@tangle-network/agent-eval 0.183.0 → 0.184.0

package/dist/{matrix-CyhW-vgJ.d.ts → matrix-CPKsRY8Y.d.ts}

@@ -1,5 +1,5 @@
 import { p as CostProvenance } from "./cost-ledger-DbQdN3nO.js";
-import { w as JudgeScore } from "./types-CS0qk_Yp.js";
+import { w as JudgeScore } from "./types-BTRcIqFr.js";
 import { o as MatrixResult } from "./index-DNgf5gyG.js";
 import { AgentProfile } from "@tangle-network/agent-interface";
 //#region src/multishot/types.d.ts
@@ -398,4 +398,4 @@ interface RunMultishotMatrixResult {
 declare function runMultishotMatrix<TPersona extends MultishotPersona>(opts: RunMultishotMatrixOptions<TPersona>): Promise<RunMultishotMatrixResult>;
 //#endregion
 export { MultishotToolExecutor as A, MultishotFatalToolError as C, MultishotShape as D, MultishotResult as E, assertMultishotShotResult as F, MultishotTransportRequest as M, MultishotTransportResponse as N, MultishotShotResultError as O, MultishotTransportToolCall as P, MultishotDriverEmptyError as S, MultishotPersona as T, JudgeRunResult as _, MultishotCellOutput as a, runJudge as b, RunMultishotMatrixResult as c, MultishotShot as d, RunMultishotOptions as f, JudgeDimension as g, JudgeConfig as h, ConversationJudgeInput as i, MultishotTransport as j, MultishotToolDefinition as k, computeCellComposite as l, DEFAULT_JUDGE_MODEL as m, CellCompositeInput as n, MultishotJudges as o, runMultishot as p, CellCompositeScore as r, RunMultishotMatrixOptions as s, ArtifactJudgeInput as t, runMultishotMatrix as u, renderDimensions as v, MultishotMessage as w, MultishotArtifact as x, renderJsonFooter as y };
-//# sourceMappingURL=matrix-CyhW-vgJ.d.ts.map
+//# sourceMappingURL=matrix-CPKsRY8Y.d.ts.map

package/dist/{matrix-CyhW-vgJ.d.ts.map → matrix-CPKsRY8Y.d.ts.map}

@@ -1 +1 @@
-{"version":3,"file":"matrix-CyhW-vgJ.d.ts","names":[],"sources":["../src/multishot/types.ts","../src/multishot/judges.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;UAIiB;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;EAIA;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;;EAEE,WAAW;EACX,SAAS;MAER;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;cCvI/D;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UCxK/B,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;;;EAIhB,iBAAiB;;;EAGjB,gBAAgB;EAChB,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UClFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,gBAAgB;;EAEhB,iBAAiB;;;EAGjB,gBAAgB;;;;;;;;;;;;;;;EAehB,UAAU,cAAc;;;;;UAMT;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}
+{"version":3,"file":"matrix-CPKsRY8Y.d.ts","names":[],"sources":["../src/multishot/types.ts","../src/multishot/judges.ts","../src/multishot/multishot.ts","../src/multishot/matrix.ts"],"mappings":";;;;;UAIiB;EACf;EACA;EACA;EACA,YAAY;IAAQ;IAAY;IAAc,MAAM;;;UAGrC;EACf;EACA;EACA;IAAc;IAAc,MAAM;;EAClC;;UAGe;EACf,YAAY;EACZ,WAAW;EACX;EACA;;;EAGA;;;;;;;;EAQA,iBAAiB;;UAGF;EACf;EACA;IACE;IACA;IACA,YAAY;;;;;;UAOC;EACf;EACA,UAAU,MAAM;EAChB,QAAQ;EACR;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA;IAAY;IAAc;;;UAGX;EACf;IAAW;IAAyB,aAAa;;EACjD;IAAU;IAAwB;;;;EAGlC;;;;EAIA;;;;;;;;KASU,sBACV,KAAK,8BACF,QAAQ;KAED,yBACV,MAAM,yBACN;;EAEE,WAAW;EACX,SAAS;MAER;EAAU;EAAiB;;UAEf;;EAEf;;GAEC;;;;;;;;UASc,eAAe,iBAAiB;;EAE/C,eAAe,SAAS;;;EAGxB,2BAA2B,SAAS;;cAGzB,kCAAkC;WACjB;EAA5B,YAA4B;;cAMjB,gCAAgC;EAC3C,YAAY;;cAMD,iCAAiC;EAC5C,YAAY;;;;;;;;;;;;;;;;;;;iBAyBE,0BAA0B,yBAAyB,SAAS;;;cCvI/D;UAEI;;EAEf;;EAEA;;UAGe,YAAY;;EAE3B;;;EAGA,WAAW;;EAEX;;EAEA,YAAY;;EAEZ;;;EAGA,cAAc,OAAO;;EAErB;;UAGe;;EAEf,OAAO;;EAEP,MAAM;;iBAGc,SAAS,QAC7B,OAAO,YAAY,SACnB,OAAO,SACN,QAAQ;;;iBAsIK,iBAAiB,eAAe;;iBAKhC,iBAAiB,eAAe;;;UCxK/B,oBAAoB,iBAAiB;EACpD,SAAS;EACT,SAAS;;;EAGT,QAAQ,eAAe;;EAEvB,QAAQ;;EAER,gBAAgB,eAAe;;;;EAI/B,mBAAmB;EACnB;EACA;EACA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,gBAAgB;;;;EAIhB,iBAAiB;;;EAGjB,gBAAgB;EAChB,SAAS;;;;;;;;;;;KAYC,cAAc,iBAAiB,qBACzC,MAAM,oBAAoB,cACvB,QAAQ;;;;;;;;;;;;;;;;;iBA2BS,aAAa,iBAAiB,kBAClD,MAAM,oBAAoB,YACzB,QAAQ;;;UClFM,uBAAuB,iBAAiB;EACvD,YAAY;EACZ,SAAS;;UAGM,mBAAmB,iBAAiB;EACnD,UAAU;EACV,SAAS;;UAGM,gBAAgB,iBAAiB;;EAEhD,cAAc,YAAY,uBAAuB;;EAEjD,aAAa,YAAY,mBAAmB;;EAE5C,iBAAiB,YAAY,mBAAmB;;EAEhD;;EAEA;;UAGe;EACf;EACA,cAAc;EACd;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;EAEF;IACE,aAAa,MAAM;MAAe;MAAc;;IAChD;;;UAIa,0BAA0B,iBAAiB;;EAE1D,UAAU;IAAQ;IAAY,OAAO;;;EAErC,UAAU;;;EAGV,QAAQ,eAAe;;EAEvB,QAAQ,gBAAgB;;EAExB,QAAQ;;EAER,gBAAgB,eAAe;;EAE/B,mBAAmB;;EAEnB;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;;EAIA,gBAAgB;;EAEhB,iBAAiB;;;EAGjB,gBAAgB;;;;;;;;;;;;;;;EAehB,UAAU,cAAc;;;;;UAMT;EACf;EACA;EACA;;UAoBe;EACf,cAAc;;EAEd,cAAc,cAAc;;EAE5B,iBAAiB,cAAc;;;;;;;iBAQjB,qBAAqB,OAAO;EAC1C;EACA;EACA;EACA;;UAuBe;EACf,QAAQ,aAAa;;iBAGD,mBAAmB,iBAAiB,kBACxD,MAAM,0BAA0B,YAC/B,QAAQ"}

package/dist/meta-eval/index.d.ts

@@ -5,7 +5,7 @@ import { r as LedgerHash } from "../canonical-CFpojCN5.js";
 import { s as TraceStore } from "../store-BErPvYBr.js";
 import { a as OutcomeFilter, i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, s as OutcomeStoreError, t as DeploymentOutcome } from "../outcome-store-CNt4iZ67.js";
 import { a as ComputedInterval } from "../ast-hI-vjW6J.js";
-import { a as RubricPredictiveValidityReport, c as CorrelationInterval, i as RubricPredictiveValidityInput, l as OutcomeReduction, n as RubricOutcomeExclusion, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity, t as OutcomeMetricSpec } from "../rubric-predictive-validity-Bmj2_cll.js";
+import { a as RubricPredictiveValidityReport, c as CorrelationInterval, i as RubricPredictiveValidityInput, l as OutcomeReduction, n as RubricOutcomeExclusion, o as RubricRanking, r as RubricOutcomePair, s as rubricPredictiveValidity, t as OutcomeMetricSpec } from "../rubric-predictive-validity-CTU8y8Cv.js";
 import { z } from "zod";
 //#region src/meta-eval/correlation-study.d.ts
 interface EvalMetricSpec {
@@ -216,6 +216,58 @@ interface EvaluatorAdmissionReport {
  * The execution owner enforces auditor separation, fresh sampling, and evidence authenticity.
  */
 declare function auditEvaluator(input: EvaluatorAuditInput): EvaluatorAdmissionReport;
+declare const probabilityAuditSchema: z.ZodObject<{
+  evaluatorDigest: z.ZodString;
+  population: z.ZodString;
+  samplingFrame: z.ZodString;
+  authority: z.ZodObject<{
+    evaluatorAuthorId: z.ZodString;
+    auditorId: z.ZodString;
+    independenceEvidenceRef: z.ZodString;
+  }, z.core.$strict>;
+  policy: z.ZodObject<{
+    confidence: z.ZodNumber;
+    maxFalseAcceptanceRate: z.ZodNumber;
+    maxFalseRejectionRate: z.ZodNumber;
+  }, z.core.$strict>;
+  thresholds: z.ZodObject<{
+    rejectAtOrBelow: z.ZodNumber;
+    acceptAtOrAbove: z.ZodNumber;
+  }, z.core.$strict>;
+  observations: z.ZodArray<z.ZodObject<{
+    id: z.ZodString;
+    independentUnitId: z.ZodString;
+    evidenceRef: z.ZodString;
+    expected: z.ZodEnum<{
+      accept: "accept";
+      reject: "reject";
+    }>;
+    exposure: z.ZodEnum<{
+      development: "development";
+      fresh: "fresh";
+    }>;
+    acceptProbability: z.ZodNullable<z.ZodNumber>;
+  }, z.core.$strict>>;
+}, z.core.$strict>;
+/** A frozen caller-owned rule over P(accept), not the provider's entropy/confidence statistic. */
+type ProbabilityPolicyAuditInput = z.infer<typeof probabilityAuditSchema>;
+interface ProbabilityPolicyAuditReport extends EvaluatorAdmissionReport {
+  probabilityPolicy: {
+    sourceEvaluatorDigest: LedgerHash;
+    thresholds: ProbabilityPolicyAuditInput['thresholds'];
+    /** Retain probabilities by observation id so the reported mapping can be reproduced. */
+    observations: Array<{
+      id: string;
+      acceptProbability: number | null;
+    }>;
+  };
+}
+/**
+ * Audit a fixed probability-to-decision rule with the existing independent-unit error bounds.
+ * No threshold search, model calls, promotion, or assumption of calibrated probabilities.
+ * Null scores and the open interval between thresholds remain unknown, never correct by default.
+ */
+declare function auditProbabilityPolicy(input: ProbabilityPolicyAuditInput): ProbabilityPolicyAuditReport;
 //#endregion
 //#region src/meta-eval/plants.d.ts
 /**
@@ -601,5 +653,5 @@ interface EvalHealthStamp {
  */
 declare function evalHealthStamp(report: SentinelReport): EvalHealthStamp;
 //#endregion
-export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, type CalibrationResult, type CandidateScore, CatchRateReport, CatchRateStatus, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorErrorRate, EvidencePerturbation, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomeMetricSpec, OutcomePair, type OutcomeReduction, OutcomeStore, OutcomeStoreError, PerturbationKind, PerturbedField, PerturbedPlant, Plant, PlantEvidence, PlantExpectation, PlantKind, PlantKindCounts, PlantManifest, PlantOutcome, type PositionalBiasResult, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SeedPlantsOptions, SeededGradingSet, type SelfPreferenceResult, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, UnseededRejection, type VerbosityBiasResult, auditEvaluator, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
+export { CalibrationBin, CalibrationOptions, CalibrationPair, CalibrationReport, type CalibrationResult, type CandidateScore, CatchRateReport, CatchRateStatus, type ContinuousAgreement, type ContinuousAgreementOptions, type ContinuousCalibrationResult, type CorrelationInterval, CorrelationResult, CorrelationStudyOptions, CorrelationStudyResult, DeploymentOutcome, EvalHealthStamp, EvalMetricSpec, type EvaluatorAdmissionPolicy, type EvaluatorAdmissionReport, type EvaluatorAuditInput, type EvaluatorAuditObservation, type EvaluatorErrorRate, EvidencePerturbation, FileSystemOutcomeStore, FileSystemOutcomeStoreOptions, type GoldenItem, InMemoryOutcomeStore, JudgeSentinelOptions, OutcomeFilter, OutcomeMetricSpec, OutcomePair, type OutcomeReduction, OutcomeStore, OutcomeStoreError, PerturbationKind, PerturbedField, PerturbedPlant, Plant, PlantEvidence, PlantExpectation, PlantKind, PlantKindCounts, PlantManifest, PlantOutcome, type PositionalBiasResult, type ProbabilityPolicyAuditInput, type ProbabilityPolicyAuditReport, RubricOutcomeExclusion, RubricOutcomePair, RubricPredictiveValidityInput, RubricPredictiveValidityReport, RubricRanking, SeedPlantsOptions, SeededGradingSet, type SelfPreferenceResult, SentinelMetricName, SentinelMetrics, SentinelReport, SentinelSetOptions, SentinelSnapshot, SentinelStore, SentinelThresholds, SentinelTrend, SnapshotMeta, UnseededRejection, type VerbosityBiasResult, auditEvaluator, auditProbabilityPolicy, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
 //# sourceMappingURL=index.d.ts.map

package/dist/meta-eval/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/evaluator-admission.ts","../../src/meta-eval/plants.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;;;UAoBiB;EACf;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;IAAgB;IAAe;;;EAE/B;;UAGe;EACf,OAAO;;EAEP,eAAe,MACb;IACE;IACA;;EAGJ;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;;EAGA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UChEM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;;iBA0CK,qBACd,qBAAqB,mBACrB,oBACA,uBACA,UAAS,qBACR;;;cC9FG,cAAY,EAAA;;;;GAMP,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;;;;;;;;;;;;;;;GASZ,EAAA,KAAA;;KAGC,4BAA4B,EAAE,aAAa;cAEjD,aAAW,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAeN,EAAA,KAAA;KAEC,sBAAsB,EAAE,aAAa;UAEhC;;EAEf;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA,UAAU;EACV;EACA;;UAGe;EACf,iBAAiB;EACjB,cAAc;EACd,aAAa;EACb,cAAc;EACd;EACA;EACA,WAAW;EACX,QAAQ;;EAER;EACA;EACA;EACA;EACA,cAAc;EACd;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF,YAAY;IAAQ;IAAY;IAA2B;;EAC3D,iBAAiB;EACjB,gBAAgB;;;;;;;iBAoEF,eAAe,OAAO,sBAAsB;;;;;;;;KCrGhD;;;;;;;;;;KAkBA;UAOK;;EAEf;EACA,MAAM;;EAEN,MAAM;;EAEN,iBAAiB;;;;;;;;;;;;;;iBAeH,YAAY;EAC1B;EACA,MAAM;EACN,MAAM;EACN,iBAAiB;IACf;;;;;UAqDa;;EAEf;;EAEA;;;KAIU;;KAGA;UAEK;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;;;;;EAML;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsJc,gBAAgB,UAAU,gBAAgB;UA8BzC;EACf,OAAO;;EAEP,UAAU;EACV,OAAO;EACP,KAAK;EACL;EACA;;;;;;;;;;;;;;;;;;;;;;iBAuBc,oBAAoB;EAClC;EACA;EACA,UAAU;;EAEV;EACA;IACE;UA0Ba;;;;;;EAMf,MAAM;;EAEN;;EAEA;;EAEA;;EAEA,QAAQ;;UAGO;;;;;;EAMf,OAAO;EACP,UAAU;;UAGK;;;;;EAKf;;;;;;EAMA;;;;;;;;;;;iBAYc,WACd,kBAAkB,cAClB,iBAAiB,SACjB,UAAS,oBACR;;;;;;UAqGc;EACf;EACA;;;;;;KAOU;UAEK;EACf;EACA;EACA;EACA;;EAEA;;;;;;;;UASe;EACf;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;;EAER;EACA;EACA;EACA;;;;;EAKA;;EAEA;;EAEA,QAAQ,QAAQ,OAAO,WAAW;;EAElC;;EAEA;EACA,UAAU;;;;;;;;;;iBAWI,UACd,kBAAkB,gBAClB,UAAU,gBACT;;;cC/lBG;KAEM,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}
+{"version":3,"file":"index.d.ts","names":[],"sources":["../../src/meta-eval/correlation-study.ts","../../src/meta-eval/calibration.ts","../../src/meta-eval/evaluator-admission.ts","../../src/meta-eval/plants.ts","../../src/meta-eval/sentinel.ts"],"mappings":";;;;;;;;;;UAoBiB;EACf;;;EAGA,WAAW,KAAK,KAAK,OAAO,eAAe;;UAG5B;EACf;EACA;;UAGe;EACf;EACA;EACA;EACA;EACA;;EAEA;IAAe;IAAe;;;EAE9B;IAAgB;IAAe;;;EAE/B;;UAGe;EACf,OAAO;;EAEP,eAAe,MACb;IACE;IACA;;EAGJ;EACA;;UAGe;;EAEf;;EAEA,gBAAgB;;EAEhB;;EAEA;;;EAGA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,aAAa,kBACb,8BACA,UAAS,0BACR,QAAQ;;;UChEM;EACf;EACA;EACA;EACA;EACA;;EAEA;;UAGe;EACf;EACA;EACA;EACA,MAAM;;EAEN;;EAEA;;UAGe;;EAEf;;EAEA;;EAEA;IAAU;IAAY;;;UAGP;EACf;EACA;;iBAGoB,iBACpB,YAAY,YACZ,cAAc,cACd,YAAY,gBACZ,uBACA,UAAS,qBACR,QAAQ;;iBA0CK,qBACd,qBAAqB,mBACrB,oBACA,uBACA,UAAS,qBACR;;;cC9FG,cAAY,EAAA;;;;GAMP,EAAA,KAAA;;KAGC,2BAA2B,EAAE,aAAa;cAEhD,mBAAiB,EAAA;;;;;;;;;;;;;;;;;GASZ,EAAA,KAAA;;KAGC,4BAA4B,EAAE,aAAa;cAEjD,aAAW,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAeN,EAAA,KAAA;KAEC,sBAAsB,EAAE,aAAa;UAEhC;;EAEf;EACA;;EAEA;;EAEA;EACA;;EAEA;;EAEA,UAAU;EACV;EACA;;UAGe;EACf,iBAAiB;EACjB,cAAc;EACd,aAAa;EACb,cAAc;EACd;EACA;EACA,WAAW;EACX,QAAQ;;EAER;EACA;EACA;EACA;EACA,cAAc;EACd;IACE;IACA;IACA;IACA;IACA;IACA;;EAEF,YAAY;IAAQ;IAAY;IAA2B;;EAC3D,iBAAiB;EACjB,gBAAgB;;;;;;;iBAoEF,eAAe,OAAO,sBAAsB;cA2GtD,wBAAsB,EAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAG1B,EAAA,KAAA;;KAGU,8BAA8B,EAAE,aAAa;UAExC,qCAAqC;EACpD;IACE,uBAAuB;IACvB,YAAY;;IAEZ,cAAc;MAAQ;MAAY;;;;;;;;;iBAStB,uBACd,OAAO,8BACN;;;;;;;;KCxOS;;;;;;;;;;KAkBA;UAOK;;EAEf;EACA,MAAM;;EAEN,MAAM;;EAEN,iBAAiB;;;;;;;;;;;;;;iBAeH,YAAY;EAC1B;EACA,MAAM;EACN,MAAM;EACN,iBAAiB;IACf;;;;;UAqDa;;EAEf;;EAEA;;;KAIU;;KAGA;UAEK;;EAEf,UAAU;;EAEV,OAAO;;EAEP,KAAK;;;;;;EAML;;EAEA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAsJc,gBAAgB,UAAU,gBAAgB;UA8BzC;EACf,OAAO;;EAEP,UAAU;EACV,OAAO;EACP,KAAK;EACL;EACA;;;;;;;;;;;;;;;;;;;;;;iBAuBc,oBAAoB;EAClC;EACA;EACA,UAAU;;EAEV;EACA;IACE;UA0Ba;;;;;;EAMf,MAAM;;EAEN;;EAEA;;EAEA;;EAEA,QAAQ;;UAGO;;;;;;EAMf,OAAO;EACP,UAAU;;UAGK;;;;;EAKf;;;;;;EAMA;;;;;;;;;;;iBAYc,WACd,kBAAkB,cAClB,iBAAiB,SACjB,UAAS,oBACR;;;;;;UAqGc;EACf;EACA;;;;;;KAOU;UAEK;EACf;EACA;EACA;EACA;;EAEA;;;;;;;;UASe;EACf;EACA;EACA;;EAEA;;UAGe;EACf,QAAQ;;EAER;EACA;EACA;EACA;;;;;EAKA;;EAEA;;EAEA,QAAQ,QAAQ,OAAO,WAAW;;EAElC;;EAEA;EACA,UAAU;;;;;;;;;;iBAWI,UACd,kBAAkB,gBAClB,UAAU,gBACT;;;cC/lBG;KAEM,6BAA6B;UAExB;;EAEf;;EAEA;;EAEA;;UAGe;;EAEf;EACA;;EAEA;;;;;;;EAOA,SAAS;;;UAIM;EACf;EACA;EACA;;;iBAcc,yBAAyB,UAAU,kBAAkB;;;;;;iBAiCrD,wBACd,QAAQ,oBAAoB,6BAC5B,MAAM,eACL;;;;;;;;iBAmBa,sBACd,QAAQ,wBAAwB,qBAChC,MAAM,eACL;UAYc;;EAEf;;;;;;;;;iBAUc,wBACd,QAAQ,kBACR,QAAQ,cACR,MAAM,cACN,UAAS,qBACR;UA4Cc;EACf,OAAO,UAAU,mBAAmB;;EAEpC,QAAQ,mBAAmB,QAAQ;;iBAGrB,sBAAsB,UAAS,qBAA0B;;;;;;;iBAqBzD,kBAAkB,eAAe;UA0ChC;;EAEf;;EAEA;;EAEA;;EAEA;;UAGe;;EAEf;EACA,aAAa;;EAEb,cAAc;;UAGC;EACf;EACA,QAAQ;;EAER,OAAO;;EAEP;;EAEA;;EAEA;EACA;EACA;;UAGe;EACf,UAAU;EACV;EACA;;;;;;;EAOA;;iBAKc,oBACd,SAAS,oBACT,MAAM,uBACL;UA6Gc;EACf;EACA;;;;;;;;;;iBAWc,gBAAgB,QAAQ,iBAAiB"}

package/dist/meta-eval/index.js

@@ -1,8 +1,8 @@
 import { n as CaptureIntegrityError, s as ValidationError } from "../errors-Dngq5h35.js";
-import { a as hashCanonical, i as compareCodeUnits } from "../canonical-DPyQ_rpt.js";
+import { a as hashCanonical, i as compareCodeUnits } from "../canonical-BsZVgBhW.js";
 import { t as mulberry32 } from "../random-Dn5fPWkt.js";
 import { a as selfPreference, i as positionalBias, n as calibrateJudgeContinuous, o as verbosityBias, r as continuousAgreement, t as calibrateJudge } from "../judge-calibration-DYmaBtJr.js";
-import { u as runMetricExtractor } from "../query-D1nLIKt7.js";
+import { u as runMetricExtractor } from "../query-DlV0pdx4.js";
 import { r as computeInterval } from "../ast-CP9ae9B0.js";
 import { t as analyzeSeries } from "../series-convergence-CjO2QdRW.js";
 import { a as reduceOutcomeMetric, i as hasVariation, n as assertUniqueObservationIds, o as validateObservationOptions, r as correlationSummary, t as rubricPredictiveValidity } from "../rubric-predictive-validity-CCK-1B7w.js";
@@ -375,6 +375,55 @@ function auditEvaluator(input) {
 		reportDigest: hashCanonical(body)
 	};
 }
+const probabilityThresholdsSchema = z.object({
+	rejectAtOrBelow: z.number().finite().min(0).max(1),
+	acceptAtOrAbove: z.number().finite().min(0).max(1)
+}).strict().refine((value) => value.rejectAtOrBelow < value.acceptAtOrAbove, { message: "rejection and acceptance thresholds must not overlap" });
+const probabilityObservationSchema = observationSchema.omit({ observed: true }).extend({ acceptProbability: z.number().finite().min(0).max(1).nullable() }).strict();
+const probabilityAuditSchema = inputSchema.extend({
+	thresholds: probabilityThresholdsSchema,
+	observations: z.array(probabilityObservationSchema)
+});
+/**
+* Audit a fixed probability-to-decision rule with the existing independent-unit error bounds.
+* No threshold search, model calls, promotion, or assumption of calibrated probabilities.
+* Null scores and the open interval between thresholds remain unknown, never correct by default.
+*/
+function auditProbabilityPolicy(input) {
+	const parsed = probabilityAuditSchema.safeParse(input);
+	if (!parsed.success) throw new ValidationError(`invalid probability policy audit: ${parsed.error.message}`);
+	const { thresholds, observations, ...source } = parsed.data;
+	observations.sort((a, b) => compareCodeUnits(a.id, b.id));
+	const probabilityPolicy = {
+		sourceEvaluatorDigest: source.evaluatorDigest,
+		thresholds,
+		observations: observations.map(({ id, acceptProbability }) => ({
+			id,
+			acceptProbability
+		}))
+	};
+	const { reportDigest: _baseDigest, ...baseReport } = auditEvaluator({
+		...source,
+		evaluatorDigest: hashCanonical({
+			kind: "probability-policy-v1",
+			sourceEvaluatorDigest: source.evaluatorDigest,
+			thresholds
+		}),
+		observations: observations.map(({ acceptProbability, ...observation }) => ({
+			...observation,
+			observed: acceptProbability === null ? "unknown" : acceptProbability >= thresholds.acceptAtOrAbove ? "accept" : acceptProbability <= thresholds.rejectAtOrBelow ? "reject" : "unknown"
+		}))
+	});
+	const body = {
+		...baseReport,
+		inputDigest: hashCanonical(parsed.data),
+		probabilityPolicy
+	};
+	return {
+		...body,
+		reportDigest: hashCanonical(body)
+	};
+}
 //#endregion
 //#region src/meta-eval/plants.ts
 /**
@@ -1130,6 +1179,6 @@ function evalHealthStamp(report) {
 	};
 }
 //#endregion
-export { FileSystemOutcomeStore, InMemoryOutcomeStore, OutcomeStoreError, auditEvaluator, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
+export { FileSystemOutcomeStore, InMemoryOutcomeStore, OutcomeStoreError, auditEvaluator, auditProbabilityPolicy, calibrateJudge, calibrateJudgeContinuous, calibrationCurve, calibrationFromPairs, catchRate, continuousAgreement, correlationStudy, definePlant, evalHealthStamp, fileSentinelStore, inMemorySentinelStore, judgeSentinelReport, perturbEvidence, plantByPerturbation, positionalBias, rubricPredictiveValidity, seedPlants, selfPreference, snapshotFromAgreement, snapshotFromCalibration, snapshotFromSentinelSet, validateSentinelSnapshot, verbosityBias };
 
 //# sourceMappingURL=index.js.map