundedProcess` is called
|
|
622
|
+
* prevents the spawn entirely — nothing runs.
|
|
623
|
+
*/
|
|
624
|
+
signal?: AbortSignal;
|
|
625
|
+
}
|
|
626
|
+
interface BoundedProcessResult {
|
|
627
|
+
/**
|
|
628
|
+
* Exit code. Forced non-zero whenever `killedByTimeout` or
|
|
629
|
+
* `killedBySignal` is set, because a SIGKILLed child can close with 0 and
|
|
630
|
+
* a killed run must never read as a pass.
|
|
631
|
+
*/
|
|
632
|
+
exitCode: number;
|
|
633
|
+
/** Captured stdout, truncated at `maxOutputBytes`. */
|
|
634
|
+
stdout: string;
|
|
635
|
+
/**
|
|
636
|
+
* Captured stderr, truncated at `maxOutputBytes`. A runner-side failure is
|
|
637
|
+
* reported in `runnerError`, not appended here; the caller decides whether
|
|
638
|
+
* the command's own stderr and the runner's diagnosis belong together.
|
|
639
|
+
*/
|
|
640
|
+
stderr: string;
|
|
641
|
+
/** Wall-clock duration in ms, measured across the whole call. */
|
|
642
|
+
wallMs: number;
|
|
643
|
+
/** The wall-clock bound was reached and the process group was killed. */
|
|
644
|
+
killedByTimeout: boolean;
|
|
645
|
+
/** `signal` aborted and the process group was killed. */
|
|
646
|
+
killedBySignal: boolean;
|
|
647
|
+
/** `maxOutputBytes` was reached and later output was discarded. */
|
|
648
|
+
outputTruncated: boolean;
|
|
649
|
+
/**
|
|
650
|
+
* Runner-side failure: the process could not be spawned, or the signal was
|
|
651
|
+
* already aborted so nothing was spawned. Absent when the command itself
|
|
652
|
+
* ran, however it exited.
|
|
653
|
+
*/
|
|
654
|
+
runnerError?: string;
|
|
655
|
+
}
|
|
656
|
+
declare function runBoundedProcess(input: BoundedProcessInput): Promise<BoundedProcessResult>;
|
|
657
|
+
//#endregion
|
|
519
658
|
//#region src/client.d.ts
|
|
520
659
|
/**
|
|
521
660
|
* ProductClient — configurable HTTP client for exercising any agent's APIs.
|
|
@@ -2904,5 +3043,5 @@ declare class PairwiseSteeringOptimizer {
|
|
|
2904
3043
|
optimize(rows: SteeringOptimizationRow[], config?: SteeringOptimizerConfig): SteeringOptimizationResult;
|
|
2905
3044
|
}
|
|
2906
3045
|
//#endregion
|
|
2907
|
-
export { AGENT_PROFILE_KINDS, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, AgentEvalError, type AgentEvalErrorCode, type AgentProfile, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileDimensionValue, type AgentProfileJson, type AgentProfileSourceInput, type Analyst, type AnalystContext, type AnalystFeedbackTrajectoryOptions, type AnalystFinding, type AnalystFindingDigest, AnalystRegistry, type AnalystRegistryOptions, type AnalystReviewDecision, type AnalystReviewRequest, type AnalystRunDigest, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeRunsOptions, type AnalyzeTracesResult, type Artifact, type ArtifactEventLike, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertCrossFamilyServedOptions, type AssertServedModelOptions, type AssertSingleBackendOptions, BOOTSTRAP_GATE_MIN_N, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BenchmarkEvaluation, type BlendWeights, type BootstrapOptions, type BootstrapResult, BudgetBreachError, BudgetGuard, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CalibrationResult, type CampaignCellFailureReceipt, type CampaignCellRetryPolicy, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignResult, type CampaignRunContext, type CampaignRunOutcome, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryEvaluation, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CandidateScore, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type CheckResult, type CheckerIdentity, type CheckerOutcome, type CliffsMagnitude, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptFinding, type ConceptSpec, ConfigError, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractSpan, type ContractVerdict, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRuntimeConfig, type ControlRuntimeError, type ControlStep, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostProvenance, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, CostReservationExceededError, type CostResult, type CostSummary, CostTracker, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateChatClientOpts, type CreateTraceAnalystOptions, CrossFamilyError, type CustomTokenPricing, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_PERMUTATIONS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, type DataAcquisitionPlan, type DatasetManifest, type DatasetOverview, type DatasetScenario, type DatasetSplit, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DefineAgentEvalOptions, type DefinedAgentEval, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DetectorEvent, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, type DriverState, type DspyRlmTraceEngineOptions, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type EquivalenceArm, type EquivalenceCheckDefinition, type EquivalenceCheckSpec, type EquivalenceChecker, type EquivalenceCheckerInput, type EquivalenceCheckerResult, type EquivalenceObligation, type EquivalenceObligationStatus, EquivalenceProtocolError, type EquivalenceRecord, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EventFilter, type EvidenceRef, type ExactAnalystRunEvent, type ExactAnalystRunResult, type ExactCapableAnalyst, type ExactRegistryRunOpts, type ExactRiskDifferenceResult, type ExperimentPlan, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExtractOptions, type ExtractResult, type ExtractedUsage, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FailedRun, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision, type GateEvidence, type GenericSpan, type GoldenItem, HARNESS_NATIVE_MODEL, type HarnessType, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, type InsightReport, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, JudgeError, type JudgeFamily, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, type JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeSensitivity, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, type LlmChargeBounds, type LlmCorrectnessCheckerOpts, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, type LlmSpan, type LlmTokenLogprob, type LlmToolCall, type LlmToolChoice, type LlmToolDefinition, type LlmUsage, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, MODEL_PRICING, type MannWhitneyResult, type MatchedPair, type MatchedRunRecordPair, type MaximumCharge, type McNemarResult, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type ModelEndpointCheck, type ModelEndpointRequest, type ModelSeats, ModelSubstitutionError, MultiLayerVerifier, type NoLeakOptions, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, type Objective, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OtlpExport, OtlpFileTraceStore, type OtlpFlatLine, type OtlpSpan, POLICY_EDIT_AXES, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, POLICY_EDIT_TARGET_SURFACES, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairRunRecordsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedDecisionMethod, type PairedDecisionShape, type PairedDecisionStatistic, type PairedDeltaTestOptions, type PairedDeltaTestResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMcNemarEvidence, type PairedMetricDelta, type PairedPromotionDecision, type PairedPromotionDecisionOptions, type PairedSignTestResult, type PairedTTestResult, PairwiseSteeringOptimizer, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, PolicyEditValidationError, type PreferenceMemoryEntry, type ProducedProposal, type ProducedState, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkValidationReport, ProductClient, type ProductClientConfig, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type PromptHandle, PromptRegistry, type ProportionInterval, type ProposalEventLike, type ProposalFinding, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposedSideEffect, type QueryTracesPage, REDACTION_VERSION, type RankTestMethod, type RankTestMethodRequest, type RankTestOptions, type RawAnalystFinding, type RawProviderEvent, type RawProviderSink, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamReport, type RedactionRule, type ReferenceReplayCaseRun, type ReferenceReplayRun, type ReferenceReplaySplit, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseTraceEvidence, type RepeatedActionOptions, type RequirementCheck, type ResearchReport, type ResearchReportOptions, type Researcher, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type RiskDifferenceResult, type RouteMap, type RoutedField, type Run, type RunCampaignOptions, type RunCommandInput, type RunCommandResult, type RunCostProvenance, type RunFilter, RunIntegrityError, type RunIntegrityReport, type RunJudgeMetadata, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSplitTag, type RunStatus, type RunTaskFailure, type RunTerminalOutcome, type RunTokenUsage, type RuntimeEventLike, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, type SandboxDriver, type SatisfiedBy, type Scenario, type ScenarioCost, type ScorePreference, type ScoreRiskDifferenceResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type SearchSpanResult, type SearchTraceResult, type SelfImproveOptions, type SelfImproveResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SeriesConvergenceOptions, type SeriesConvergenceResult, type SeriesDistribution, ServedCrossFamilyError, type ServedModelCheck, type ServedModelPolicy, type ServedModelVerdict, type Severity, type SignTestAlternative, type SingleBackendDivergence, type SingleBackendReport, type Span, type SpanFilter, type SpanHandle, type SpanMatchRecord, type SplitCoverage, type SteeringBundle, type SteeringChange, type SteeringOptimizationResult, type SteeringOptimizationRow, type StopDecision, type StrategyChecker, type StreamingDetector, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SynthesisTarget, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, type TaskGold, type TaskHeadroom, type ToolCallEventLike, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAnalysisEngine, type TraceAnalysisEngineResult, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystDefinition, type TraceAnalystFilters, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, TraceEmitter, type TraceEvent, type TraceInsightReadiness, type TraceInsightSuite, type TraceStore, type Trajectory, type TrajectoryStep, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, UNKNOWN_MODEL, type UserQuestion, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, type VerbosityBiasResult, type Verdict, type VerdictCacheStore, type VerdictCertification, type Verification, type VerificationReport, type VerificationStrategyProfile, type VerificationStrategySource, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type ViteDeployRunnerInput, WILCOXON_EXACT_MAX_N, type WeightedCompositeInput, type WeightedCompositeResult, type WilcoxonSignedRankResult, type WranglerDeployRunnerInput, acquisitionPlansForKnowledgeGaps, admitPolicyEdit, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, aggregateJudgeVerdicts, aggregateRunScore, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeRuns, analyzeSeries, analyzeTraces, applyPolicyEditToSurface, argHash, assertCapabilityHeadroom, assertCrossFamily, assertCrossFamilyServed, assertNoHiddenLeak, assertNoJudgeVerdict, assertProductBenchmarkRun, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assertSingleBackend, assignFeedbackSplit, benjaminiHochberg, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildReflectionPrompt, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, canonicalJson, capabilityHeadroom, captureFetchToRawSink, certificationEvidenceDigest, checkCanaries, checkServedModel, checkTraceContracts, clamp01, classifyFailure, cliffsDelta, cohensD, comparePairedArms, completionVerdict, computeExperimentStats, computeFindingId, computePolicyEditId, computeToolUseMetrics, confidenceInterval, contentHash, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createLlmReviewer, createTraceAnalyst, decidePairedPromotion, defaultBlendWeights, defineAgentEval, defineEquivalenceCheck, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, domainEvidencePattern, dominates, eProcess, ensembleJudge, equivalenceVerdict, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, expandProfileAxes, exportProductBenchmark, exportProductBenchmarkRuns, exportRunAsOtlp, extractErrorCount, extractProducedState, extractUsage, extractUsageFromSse, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, fileVerdictCache, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gradeOnHidden, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashContent, hashJson, hiddenGrade, holm, improvementVerdict, inMemoryReviewStore, inferDomainKeywords, interRaterReliability, interpretCliffs, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isModelPriced, isPolicyEdit, isRunRecord, isToolSpan, isTransientLlmError, jsonShape, jsonlReviewStore, jsonlRunRecordBackend, judgeAgreementView, judgeFamily, judgeSpans, knowledgeReadinessTracePayload, leaderboard, llmJudge, loadScorecard, localCommandRunner, makeFinding, makePolicyEdit, makePolicyEditCandidateRecord, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, minimumPairsForPairedDeltaTest, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, notBlocked, objectiveEval, observeAll, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedCohensDz, pairedDeltaTest, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseReflectionResponse, parseRunRecordSafe, partialCredit, partitionHeldOut, passAtK, pearsonR, policyEditFromFinding, policyEditsFromFindings, preflightModels, productBenchmarkRepoIdentity, index_d_exports as profile, projectRuntimeTrajectoryEvidence, proposeSynthesisTargets, ranks, readProductBenchmarkManifest, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redactString, regexMatches, renderPreferenceMemoryMarkdown, repeatedActionDetector, requiredPairedSampleSize, requiredSampleSize, resolveModelPricing, resolveSeat, roundTripRunRecord, routeFields, runAgentControlLoop, runCampaign, runCanaries, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runIntentMatchJudge, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runProposeReview, runProposeReviewAsControlLoop, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreKnowledgeReadiness, scorePolicyEditReadiness, scoreRedTeamOutput, scoreTraceInsightReadiness, seatPresets, selfImprove, servedModelAcceptable, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, textInSnapshot, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, traceContract, transientDispatchFailure, urlContains, userQuestionsForKnowledgeGaps, validatePolicyEdit, validatePolicyEditCandidateRecord, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyCompletion, viteDeployRunner, weightedComposite, weightedMean, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, wranglerDeployRunner };
|
|
3046
|
+
export { AGENT_PROFILE_KINDS, ATTESTATION_ALGORITHM, type ActionExecutionPolicy, type ActionPolicyDecision, type ActionableSideInfo, type ActiveLearningOptions, AgentEvalError, type AgentEvalErrorCode, type AgentProfile, type AgentProfileCell, type AgentProfileCellInput, type AgentProfileDimensionValue, type AgentProfileJson, type AgentProfileSourceInput, type Analyst, type AnalystContext, type AnalystFeedbackTrajectoryOptions, type AnalystFinding, type AnalystFindingDigest, AnalystRegistry, type AnalystRegistryOptions, type AnalystReviewDecision, type AnalystReviewRequest, type AnalystRunDigest, type AnalystRunEvent, type AnalystRunInputs, type AnalystRunResult, type AnalystRunSummary, type AnalystSeverity, type AnalystUsageReceipt, type AnalyzeRunsOptions, type AnalyzeTracesResult, type Artifact, type ArtifactEventLike, type AssertCapabilityHeadroomOptions, type AssertCrossFamilyOptions, type AssertCrossFamilyServedOptions, type AssertServedModelOptions, type AssertSingleBackendOptions, type AttestationProvenance, type AttestationVerification, type AttestedReport, BOOTSTRAP_GATE_MIN_N, type BackendDescriptor, BackendIntegrityError, type BackendIntegrityReport, type BenchmarkEvaluation, type BlendWeights, type BootstrapOptions, type BootstrapResult, type BoundedProcessInput, type BoundedProcessResult, BudgetBreachError, BudgetGuard, type BudgetLedgerEntry, type BudgetPolicy, type BudgetSpec, CODING_HARNESSES, type CalibrationResult, type CampaignCellFailureReceipt, type CampaignCellRetryPolicy, type CampaignFactoryParams, type CampaignIntegrityPolicy, type CampaignResult, type CampaignRunContext, type CampaignRunOutcome, type CampaignScenario, type CampaignVariant, type CanaryAlert, type CanaryEvaluation, type CanaryKind, type CanaryLeak, type CanaryOptions, type CanaryReport, type CandidateScore, type CapabilityHeadroomOptions, type CapabilityHeadroomResult, type CellVerdict, type ChannelRollup, type ChatCallOpts, type ChatClient, type ChatRequest, type ChatResponse, type CheckResult, type CheckerIdentity, type CheckerOutcome, type CliffsMagnitude, type CommandRunner, type ComparePairedArmsOptions, type CompletionCriterion, type CompletionRequirement, type CompletionVerdict, type ConceptFinding, type ConceptSpec, ConfigError, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type ContractCheckResult, type ContractSpan, type ContractVerdict, type ControlActionOutcome, type ControlBudget, type ControlContext, type ControlDecision, type ControlEvalResult, type ControlRunResult, type ControlRuntimeConfig, type ControlRuntimeError, type ControlStep, type CorpusAgreementOptions, type CorpusAgreementPerDimension, type CorpusAgreementReport, type CorpusScoreRecord, type CorrectnessChecker, CostAccountingIncompleteError, CostCallConflictError, CostCeilingReachedError, type CostChannel, type CostEntry, CostLedger, type CostLedgerFilter, type CostLedgerHandle, type CostLedgerOptions, type CostLedgerPersistence, CostLedgerPersistenceError, type CostLedgerSummary, type CostProvenance, type CostReceipt, CostReceiptCaptureError, type CostReceiptInput, CostReservationExceededError, type CostResult, type CostSummary, CostTracker, type CostUsage, type CounterfactualContext, type CounterfactualMutation, type CounterfactualResult, type CounterfactualRunner, type CreateChatClientOpts, type CreateTraceAnalystOptions, CrossFamilyError, type CustomTokenPricing, DECISION_PAIRED_DELTA_STATISTIC, DEFAULT_PERMUTATIONS, DEFAULT_REDACTION_RULES, DEFAULT_RED_TEAM_CORPUS, DEFAULT_TRACE_ANALYST_BUDGETS, DEFAULT_TRACE_ANALYST_KINDS, type DataAcquisitionPlan, type DatasetManifest, type DatasetOverview, type DatasetScenario, type DatasetSplit, type DefaultAnalystRegistryOptions, type DefaultVerdict, type DefineAgentEvalOptions, type DefinedAgentEval, type DeployGateLayerInput, type DeployRunResult, type DeployRunner, type DetectorEvent, type DetectorSignal, type DiffPolicy, type DiffScorecardOptions, type DirEntry, type DiscoverPersonasOptions, type DiscoveredPersona, type DriverState, type DspyRlmTraceEngineOptions, type EProcess, type EProcessOptions, type EProcessState, type EProcessStep, ERROR_COUNT_PATTERNS, type EnsembleAggregate, type EnsembleJudgeOptions, type EquivalenceArm, type EquivalenceCheckDefinition, type EquivalenceCheckSpec, type EquivalenceChecker, type EquivalenceCheckerInput, type EquivalenceCheckerResult, type EquivalenceObligation, type EquivalenceObligationStatus, EquivalenceProtocolError, type EquivalenceRecord, type ErrorCluster, type ErrorCountPattern, type ErrorStreakOptions, type EvalCampaignOptions, type EvalCampaignResult, type EventFilter, type EvidenceRef, type ExactAnalystRunEvent, type ExactAnalystRunResult, type ExactCapableAnalyst, type ExactRegistryRunOpts, type ExactRiskDifferenceResult, type ExperimentPlan, type ExperimentRep, type ExperimentResult, type ExperimentStats, type ExtractOptions, type ExtractResult, type ExtractedUsage, FAILURE_CLASSES, FAILURE_MODE_KIND_SPEC, type FailedRun, type FailureClass, type FailureClassification, type FailureContext, type FailureMode, type FailureRule, type FeedbackArtifactType, type FeedbackAttempt, type FeedbackLabel, type FeedbackLabelKind, type FeedbackLabelSource, type FeedbackOptimizerRow, type FeedbackOutcome, type FeedbackSplitPolicy, type FeedbackTask, type FeedbackTrajectory, type FeedbackTrajectoryFilter, type FeedbackTrajectoryStore, type FieldDestination, FileSystemFeedbackTrajectoryStore, FileSystemRawProviderSink, FileSystemTraceStore, type Finding, type FindingSubject, type FindingSubjectKind, type FindingToPolicyEditOptions, type FindingsDiff, FindingsStore, type GainDistributionBin, type GainDistributionFigureSpec, type GainDistributionOptions, type GateDecision, type GateEvidence, type GenericSpan, type GoldenItem, HARNESS_NATIVE_MODEL, type HarnessType, type HeadroomInput, HeldOutGate, type HeldOutGateConfig, type HeldOutGateRejectionCode, type HeldOutPartition, type HiddenCriteriaGrader, type HiddenGradeResult, type HiddenLeak, type ImprovementThresholds, type ImprovementVerdictResult, InMemoryFeedbackTrajectoryStore, InMemoryRawProviderSink, InMemoryTraceStore, type InsightReport, type IntentMatchInput, type IntentMatchOptions, type IntentMatchResult, type InterimReleaseConfidence, type InterimReleaseConfidenceInput, JudgeError, type JudgeFamily, type JudgeRetryOutcome, type JudgeRetryPolicy, type JudgeRubric, type JudgeScore, type JudgeScoreInput, type JudgeScoresRecord, type JudgeSpan, type JudgeVerdict, type KeywordConceptSpec, type KeywordCoverageFinding, type KeywordCoverageOptions, type KeywordCoverageResult, type KnowledgeAcquisitionMode, type KnowledgeBundle, type KnowledgeFreshness, type KnowledgeImportance, type KnowledgeReadinessReport, type KnowledgeRequirement, type KnowledgeRequirementCategory, type KnowledgeSensitivity, type Layer, type LayerResult, type LayerStatus, type LeaderboardOptions, type LeaderboardRow, LlmCallError, type LlmCallMetadata, type LlmCallRequest, type LlmCallResult, type LlmChargeBounds, type LlmCorrectnessCheckerOpts, type LlmJudgeOptions, type LlmMessage, LlmResponseError, type LlmReviewerConfig, type LlmSpan, type LlmTokenLogprob, type LlmToolCall, type LlmToolChoice, type LlmToolDefinition, type LlmUsage, MANN_WHITNEY_EXACT_MAX_STATES, MANN_WHITNEY_EXACT_MAX_WORK, MODEL_PRICING, type MannWhitneyResult, type MatchedPair, type MatchedRunRecordPair, type MaximumCharge, type McNemarResult, type MintRolloutOptions, type MintRolloutResult, type MintedRolloutLine, type ModelEndpointCheck, type ModelEndpointRequest, type ModelSeats, ModelSubstitutionError, MultiLayerVerifier, type NoLeakOptions, NoopRawProviderSink, NotFoundError, OUTPUT_VALUE, type Objective, type Oracle, type OracleObservation, type OracleReport, type OracleResult, type OtlpExport, OtlpFileTraceStore, type OtlpFlatLine, type OtlpSpan, POLICY_EDIT_AXES, POLICY_EDIT_CANDIDATE_RECORD_SCHEMA, POLICY_EDIT_TARGET_SURFACES, type PaidCallResult, type PairArmsOptions, type PairArmsResult, type PairRunRecordsResult, type PairedArmRow, type PairedArmsComparison, type PairedBootstrapOptions, type PairedBootstrapResult, type PairedCorrectness, type PairedDecisionMethod, type PairedDecisionShape, type PairedDecisionStatistic, type PairedDeltaTestOptions, type PairedDeltaTestResult, type PairedEvalueOptions, type PairedEvalueSequence, type PairedEvalueStep, type PairedMcNemarEvidence, type PairedMetricDelta, type PairedPromotionDecision, type PairedPromotionDecisionOptions, type PairedSignTestResult, type PairedTTestResult, PairwiseSteeringOptimizer, type ParetoPoint, type ParetoResult, type PartitionHeldOutOptions, type PendingCostCall, type PendingCostCallView, type PersistedFinding, type PersonaConfig, type PolicyEdit, type PolicyEditAdmission, type PolicyEditAdmissionOptions, type PolicyEditAxis, type PolicyEditCandidateRecord, type PolicyEditChange, type PolicyEditExpectedGain, type PolicyEditGainDirection, type PolicyEditGainUnit, type PolicyEditInit, type PolicyEditRisk, type PolicyEditSchemaVersion, type PolicyEditSource, type PolicyEditTarget, type PolicyEditTargetSurface, PolicyEditValidationError, type PreferenceMemoryEntry, type ProducedProposal, type ProducedState, type ProductBenchmarkExportOptions, type ProductBenchmarkExportResult, type ProductBenchmarkManifest, type ProductBenchmarkSingleRunExportOptions, type ProductBenchmarkSplit, type ProductBenchmarkValidationReport, ProductClient, type ProductClientConfig, type ProfileAxisSpec, type ProjectRuntimeTrajectoryEvidenceOptions, type PromptHandle, PromptRegistry, type ProportionInterval, type ProposalEventLike, type ProposalFinding, type ProposeFn, type ProposeInput, type ProposeOutput, type ProposeReviewConfig, type ProposeReviewControlAction, type ProposeReviewControlConfig, type ProposeReviewControlResult, type ProposeReviewControlState, type ProposeReviewReport, type ProposedSideEffect, type QueryTracesPage, REDACTION_VERSION, type RankTestMethod, type RankTestMethodRequest, type RankTestOptions, type RawAnalystFinding, type RawProviderEvent, type RawProviderSink, type RecordRunsOptions, type RedTeamCase, type RedTeamCategory, type RedTeamFinding, type RedTeamReport, type RedactionRule, type ReferenceReplayCaseRun, type ReferenceReplayRun, type ReferenceReplaySplit, type ReflectionContext, type ReflectionProposal, type RegistryRunOpts, type ReleaseConfidenceInput, type ReleaseConfidenceIssue, type ReleaseConfidenceMetrics, type ReleaseConfidenceScorecard, type ReleaseTraceEvidence, type RepeatedActionOptions, type RequirementCheck, type ResearchReport, type ResearchReportOptions, type Researcher, type Review, type ReviewFn, type ReviewInput, type ReviewMemoryEntry, type ReviewMemoryStore, type RiskDifferenceResult, type RouteMap, type RoutedField, type Run, type RunCampaignOptions, type RunCommandInput, type RunCommandResult, type RunCostProvenance, type RunFilter, RunIntegrityError, type RunIntegrityReport, type RunJudgeMetadata, type RunOutcome, type RunPaidCallInput, type RunRecord, type RunRecordBackend, RunRecordValidationError, type RunScore, type RunScoreWeights, type RunSplitTag, type RunStatus, type RunTaskFailure, type RunTerminalOutcome, type RunTokenUsage, type RuntimeEventLike, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryEvidenceSummary, type RuntimeTrajectoryHookEvent, type RuntimeTrajectoryRecord, type RuntimeTrajectoryRunRecord, SEMANTIC_CONCEPT_JUDGE_VERSION, type SandboxDriver, type SatisfiedBy, type Scenario, type ScenarioCost, type ScorePreference, type ScoreRiskDifferenceResult, type Scorecard, type ScorecardCell, type ScorecardCellDiff, type ScorecardDiff, type ScorecardEntry, type ScorecardLogLine, type SearchSpanResult, type SearchTraceResult, type SelfImproveOptions, type SelfImproveResult, type SemanticConceptJudgeInput, type SemanticConceptJudgeOptions, type SemanticConceptJudgeResult, type SequentialDecision, type SeriesConvergenceOptions, type SeriesConvergenceResult, type SeriesDistribution, ServedCrossFamilyError, type ServedModelCheck, type ServedModelPolicy, type ServedModelVerdict, type Severity, type SignTestAlternative, type SingleBackendDivergence, type SingleBackendReport, type Span, type SpanFilter, type SpanHandle, type SpanMatchRecord, type SplitCoverage, type SteeringBundle, type SteeringChange, type SteeringOptimizationResult, type SteeringOptimizationRow, type StopDecision, type StrategyChecker, type StreamingDetector, type SummaryTable, type SummaryTableOptions, type SummaryTableRow, type SynthesisTarget, TRACE_ANALYST_TRUNCATION_MARKER_PREFIX, type TaskGold, type TaskHeadroom, type ToolCallEventLike, type ToolSpan, type ToolStats, type ToolUseMetrics, type ToolUseOptions, type TraceAnalysisEngine, type TraceAnalysisEngineResult, type TraceAnalysisStore, type TraceAnalystByteBudgets, type TraceAnalystDefinition, type TraceAnalystFilters, type TraceAnalystSpan, type TraceAnalystSpanKind, type TraceAnalystSpanStatus, type TraceAnalystTraceSummary, type TraceContract, TraceEmitter, type TraceEvent, type TraceInsightReadiness, type TraceInsightSuite, type TraceStore, type Trajectory, type TrajectoryStep, type TreatmentGate, type TreatmentGateInput, type TreatmentGateOptions, type TrialTrace, UNKNOWN_MODEL, type UserQuestion, VERIFICATION_STRATEGIES, VERIFICATION_STRATEGY_SOURCES, ValidationError, type VerbosityBiasResult, type Verdict, type VerdictCacheStore, type VerdictCertification, type Verification, type VerificationReport, type VerificationStrategyProfile, type VerificationStrategySource, type VerifyFn, type VerifyOptions, type ViewSpansResult, type ViewTraceOversized, type ViewTraceResult, type ViteDeployRunnerInput, WILCOXON_EXACT_MAX_N, type WeightedCompositeInput, type WeightedCompositeResult, type WilcoxonSignedRankResult, type WranglerDeployRunnerInput, acquisitionPlansForKnowledgeGaps, admitPolicyEdit, agentProfileCellHashMaterial, agentProfileCellKey, agentProfileHash, agentProfileId, aggregateJudgeVerdicts, aggregateRunScore, analystFindingDigest, analystRunDigest, analystRunToFeedbackTrajectory, analystRunToReviewRequests, analyzeAntiSlop, analyzeRuns, analyzeSeries, analyzeTraces, applyPolicyEditToSurface, argHash, assertCapabilityHeadroom, assertCrossFamily, assertCrossFamilyServed, assertNoHiddenLeak, assertNoJudgeVerdict, assertProductBenchmarkRun, assertRealBackend, assertRunCaptured, assertServedModel, assertServedModels, assertSingleBackend, assignFeedbackSplit, attest, benjaminiHochberg, blendHeldout, blockingKnowledgeEval, bonferroni, bootstrapCi, budgetBreachView, buildAgentProfileCell, buildDefaultAnalystRegistry, buildEquivalenceRecord, buildReflectionPrompt, buildTraceInsightContext, buildTraceInsightPrompt, buildTrajectory, calibrateJudge, calibrateJudgeContinuous, canonicalJson, capabilityHeadroom, captureFetchToRawSink, certificationEvidenceDigest, checkCanaries, checkServedModel, checkTraceContracts, clamp01, classifyFailure, cliffsDelta, cohensD, comparePairedArms, completionVerdict, computeExperimentStats, computeFindingId, computePolicyEditId, computeToolUseMetrics, confidenceInterval, contentHash, continuousAgreement, controlRunToFeedbackTrajectory, corpusInterRaterAgreement, corpusInterRaterAgreementFromJudgeScores, costForTokenPricing, costForUsage, costReceiptFromLlm, costReceiptFromLlmError, createAntiSlopJudge, createBoundedTraceAnalysisStore, createChatClient, createDspyRlmTraceEngine, createFeedbackTrajectory, createLlmCorrectnessChecker, createLlmReviewer, createTraceAnalyst, decidePairedPromotion, defaultBlendWeights, defineAgentEval, defineEquivalenceCheck, deployGateLayer, describeTraceInsightScope, diffFindings, diffScorecard, discoverPersonas, domainEvidencePattern, dominates, eProcess, ensembleJudge, equivalenceVerdict, errorStreakDetector, estimateCost, estimateTokens, evaluateActionPolicy, evaluateInterimReleaseConfidence, evaluateOracles, evaluateReleaseConfidence, expandProfileAxes, exportProductBenchmark, exportProductBenchmarkRuns, exportRunAsOtlp, extractErrorCount, extractProducedState, extractUsage, extractUsageFromSse, failureClusterView, feedbackTrajectoriesToDatasetScenarios, feedbackTrajectoriesToOptimizerRows, feedbackTrajectoryToOptimizerRow, fileVerdictCache, formatScorecardDiff, gainHistogram, gateTreatmentApplied, gradeOnHidden, gradeSemanticStatus, groupRunsByAgentProfileCell, harnessAxisOf, hashContent, hashJson, hiddenGrade, holm, improvementVerdict, inMemoryReviewStore, inferDomainKeywords, interRaterReliability, interpretCliffs, iqr, isBinaryOutcomeVector, isJudgeSpan, isLlmSpan, isModelPriced, isPolicyEdit, isRunRecord, isToolSpan, isTransientLlmError, jsonShape, jsonlReviewStore, jsonlRunRecordBackend, judgeAgreementView, judgeFamily, judgeSpans, knowledgeReadinessTracePayload, leaderboard, llmJudge, loadScorecard, localCommandRunner, makeFinding, makePolicyEdit, makePolicyEditCandidateRecord, makeProposalFinding, manifestContentDigest, mannWhitneyU, maximumChargeForLlmRequest, mcnemar, mcnemarPower, mcnemarRequiredN, minimumPairsForPairedDeltaTest, mintRolloutRows, modelHasSnapshot, modelPriceKey, mulberry32, notBlocked, objectiveEval, observeAll, otlpTextToTraceAnalysisStore, pairArms, pairRunRecords, pairedBinaryScale, pairedBootstrap, pairedCohensDz, pairedDeltaTest, pairedDeltaTieFraction, pairedEvalueSequence, pairedMde, pairedRiskDifference, pairedRiskDifferenceExact, pairedRiskDifferenceScore, pairedSignTest, pairedTTest, paretoChart, paretoFrontier, parseReflectionResponse, parseRunRecordSafe, partialCredit, partitionHeldOut, passAtK, pearsonR, policyEditFromFinding, policyEditsFromFindings, preflightModels, productBenchmarkRepoIdentity, index_d_exports as profile, projectRuntimeTrajectoryEvidence, proposeSynthesisTargets, ranks, readProductBenchmarkManifest, recordRuns, recordRunsToScorecard, redTeamDataset, redTeamReport, redactString, regexMatches, renderPreferenceMemoryMarkdown, repeatedActionDetector, requiredPairedSampleSize, requiredSampleSize, resolveModelPricing, resolveSeat, roundTripRunRecord, routeFields, runAgentControlLoop, runBoundedProcess, runCampaign, runCanaries, runCounterfactual, runEquivalenceCheck, runEvalCampaign, runIntentMatchJudge, runKeywordCoverageJudge, runKeywordCoverageJudgeUrl, runProposeReview, runProposeReviewAsControlLoop, runSemanticConceptJudge, runTaskScore, runsForScenario, scoreKnowledgeReadiness, scorePolicyEditReadiness, scoreRedTeamOutput, scoreTraceInsightReadiness, seatPresets, selfImprove, servedModelAcceptable, spearmanR, stripFencedJson, subjectiveEval, summarizeBackendIntegrity, summarizeNumberSeries, summarizePreferenceMemory, summaryTable, textInSnapshot, toAgentProfileJson, tokenizeDomainWords, toolSpansToTraceAnalysisStore, toolWasteView, traceContract, transientDispatchFailure, urlContains, userQuestionsForKnowledgeGaps, validatePolicyEdit, validatePolicyEditCandidateRecord, validateRunRecord, verbosityBias, verifyAgentProfileCell, verifyAttestation, verifyCompletion, viteDeployRunner, weightedComposite, weightedMean, wilcoxonSignedRank, wilson, withAssignedFeedbackSplit, withHeldoutBlend, withJudgeRetry, wranglerDeployRunner };
|
|
2908
3047
|
//# sourceMappingURL=index.d.ts.map
|