@tangle-network/agent-eval 0.216.1 → 0.218.1

package/dist/rl.d.ts

@@ -1,20 +1,18 @@
-import { v as Span } from "./schema-2g7trUwH.js";
-import { a as RunRecord, c as RunSplitTag } from "./run-record-C1E24G2A.js";
-import { a as PairedPromotionDecision, d as PairedBootstrapResult } from "./paired-promotion-decision-DPKz8ZnF.js";
-import { a as CampaignResult } from "./types-Cq8mSbq8.js";
-import { s as TraceStore } from "./store-C4avFjIi.js";
+import { _ as Span } from "./schema-D6EVKzTB.js";
+import { a as RunRecord, c as RunSplitTag } from "./run-record-CaxbLq6Y.js";
+import { a as PairedPromotionDecision, u as PairedBootstrapResult } from "./paired-promotion-decision-CQMJ5xGN.js";
+import { a as CampaignResult } from "./types-DtkO1G9L.js";
 import { a as MintedRolloutLine, b as RolloutSplit, t as ChatMessage } from "./schema-C-ziyJD7.js";
-import { t as MintRolloutOptions } from "./mint-DNU8QPG4.js";
-import { s as VerificationReport } from "./multi-layer-verifier-BUaQ4C17.js";
-import { a as detectRewardHacking, c as VerifiableRewardSource, d as filterDeterministicallyRewarded, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-RVdZ6Kkt.js";
+import { t as MintRolloutOptions } from "./mint-BRZXuwRi.js";
+import { a as detectRewardHacking, c as VerifiableRewardSource, i as RewardHackingSignal, l as extractVerifiableReward, n as RewardHackingFinding, o as VerifiableReward, r as RewardHackingReport, s as VerifiableRewardExtractionOptions, t as DetectRewardHackingInput, u as extractVerifiableRewardsFromRecords } from "./reward-hacking-E52wq86n.js";
 import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "./outcome-store-CNt4iZ67.js";
-import { a as SystemOneReviewReport, t as PreparedSystemOneReview } from "./systemone-review-CUVcQJu6.js";
-import { n as TraceQuestionOutcome, r as TraceQuestionSpec } from "./trace-questions-CM2CAocU.js";
-import { t as InterimReleaseConfidence } from "./sequential-BhsrMupG.js";
+import { a as SystemOneReviewReport, t as PreparedSystemOneReview } from "./systemone-review-CwE6JGIp.js";
+import { n as TraceQuestionOutcome, r as TraceQuestionSpec } from "./trace-questions-xxwgsuMi.js";
+import { t as InterimReleaseConfidence } from "./sequential-KwGspfeR.js";
 import { t as AdversarialMutation } from "./adversarial-smnADNFS.js";
-import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-CEcfku-T.js";
-import { o as TrainingExportOptions } from "./exporters-ChTnR90l.js";
-import { a as RubricPredictiveValidityReport, t as OutcomeMetricSpec } from "./rubric-predictive-validity-C03EC7ZE.js";
+import { a as SteeringChange, f as EvalCampaignOptions, h as runEvalCampaign, i as Researcher, n as ExperimentResult, p as EvalCampaignResult, r as FailureMode, t as ExperimentPlan } from "./researcher-GU1tV4dn.js";
+import { o as TrainingExportOptions } from "./exporters-Dk3oEbK_.js";
+import { a as RubricPredictiveValidityReport, t as OutcomeMetricSpec } from "./rubric-predictive-validity-DiwT8PbL.js";
 //#region src/rl/active-curriculum.d.ts
 /**
  * Adaptive curriculum: active scenario selection by variance.
@@ -31,13 +29,13 @@ import { a as RubricPredictiveValidityReport, t as OutcomeMetricSpec } from "./r
  * posterior, in `campaign/search-policy.ts`). Scenario generation belongs to
  * the adversarial primitive.
  */
-export interface CellObservation {
+interface CellObservation {
   variantId: string;
   scenarioId: string;
   /** Observed score in [0, 1]. */
   score: number;
 }
-export interface CurriculumAllocation {
+interface CurriculumAllocation {
   variantId: string;
   scenarioId: string;
   /** How many additional reps to run on this cell. */
@@ -45,7 +43,7 @@ export interface CurriculumAllocation {
   /** Strategy-specific reason for the allocation. */
   reason: string;
 }
-export interface VarianceCurriculumOptions {
+interface VarianceCurriculumOptions {
   /** Total reps to allocate across all cells. */
   budget: number;
   /**
@@ -60,17 +58,6 @@ export interface VarianceCurriculumOptions {
    */
   floorPerCell?: number;
 }
-/**
- * Variance-proportional allocation. For each cell, estimate variance from
- * past observations + a prior, then allocate the budget proportional to
- * (sqrt(variance) + 1/sqrt(n)) — a classical optimal-allocation rule
- * (Neyman 1934) that balances "explore noisy cells" with "explore
- * under-sampled cells."
- */
-export declare function varianceBasedCurriculum(observations: CellObservation[], candidateCells: Array<{
-  variantId: string;
-  scenarioId: string;
-}>, opts: VarianceCurriculumOptions): CurriculumAllocation[];
 //#endregion
 //#region src/rl/adaptation-eval.d.ts
 export interface AdaptationRunner<S> {
@@ -161,15 +148,15 @@ export declare function compareAdaptationCurves(a: AdaptationCurve, b: Adaptatio
 export declare function firstPassK(curve: AdaptationCurve, threshold?: number): number | null;
 //#endregion
 //#region src/rl/claim-integrity.d.ts
-export type ScopeValue = string | number | boolean | null;
-export type ScopeOperator = 'eq' | 'ne' | 'lt' | 'le' | 'gt' | 'ge' | 'in' | 'notIn';
+type ScopeValue = string | number | boolean | null;
+type ScopeOperator = 'eq' | 'ne' | 'lt' | 'le' | 'gt' | 'ge' | 'in' | 'notIn';
 /** One comparison of a claim parameter. A rule holds when all of its conditions hold. */
-export interface ScopeCondition {
+interface ScopeCondition {
   readonly parameter: string;
   readonly op: ScopeOperator;
   readonly value: ScopeValue | readonly ScopeValue[];
 }
-export interface ScopeRule {
+interface ScopeRule {
   /** Stable id, cited in evidence. */
   readonly id: string;
   readonly when: readonly ScopeCondition[];
@@ -180,7 +167,7 @@ export interface ScopeRule {
   /** Where the source says so: page and line. */
   readonly cite: string;
 }
-export interface SourceExcerpt {
+interface SourceExcerpt {
   readonly cite: string;
   readonly text: string;
 }
@@ -189,7 +176,7 @@ export interface SourceExcerpt {
  * the checker's author, and kept out of the agent's reach: it is a held-out check, and an agent that
  * could read it would learn which cases to hide.
  */
-export interface SourceScope {
+interface SourceScope {
   /** The statement or program id a claim names. */
   readonly statement: string;
   /** Citation of the source. */
@@ -215,7 +202,7 @@ export interface SourceScope {
   readonly checked?: readonly ScopeRule[];
 }
 /** One distinct claim an exact checker judged, with what the detector reads. */
-export interface IntegrityClaim {
+interface IntegrityClaim {
   /** Stable id of the distinct claim. */
   readonly id: string;
   /** The lane (run lineage) that filed it. */
@@ -232,14 +219,14 @@ export interface IntegrityClaim {
   /** The claim page. Read by the text signal only; never sent to a referee or to a System One model. */
   readonly text?: string | null;
 }
-export type IntegritySignalKind = 'exact' | 'anomaly' | 'text' | 'model' | 'context';
-export type IntegritySignalId = 'scope-declared' | 'excluded-case' | 'listed-in-source' | 'inside-checked-range' | 'parameter-farming' | 'grader-reference' | 'referee' | 'systemone' | 'trajectory-monitor' | 'checker-read-first';
-export interface IntegrityEvidence {
+type IntegritySignalKind = 'exact' | 'anomaly' | 'text' | 'model' | 'context';
+type IntegritySignalId = 'scope-declared' | 'excluded-case' | 'listed-in-source' | 'inside-checked-range' | 'parameter-farming' | 'grader-reference' | 'referee' | 'systemone' | 'trajectory-monitor' | 'checker-read-first';
+interface IntegrityEvidence {
   readonly kind: 'source' | 'claim' | 'lane' | 'trace' | 'model';
   readonly ref: string;
   readonly quote?: string;
 }
-export interface IntegritySignal {
+interface IntegritySignal {
   readonly id: IntegritySignalId;
   readonly kind: IntegritySignalKind;
   /** `true` against the claim; `false` evaluated with nothing against it; `null` not evaluated. */
@@ -257,7 +244,7 @@ export declare function ruleHolds(rule: ScopeRule, claim: Pick<IntegrityClaim, '
  * source reports checked (claims of new only; an anomaly, since sources do err).
  */
 export declare function scopeSignals(claim: IntegrityClaim, scope: SourceScope | undefined): IntegritySignal[];
-export interface FarmingOptions {
+interface FarmingOptions {
   /** Distinct claims that must share one parameter value while another parameter varies. Default 3. */
   readonly minimumRepeats?: number;
 }
@@ -295,7 +282,7 @@ export declare const GRADER_REFERENCE_PHRASES: readonly [{
 /** Whether a claim page names the checker or the printed letter as its justification. */
 export declare function graderReferenceSignal(text: string | null | undefined): IntegritySignal;
 /** One recorded action of the lane's agents, oldest first. */
-export interface TrajectoryEvent {
+interface TrajectoryEvent {
   readonly at: string | number;
   readonly ref: string;
   /** The action as text: a tool name and its arguments. */
@@ -311,7 +298,7 @@ export declare function checkerReadSignal(events: readonly TrajectoryEvent[], op
   firstClaimAt: string | number | null;
 }): IntegritySignal;
 declare const REFEREE_VERDICTS: readonly ['holds', 'excluded', 'trivial', 'known', 'unclear'];
-export type RefereeVerdict = (typeof REFEREE_VERDICTS)[number];
+type RefereeVerdict = (typeof REFEREE_VERDICTS)[number];
 /** The referee's messages for one claim, for any chat model. */
 export declare function sourceScopeRefereeMessages(claim: IntegrityClaim, scope: SourceScope): ({
   role: 'system';
@@ -320,7 +307,7 @@ export declare function sourceScopeRefereeMessages(claim: IntegrityClaim, scope:
   role: 'user';
   content: string;
 })[];
-export interface SourceScopeRefereeAnswer {
+interface SourceScopeRefereeAnswer {
   readonly verdict: RefereeVerdict;
   readonly cite: string | null;
   readonly quote: string | null;
@@ -360,8 +347,8 @@ declare const claimIntegritySystemOneQuestions: {
     readonly instructions: "Is the claim's case inside the statement the source means, so that a verified result there would count for the source's authors?";
   };
 };
-export type ClaimIntegritySystemOneQuestions = typeof claimIntegritySystemOneQuestions;
-export interface ClaimIntegritySystemOneOptions {
+type ClaimIntegritySystemOneQuestions = typeof claimIntegritySystemOneQuestions;
+interface ClaimIntegritySystemOneOptions {
   readonly model: string;
   /**
    * Support mass at or above which a question counts against the claim. The caller calibrates it:
@@ -392,12 +379,11 @@ export declare function monitorSignal(outcomes: readonly TraceQuestionOutcome[]
 } | null, options?: {
   costUsd?: number | null;
 }): IntegritySignal;
-export declare const CLAIM_INTEGRITY_VERSION = "2";
-export interface ClaimIntegrityPolicy {
+interface ClaimIntegrityPolicy {
   /** Signals whose firing is recorded but moves no verdict, such as a judge not yet calibrated. */
   readonly recordOnly?: readonly IntegritySignalId[];
 }
-export interface ClaimIntegrityVerdict {
+interface ClaimIntegrityVerdict {
   readonly verdict: 'pass' | 'hack' | 'unknown';
   readonly reasons: readonly string[];
   /** Whether a trajectory monitor should read the lane before a person decides. */
@@ -444,7 +430,7 @@ export declare function claimIntegrityVerdict(signals: readonly IntegritySignal[
  * or LLM-call count. We accept whatever the caller provides; the curve
  * is on whatever axis they pick.
  */
-export interface ComputeCurveBudget {
+interface ComputeCurveBudget {
   /** Identifier — for the report. Common: '1x', '4x', '16x'. */
   id: string;
   /** Numeric value on the chosen axis (tokens, calls, USD, ms — caller picks). */
@@ -452,7 +438,7 @@ export interface ComputeCurveBudget {
   /** Free-form metadata (the caller can carry per-budget config). */
   meta?: Record<string, unknown>;
 }
-export interface ComputeCurvePoint {
+interface ComputeCurvePoint {
   budgetId: string;
   cost: number;
   score: number;
@@ -463,7 +449,7 @@ export interface ComputeCurvePoint {
   /** Any extra metrics the runner returned. */
   metrics?: Record<string, number>;
 }
-export interface ComputeCurve {
+interface ComputeCurve {
   candidateId: string;
   points: ComputeCurvePoint[];
   /** Rough exponent fit: score ≈ a + b * log(cost). Useful for "how steep is the curve?" */
@@ -471,7 +457,7 @@ export interface ComputeCurve {
   /** Best (highest-score) point on the curve. */
   best: ComputeCurvePoint;
 }
-export interface RunComputeCurveOptions {
+interface RunComputeCurveOptions {
   candidateId: string;
   budgets: ComputeCurveBudget[];
   /**
@@ -485,8 +471,7 @@ export interface RunComputeCurveOptions {
     metrics?: Record<string, number>;
   }>;
 }
-export declare function runComputeCurve(opts: RunComputeCurveOptions): Promise<ComputeCurve>;
-export interface ComputeBestOfNOptions<O> {
+interface ComputeBestOfNOptions<O> {
   /** Number of independent samples to draw. */
   n: number;
   /** Sampler — produces one rollout. */
@@ -494,7 +479,7 @@ export interface ComputeBestOfNOptions<O> {
   /** Score one rollout. */
   scoreFn: (rollout: O) => Promise<number> | number;
 }
-export interface ComputeBestOfNResult<O> {
+interface ComputeBestOfNResult<O> {
   best: O;
   bestScore: number;
   scores: number[];
@@ -504,13 +489,13 @@ export interface ComputeBestOfNResult<O> {
 }
 /** The simplest test-time scaling primitive. */
 export declare function bestOfN<O>(opts: ComputeBestOfNOptions<O>): Promise<ComputeBestOfNResult<O>>;
-export interface SelfConsistencyOptions<O> {
+interface SelfConsistencyOptions<O> {
   n: number;
   sample: (sampleIdx: number) => Promise<O>;
   /** Extract the canonical answer key (string) from a rollout. */
   answerKey: (rollout: O) => string;
 }
-export interface SelfConsistencyResult<O> {
+interface SelfConsistencyResult<O> {
   /** Modal answer (the majority vote). */
   answer: string;
   /** Fraction of samples voting for the modal answer in [0, 1]. */
@@ -522,24 +507,19 @@ export interface SelfConsistencyResult<O> {
   /** All rollouts. */
   rollouts: O[];
 }
-/**
- * Self-consistency / majority-vote test-time scaling. For tasks with a
- * small categorical answer space (math problems, multiple choice).
- */
-export declare function selfConsistency<O>(opts: SelfConsistencyOptions<O>): Promise<SelfConsistencyResult<O>>;
 /**
  * Pareto frontier over (candidate, compute, score) tuples. A point is on
  * the frontier iff no other point dominates it in both score (higher
  * better) and cost (lower better). Returns the frontier sorted ascending
  * by cost.
  */
-export interface ParetoPointInput {
+interface ParetoPointInput {
   candidateId: string;
   budgetId: string;
   cost: number;
   score: number;
 }
-export declare function paretoFrontier(points: ParetoPointInput[]): ParetoPointInput[];
+export declare function paretoCostScoreFrontier(points: ParetoPointInput[]): ParetoPointInput[];
 //#endregion
 //#region src/rl/contamination.d.ts
 /**
@@ -551,15 +531,15 @@ export declare function paretoFrontier(points: ParetoPointInput[]): ParetoPointI
  * does not identify contamination as the cause. Per-item differences have no
  * calibrated sampling null and carry no p-values or q-values.
  */
-export type ScenarioPerturbationKind = 'rename_variables' | 'shuffle_order' | 'paraphrase' | 'inject_irrelevant_clause' | 'custom';
-export interface ScenarioPerturbation<S> {
+type ScenarioPerturbationKind = 'rename_variables' | 'shuffle_order' | 'paraphrase' | 'inject_irrelevant_clause' | 'custom';
+interface ScenarioPerturbation<S> {
   kind: ScenarioPerturbationKind;
   /** Apply to one scenario, return its perturbed sibling. */
   apply: (scenario: S) => Promise<S> | S;
   /** Optional id — for the report. */
   id?: string;
 }
-export interface ContaminationProbeInput<S> {
+interface ContaminationProbeInput<S> {
   /** Stable, unique identity of every original scenario. */
   scenarioId: (s: S) => string;
   /** Original scenarios. */
@@ -577,7 +557,7 @@ export interface ContaminationProbeInput<S> {
    */
   scoreFn: (s: S) => Promise<number>;
 }
-export interface ContaminationProbeOptions {
+interface ContaminationProbeOptions {
   /** Drop scores below this from the probe; treats partial failures separately. Default 0. */
   scoreFloor?: number;
   /** Significance threshold for the single global paired test. Default 0.05. */
@@ -588,7 +568,7 @@ export interface ContaminationProbeOptions {
    */
   minMedianDrop?: number;
 }
-export interface ContaminationProbeReport {
+interface ContaminationProbeReport {
   perScenario: Array<{
     scenarioId: string;
     originalScore: number;
@@ -611,31 +591,6 @@ export interface ContaminationProbeReport {
   excludedScenarioIds: string[];
 }
 export declare function runContaminationProbe<S>(input: ContaminationProbeInput<S>, opts?: ContaminationProbeOptions): Promise<ContaminationProbeReport>;
-/**
- * Identifier-rename perturbation for code/text scenarios. Replaces every
- * occurrence of the listed identifiers with synthesized aliases. Use when
- * the scenario's structural difficulty is independent of variable names
- * (e.g. SWE-Bench-style coding tasks).
- */
-export declare function renameVariables<S extends {
-  prompt: string;
-}>(identifiers: string[], rename?: (name: string, idx: number) => string): ScenarioPerturbation<S>;
-/**
- * Order-shuffle perturbation. Reshuffles a list-shaped section of the
- * prompt (for QA scenarios that present options A/B/C/D — answer depends
- * on the option labels, not order). Caller provides the section extractor.
- */
-export declare function shuffleOrder<S extends {
-  prompt: string;
-}>(shuffleSection: (prompt: string, rng: () => number) => string, seed: number): ScenarioPerturbation<S>;
-/**
- * Inject-irrelevant-clause perturbation. Adds a benign sentence that
- * shouldn't change the answer. Tests for "did the model just memorize
- * the input string."
- */
-export declare function injectIrrelevantClause<S extends {
-  prompt: string;
-}>(clause: string, position?: 'prefix' | 'suffix'): ScenarioPerturbation<S>;
 //#endregion
 //#region src/rl/rollout-input.d.ts
 /**
@@ -790,7 +745,7 @@ export declare function toAnthropicFormat(triples: PreferenceTriple[], context:
 }>;
 //#endregion
 //#region src/rl/process-reward.d.ts
-export interface StepReward {
+interface StepReward {
   /** Trace span this reward attaches to. */
   spanId: string;
   runId: string;
@@ -812,13 +767,13 @@ export interface StepReward {
   /** Optional weight — how much this step contributes to credit assignment. */
   weight?: number;
 }
-export interface StepScorer {
+interface StepScorer {
   /** Span kinds this scorer applies to. */
   appliesTo: Span['kind'][];
   /** Returns null to skip the span; returns a `StepReward` shape (without index/runId/spanId, which are filled in). */
   score(span: Span): Promise<Omit<StepReward, 'spanId' | 'runId' | 'stepIndex'>> | null | undefined;
 }
-export interface ExtractStepRewardsOptions {
+interface ExtractStepRewardsOptions {
   /**
    * Ordered list of scorers. Each span runs through scorers in order;
    * the first non-null result wins. If no scorer applies, the span is
@@ -828,8 +783,7 @@ export interface ExtractStepRewardsOptions {
   /** Optional filter — return null to drop the span entirely before scoring. */
   preFilter?: (span: Span) => boolean;
 }
-export declare function extractStepRewards(store: TraceStore, runId: string, opts: ExtractStepRewardsOptions): Promise<StepReward[]>;
-export interface RunwiseStepSummary {
+interface RunwiseStepSummary {
   runId: string;
   totalSteps: number;
   meanReward: number;
@@ -841,8 +795,7 @@ export interface RunwiseStepSummary {
   worstStepDelta: number;
   worstStepIndex: number | null;
 }
-export declare function runwiseStepRewardSummary(stepRewards: StepReward[]): RunwiseStepSummary;
-export interface PrmTrainingTriple {
+interface PrmTrainingTriple {
   /** Prefix run-id (or composite key) — the trajectory up to step k-1. */
   prefixRunId: string;
   prefixStepIndex: number;
@@ -877,13 +830,13 @@ export declare function prmTrainingPairs(stepRewardsByRun: Map<string, StepRewar
 }): PrmTrainingTriple[];
 //#endregion
 //#region src/rl/exporters.d.ts
-export interface DpoLookups {
+interface DpoLookups {
   /** Resolve the prompt text for a run (typically from a trace store / raw event sink). */
   promptOf: (runId: string) => string | Promise<string>;
   /** Resolve the assistant completion text for a run. */
   completionOf: (runId: string) => string | Promise<string>;
 }
-export interface DpoExportRow {
+interface DpoExportRow {
   prompt: string;
   chosen: string;
   rejected: string;
@@ -893,7 +846,7 @@ export interface DpoExportRow {
   meta?: Record<string, unknown>;
 }
 /** The minted lines for the runs a `PreferenceTriple` names on each side. */
-export type DpoLineContext = RolloutLineContext;
+type DpoLineContext = RolloutLineContext;
 /**
  * Convert preference triples to TRL-compatible DPO rows. The shape
  * `{prompt, chosen, rejected}` is the canonical HuggingFace DPODataset
@@ -910,17 +863,15 @@ export type DpoLineContext = RolloutLineContext;
  * why dropping, not zeroing, is the right disposition for a preference pair.
  */
 export declare function toDpoRows(triples: PreferenceTriple[], lookups: DpoLookups, context: DpoLineContext): Promise<DpoExportRow[]>;
-/** Serialize DPO rows as JSONL. One line per row. */
-export declare function toDpoJsonl(rows: DpoExportRow[]): string;
 /** The one training policy, owned by `rollout/exporters`; the `/rl` name for it. */
-export type TrainingLineSelectionOptions = TrainingExportOptions;
-export interface GrpoLookups extends Pick<TrainingLineSelectionOptions, 'allowHeldOutTrainingData' | 'splitFilter'> {
+type TrainingLineSelectionOptions = TrainingExportOptions;
+interface GrpoLookups extends Pick<TrainingLineSelectionOptions, 'allowHeldOutTrainingData' | 'splitFilter'> {
   /** Resolve the prompt text for a rollout, keyed by `line.run_id`. */
   promptOf: (runId: string) => string | Promise<string>;
   /** Resolve the assistant completion text for a rollout. */
   completionOf: (runId: string) => string | Promise<string>;
 }
-export interface GrpoExportRow {
+interface GrpoExportRow {
   prompt: string;
   completions: string[];
   rewards: number[];
@@ -943,11 +894,10 @@ export interface GrpoExportRow {
  * On the line path a realness-gated line stays in its group at reward 0 rather
  * than being dropped. 0 is the honest label for a faked success and is usable
  * signal; removing the line would also move the group's baseline, just in the
- * other direction. (SFT differs — see `toSftRows`.)
+ * other direction. (SFT differs — see `toSftRowsWithLookups`.)
  */
 export declare function toGrpoRows(lines: MintedRolloutLine[], lookups: GrpoLookups): Promise<GrpoExportRow[]>;
-export declare function toGrpoJsonl(rows: GrpoExportRow[]): string;
-export interface SftLookups extends TrainingLineSelectionOptions {
+interface SftLookups extends TrainingLineSelectionOptions {
   /**
    * Resolve the prompt text for a GAP line (one minted without a transcript),
    * keyed by `line.run_id`. A line with a captured transcript never consults it.
@@ -960,7 +910,7 @@ export interface SftLookups extends TrainingLineSelectionOptions {
   /** Extra filter on top of the realness gate (e.g., low score, failed cases). */
   include?: (line: MintedRolloutLine) => boolean;
 }
-export interface SftExportRow {
+interface SftExportRow {
   /** The canonical chat transcript, tool turns and nullable content included. */
   messages: ChatMessage[];
   meta?: Record<string, unknown>;
@@ -982,9 +932,9 @@ export interface SftExportRow {
  * says which path a row took. A captured transcript whose every turn is copied
  * context yields no row rather than a lookup-built substitute.
  */
-export declare function toSftRows(lines: MintedRolloutLine[], lookups: SftLookups): Promise<SftExportRow[]>;
+export declare function toSftRowsWithLookups(lines: MintedRolloutLine[], lookups: SftLookups): Promise<SftExportRow[]>;
 export declare function toSftJsonl(rows: SftExportRow[]): string;
-export interface PrmLookups {
+interface PrmLookups {
   /** Resolve the prompt text for a run. */
   promptOf: (runId: string) => string | Promise<string>;
   /** Resolve the trajectory step text for a (runId, spanId) pair. */
@@ -992,7 +942,7 @@ export interface PrmLookups {
   /** Optional: sequence of prefix span ids leading up to the divergence. */
   prefixOf?: (runId: string, prefixStepIndex: number) => string[] | Promise<string[]>;
 }
-export interface PrmExportRow {
+interface PrmExportRow {
   prompt: string;
   /** Span ids for the steps before divergence — caller resolves text via `stepTextOf`. */
   prefixSpanIds: string[];
@@ -1004,7 +954,7 @@ export interface PrmExportRow {
   marginScore: number;
   meta?: Record<string, unknown>;
 }
-export interface PrmLineContext extends RolloutLineContext {
+interface PrmLineContext extends RolloutLineContext {
   /**
    * The `maxSteps` cap the lines were minted with, if any.
    *
@@ -1036,8 +986,7 @@ export interface PrmLineContext extends RolloutLineContext {
  * the exact step the gaming happened.
  */
 export declare function toPrmRows(triples: PrmTrainingTriple[], lookups: PrmLookups, context: PrmLineContext): Promise<PrmExportRow[]>;
-export declare function toPrmJsonl(rows: PrmExportRow[]): string;
-export interface StepRewardJsonlRow {
+interface StepRewardJsonlRow {
   runId: string;
   spanId: string;
   stepIndex: number;
@@ -1045,26 +994,15 @@ export interface StepRewardJsonlRow {
   determinism: 'deterministic' | 'probabilistic';
   weight: number;
 }
-/**
- * Step-level reward rows as JSONL.
- *
- * `context` is REQUIRED for the same reason it is on `toDpoRows` and
- * `toPrmRows`: this is a line-less input carrying a reward number. Steps
- * belonging to a realness-gated run are dropped rather than zeroed — a
- * per-step reward of 0 across a whole trajectory is a claim that every step was
- * bad, which is a different (and false) statement from "this run's success was
- * fabricated, so its step-level credit assignment is meaningless".
- */
-export declare function stepRewardsToJsonl(stepRewards: StepReward[], context: RolloutLineContext): string;
 //#endregion
 //#region src/rl/dataset.d.ts
-export type RewardKind = 'deterministic' | 'probabilistic' | 'mixed';
+type RewardKind = 'deterministic' | 'probabilistic' | 'mixed';
 declare const DATASET_FORMATS: readonly ['grpo', 'sft', 'dpo'];
-export type DatasetFormat = (typeof DATASET_FORMATS)[number];
+type DatasetFormat = (typeof DATASET_FORMATS)[number];
 export declare function validateDatasetFormats(value: unknown): DatasetFormat[];
 /** Caller-declared context — the qualitative half of the datasheet that can't
  *  be computed from records. */
-export interface RlDatasetConfig {
+interface RlDatasetConfig {
   name: string;
   version: string;
   /** Product/task domain, e.g. 'legal-m&a', 'tax-1040'. */
@@ -1093,7 +1031,7 @@ export interface RlDatasetConfig {
     verifiableRewardFilter?: boolean;
   };
 }
-export interface RewardStats {
+interface RewardStats {
   n: number;
   mean: number | null;
   median: number | null;
@@ -1101,7 +1039,7 @@ export interface RewardStats {
   max: number | null;
   std: number | null;
 }
-export interface RlDatasetStats {
+interface RlDatasetStats {
   records: number;
   /** Rollouts carrying an explicit task-quality score. */
   scoredRecords: number;
@@ -1125,12 +1063,12 @@ export interface RlDatasetStats {
    */
   rolloutsWithoutCost: number;
 }
-export interface RlDatasetManifest extends RlDatasetConfig {
+interface RlDatasetManifest extends RlDatasetConfig {
   formats: DatasetFormat[];
   rowCounts: Partial<Record<DatasetFormat, number>>;
   stats: RlDatasetStats;
 }
-export interface RlDatasetBundle {
+interface RlDatasetBundle {
   manifest: RlDatasetManifest;
   /** Relative filename -> contents. Write these to a directory to publish. */
   files: Record<string, string>;
@@ -1147,33 +1085,24 @@ export declare function buildRlDataset(lines: MintedRolloutLine[], lookups: Grpo
   triples: PreferenceTriple[];
   lookups: DpoLookups;
 }): Promise<RlDatasetBundle>;
-/** Render the "Datasheet for Datasets" card that a buyer reads. */
-export declare function datasheetToMarkdown(m: RlDatasetManifest): string;
 //#endregion
 //#region src/rl/corpus.d.ts
 /** A corpus record is a RunRecord carrying the trajectory text the harness
  *  captured. `prompt`/`completion` are top-level (the validator ignores extras). */
-export type CorpusRecord = RunRecord & {
+type CorpusRecord = RunRecord & {
   prompt?: string;
   completion?: string;
 };
-export interface CorpusAppendResult {
+interface CorpusAppendResult {
   appended: number;
   /** Skipped because a record with the same runId was already in the corpus
    *  (idempotent appends — NOT re-run collapsing; re-runs get fresh runIds). */
   skipped: number;
   total: number;
 }
-/**
- * Append graded records to the corpus (append-only JSONL). Deduplicates by
- * `runId` against what's already on disk so re-running the same harness is
- * idempotent. Creates the file and parent dir. This is the call every eval
- * harness makes by default after producing its records.
- */
-export declare function appendToCorpus(records: CorpusRecord[], corpusPath: string): CorpusAppendResult;
 /** Read the full corpus. Returns [] if the corpus does not exist yet. */
 export declare function readCorpus(corpusPath: string): CorpusRecord[];
-export interface HarvestOptions {
+interface HarvestOptions {
   /** Keep only records scoring >= this (rejection-sampling for SFT). */
   minScore?: number;
   /** Keep only these source splits. Held-out rows still require the explicit override below. */
@@ -1181,20 +1110,6 @@ export interface HarvestOptions {
   /** Permit held-out rows in training files. Default false. */
   allowHeldOutTrainingData?: boolean;
 }
-/**
- * Harvest the accumulated corpus into a publishable RL dataset bundle. Reads
- * trajectory text from each record's top-level `prompt`/`completion`; records
- * missing either are excluded (a graded score with no trajectory can't train).
- * Optionally filters by score / split. Throws (via buildRlDataset) if nothing
- * survives — an empty dataset must never be published.
- *
- * `minScore` is applied to the GATED reward (`trainingScore`), so a gamed run
- * cannot buy its way into the published bundle with its claimed score —
- * `minScore` is exactly the door a reward-hacked run would otherwise clear for
- * SFT. Unscored records are dropped before packaging: a missing label is not a
- * zero, and it is not publishable either.
- */
-export declare function buildDatasetFromCorpus(corpusPath: string, config: RlDatasetConfig, opts?: HarvestOptions): Promise<RlDatasetBundle>;
 //#endregion
 //#region src/rl/off-policy.d.ts
 /**
@@ -1236,7 +1151,7 @@ export declare function buildDatasetFromCorpus(corpusPath: string, config: RlDat
  * complement with `replayCampaign` (exact replay where the request hashes
  * match) for high-confidence answers and OPE for the gap.
  */
-export interface OffPolicyTrajectory {
+interface OffPolicyTrajectory {
   /** Stable id, for traceability through the dataset. */
   runId: string;
   /** Reward observed under the behavior policy (the realized outcome). */
@@ -1267,13 +1182,13 @@ export interface OffPolicyTrajectory {
    */
   vHatTarget?: number | null;
 }
-export interface OffPolicyContributionCounts {
+interface OffPolicyContributionCounts {
   /** Contributions using the contextual-bandit doubly-robust formula. */
   dr: number;
   /** Contributions using exact IPS because no reward-model estimate was supplied. */
   ipsFallback: number;
 }
-export interface OffPolicyEstimate {
+interface OffPolicyEstimate {
   /** Estimated value of the target policy. */
   value: number;
   /** Standard error of the estimate. */
@@ -1290,7 +1205,7 @@ export interface OffPolicyEstimate {
   /** Populated by `doublyRobust` to expose which formula each row used. */
   contributionCounts?: OffPolicyContributionCounts;
 }
-export interface OffPolicyOptions {
+interface OffPolicyOptions {
   /**
    * Cap importance weights at this value (Ionides 2008 truncated IS) to
    * trade unbiasedness for variance reduction. Default `Infinity` (no cap).
@@ -1336,16 +1251,6 @@ export declare function selfNormalizedImportanceWeighting(trajectories: OffPolic
  * fitting and evaluating Q on the same outcomes leaks the answer.
  */
 export declare function doublyRobust(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): OffPolicyEstimate;
-/**
- * Convenience: run all three estimators and return them side-by-side.
- * The recommended diagnostic — agreement across estimators is a much
- * stronger signal than any single one.
- */
-export declare function offPolicyEstimateAll(trajectories: OffPolicyTrajectory[], opts?: OffPolicyOptions): {
-  ips: OffPolicyEstimate;
-  snips: OffPolicyEstimate;
-  dr: OffPolicyEstimate;
-};
 //#endregion
 //#region src/rl/predictive-validity-researcher.d.ts
 export interface PredictiveValidityResearcherOptions {
@@ -1494,7 +1399,7 @@ export interface RLCampaignResult {
 export declare function runRLCampaign<V>(opts: RunRLCampaignOptions<V>): Promise<RLCampaignResult>;
 //#endregion
 //#region src/rl/run-record-adapters.d.ts
-export interface AdapterContext {
+interface AdapterContext {
   /** Logical experiment id — typically the campaign or sweep identifier. */
   experimentId: string;
   /** Snapshot model id (e.g. `claude-sonnet-4-6@2025-04-15`). */
@@ -1522,73 +1427,17 @@ export interface AdapterContext {
 export declare function campaignToRunRecords(campaign: CampaignResult, ctx: AdapterContext & {
   candidateId?: string;
 }): RunRecord[];
-/**
- * Convert a `MultiLayerVerifier` `VerificationReport` into a `RunRecord`.
- * A split score is emitted only when `report.taskScore` proves the configured
- * scoring panel completed. Partial scores remain in `outcome.raw` for
- * diagnosis. Layer errors and timeouts become judge or execution telemetry;
- * only a scored `fail` layer may produce task-failure detail.
- */
-export declare function verificationReportToRunRecord(report: VerificationReport, ctx: AdapterContext & {
-  candidateId: string;
-  scenarioId: string;
-}, opts?: {
-  runId?: string;
-}): RunRecord;
 //#endregion
 //#region src/rl/sim-fidelity.d.ts
 /** Extracts a flat behavioral feature map from one record. `string` values
  *  are categorical, `number` values are quantile-bucketed over the union of
  *  both sides, `null` means the feature is absent on this record and is
  *  counted explicitly as its own category (never silently dropped). */
-export type BehaviorFeatures = (record: RunRecord) => Record<string, string | number | null>;
-/** Reserved histogram category for `null` feature values. A capture-rate
- *  difference (one side instruments a signal, the other does not) registers
- *  as divergence by design: a simulator that produces no tool traces is not
- *  representative of production that does. */
-export declare const ABSENT_CATEGORY = "(absent)";
-/**
- * Default feature set — ONLY fields verified present on both simulated and
- * production records:
- *
- *   - `score`, `wall_ms`, `output_tokens` — mandatory per the `RunRecord`
- *     validator (non-finite values read as absent rather than poisoning a
- *     bucket).
- *   - `failure_class` — optional taxonomy field; absent counted explicitly.
- *   - `turn_count`, `tool_errors`, `tool_error_recovery` — derived from the
- *     `outcome.raw` counters the intake adapters and eval harnesses write
- *     (`turns_completed`, `assistant_messages`, `tool_errors`,
- *     `turns_aborted`); absent on records whose producer did not capture
- *     them, counted explicitly.
- *   - `completion_length` — from the optional `CorpusRecord` trajectory
- *     text; the message-length proxy when records come from the corpus.
- *
- * `RunRecord` carries event COUNTS, not event ordering, so
- * `tool_error_recovery` is a counts-only derivation: errors occurred and the
- * run still completed cleanly ('recovered') vs aborted or classified as a
- * failure ('unrecovered') — not a literal error→retry sequence check.
- */
-export declare const defaultBehaviorFeatures: BehaviorFeatures;
-/**
- * Jensen-Shannon divergence between two categorical histograms (raw counts;
- * normalized internally). Log base 2 → bounded [0,1]: 0 = identical
- * distributions, 1 = disjoint support. Symmetric, defined even where the
- * supports differ — exactly the regime sim-vs-production comparison lives in.
- * Throws on zero-mass or negative/non-finite counts: an empty histogram has
- * no distribution and a silent 0 would read as "perfectly representative".
- */
-export declare function jsDivergence(p: Record<string, number>, q: Record<string, number>): number;
-/**
- * Deterministic quantile edges over a value set (the UNION of both sides, so
- * sim and production land in the same buckets). Linear interpolation between
- * order statistics; duplicate edges from heavy ties collapse into fewer,
- * wider buckets. Returns `bucketCount - 1` edges before deduplication.
- */
-export declare function quantileEdges(values: number[], bucketCount?: number): number[];
+type BehaviorFeatures = (record: RunRecord) => Record<string, string | number | null>;
 /** Stable half-open bucket label for a value against quantile edges:
  *  `[-inf,e0)`, `[e0,e1)`, …, `[eLast,+inf)`. */
 export declare function bucketLabel(value: number, edges: number[]): string;
-export interface FeatureShift {
+interface FeatureShift {
   /** Category label (a string value, a numeric bucket, or `ABSENT_CATEGORY`). */
   value: string;
   /** Probability of this category among ALL simulated records (nulls included
@@ -1597,7 +1446,7 @@ export interface FeatureShift {
   /** Probability among ALL production records. */
   pProd: number;
 }
-export interface FeatureDivergence {
+interface FeatureDivergence {
   feature: string;
   /** Jensen-Shannon divergence in [0,1] for this feature. */
   divergence: number;
@@ -1608,8 +1457,8 @@ export interface FeatureDivergence {
   /** Non-null observations on the production side. */
   nProd: number;
 }
-export type FidelityVerdict = 'representative' | 'skewed' | 'insufficient-data';
-export interface FidelityReport {
+type FidelityVerdict = 'representative' | 'skewed' | 'insufficient-data';
+interface FidelityReport {
   perDimension: FeatureDivergence[];
   /** 1 − mean divergence over features with sufficient data. NaN when the
    *  verdict is 'insufficient-data' — a 0 would read as "maximally skewed"
@@ -1622,20 +1471,13 @@ export interface FidelityReport {
    *  'skewed' below, 'insufficient-data' when no feature met minN. */
   verdict: FidelityVerdict;
 }
-export interface SimFidelityOptions {
+interface SimFidelityOptions {
   /** Feature extractor. Defaults to `defaultBehaviorFeatures`. */
   features?: BehaviorFeatures;
   /** Minimum non-null observations per side per feature. Default 20. */
   minNPerFeature?: number;
 }
-/**
- * Compare a simulator's RunRecords against production RunRecords, feature by
- * feature. Numeric features are bucketed by deterministic quantiles of the
- * union; nulls count as an explicit `ABSENT_CATEGORY`. Throws on empty
- * inputs — "no records" is a wiring error, not a distribution.
- */
-export declare function simFidelityReport(simulated: RunRecord[], production: RunRecord[], opts?: SimFidelityOptions): FidelityReport;
-export interface EasyModeOptions {
+interface EasyModeOptions {
   /** A run passes when its score (holdout, else search) >= this. Default 0.5
    *  — matches the pass-threshold convention across the rl/ primitives. */
   passThreshold?: number;
@@ -1643,7 +1485,7 @@ export interface EasyModeOptions {
    *  a 10-point inflation is enough to flip most promotion gates. */
   inflationTolerance?: number;
 }
-export interface EasyModeReport {
+interface EasyModeReport {
   simPassRate: number;
   prodPassRate: number;
   /** simPassRate − prodPassRate. Positive = the simulator is easier than reality. */
@@ -1652,14 +1494,6 @@ export interface EasyModeReport {
    *  simulator overstate production performance. */
   inflated: boolean;
 }
-/**
- * The headline simulator failure mode as its own named artifact: a simulator
- * that creates "easy mode" inflates pass-rate relative to production, and
- * every score measured against it overstates reality. Throws on empty inputs
- * and on records carrying neither score — a silently-skipped record would
- * bias the very rate this check exists to keep honest.
- */
-export declare function easyModeCheck(simulated: RunRecord[], production: RunRecord[], opts?: EasyModeOptions): EasyModeReport;
 //#endregion
 //#region src/rl/tournament.d.ts
 /**
@@ -1688,7 +1522,7 @@ export declare function easyModeCheck(simulated: RunRecord[], production: RunRec
  * tournament eval as the most sample-efficient and most rank-stable
  * method when you have many candidates.
  */
-export interface PairwiseOutcome {
+interface PairwiseOutcome {
   /** Winner candidate id. */
   winner: string;
   /** Loser candidate id. */
@@ -1705,7 +1539,7 @@ export interface PairwiseOutcome {
    */
   weight?: number;
 }
-export interface BradleyTerryRating {
+interface BradleyTerryRating {
   candidateId: string;
   /** Latent strength θ ≥ 0 from the BT MLE. */
   strength: number;
@@ -1716,7 +1550,7 @@ export interface BradleyTerryRating {
   /** Win count (+ 0.5 per draw). */
   wins: number;
 }
-export interface BradleyTerryFit {
+interface BradleyTerryFit {
   ratings: BradleyTerryRating[];
   /** Iterations of the MM algorithm before convergence. */
   iterations: number;
@@ -1746,23 +1580,19 @@ export declare function fitBradleyTerry(outcomes: PairwiseOutcome[], opts?: {
  * `applyEloUpdate` mutates the map in place and returns the deltas so
  * the caller can log per-comparison rating changes.
  */
-export interface EloOptions {
+interface EloOptions {
   /** Default rating for unseen candidates. Default 1500. */
   defaultRating?: number;
   /** K-factor controls the step size. Default 32 (FIDE-ish). */
   kFactor?: number;
 }
-export declare function applyEloUpdate(ratings: Map<string, number>, outcome: PairwiseOutcome, opts?: EloOptions): {
-  winnerDelta: number;
-  loserDelta: number;
-};
 /**
  * Build pairwise outcomes from the campaign artifact: for every scenario
  * shared by two candidates, the higher-scoring run wins. Useful when you
  * want a tournament view of an existing campaign without an additional
  * pairwise judge call.
  */
-export interface BuildPairwiseFromCampaignInput {
+interface BuildPairwiseFromCampaignInput {
   runs: Array<{
     candidateId: string;
     /** Stable identifier for the matching unit (typically scenarioId). */
@@ -1775,7 +1605,6 @@ export interface BuildPairwiseFromCampaignInput {
    */
   drawMargin?: number;
 }
-export declare function buildPairwiseFromCampaign(input: BuildPairwiseFromCampaignInput): PairwiseOutcome[];
 //#endregion
 //#region src/rl/verified-findings-dataset.d.ts
 /**
@@ -1808,9 +1637,9 @@ export declare function buildPairwiseFromCampaign(input: BuildPairwiseFromCampai
  * per-case files contribute prefix-divergence detail and run ids only, and
  * are cross-checked against the report where they overlap.
  */
-export declare const VERIFIED_FINDING_SCHEMA = "agent-eval/verified-finding@0";
+declare const VERIFIED_FINDING_SCHEMA = "agent-eval/verified-finding@0";
 /** One case row from a replay-verify `batch-report.json`. */
-export interface ReplayBatchCase {
+interface ReplayBatchCase {
   corpus: string;
   trajId: string;
   image: string;
@@ -1837,7 +1666,7 @@ export interface ReplayBatchCase {
   fix: ReplayBatchFix | null;
   wallMs: number;
 }
-export interface ReplayBatchFix {
+interface ReplayBatchFix {
   attempted: boolean;
   sampledOut: boolean;
   command: string | null;
@@ -1845,12 +1674,12 @@ export interface ReplayBatchFix {
   armBExit: number | null;
   failureVanished: boolean | null;
 }
-export interface ReplayBatchReport {
+interface ReplayBatchReportFile {
   generatedAt: string;
   cases: ReplayBatchCase[];
 }
 /** Gold label entry for one trajectory (CodeTraceBench annotation format). */
-export interface GoldLabelEntry {
+interface GoldLabelEntry {
   traj_id: string;
   solved: boolean;
   step_count: number;
@@ -1864,12 +1693,12 @@ export interface GoldLabelEntry {
   }>;
 }
 /** One step from a normalized trajectory `steps.json`. */
-export interface NormalizedStep {
+interface NormalizedStep {
   step_id: number;
   action: string;
   observation?: string | null;
 }
-export interface PrefixDivergence {
+interface RecordedPrefixDivergence {
   step: number;
   /** `unknown-expectation` marks a step the recording carries no returncode
    *  for: it could not be confirmed, so it is not agreement. */
@@ -1879,19 +1708,19 @@ export interface PrefixDivergence {
   actualExit: number;
 }
 /** Optional extract from a per-case `replay-verdict.json` (arm A detail only). */
-export interface CaseVerdictDetail {
+interface CaseVerdictDetail {
   k: number;
   prefixExecuted: number;
   recordedReturncode: number;
   signatureBasis: string | null;
-  prefixDivergences: PrefixDivergence[];
+  prefixDivergences: RecordedPrefixDivergence[];
   armACommand: string | null;
   runIds: {
     original: string | null;
     armA: string | null;
   };
 }
-export interface TrajectoryStep {
+interface DatasetTrajectoryStep {
   stepId: number;
   action: string;
   observation: string | null;
@@ -1899,15 +1728,15 @@ export interface TrajectoryStep {
   observationTruncated: boolean;
   observationChars: number;
 }
-export type FixOutcome = 'flipped' | 'not-flipped' | 'generation-failed' | 'not-attempted';
-export interface VerifiedFindingFix {
+type FixOutcome = 'flipped' | 'not-flipped' | 'generation-failed' | 'not-attempted';
+interface VerifiedFindingFix {
   outcome: FixOutcome;
   command: string | null;
   llmError: string | null;
   armBExit: number | null;
   failureVanished: boolean | null;
 }
-export interface VerifiedFindingRow {
+interface VerifiedFindingRow {
   schema: typeof VERIFIED_FINDING_SCHEMA;
   /** `<runId>/<corpus>/<trajId>` — unique across batches. */
   caseId: string;
@@ -1938,7 +1767,7 @@ export interface VerifiedFindingRow {
       start: number;
       end: number;
     };
-    steps: TrajectoryStep[];
+    steps: DatasetTrajectoryStep[];
   };
   verification: {
     reproduced: boolean;
@@ -1953,7 +1782,7 @@ export interface VerifiedFindingRow {
     /** Steps the recording carries no returncode for. Nonzero here means part
      *  of the prefix replay was never confirmed against the recording. */
     prefixUnknownExpectations: number;
-    prefixDivergenceDetail: PrefixDivergence[] | null;
+    prefixDivergenceDetail: RecordedPrefixDivergence[] | null;
     armAExit: number | null;
     armAReturncodeMatch: boolean;
     armACommand: string | null;
@@ -1976,7 +1805,7 @@ export interface VerifiedFindingRow {
     armARunId: string | null;
   };
 }
-export interface VerifiedFindingsSummary {
+interface VerifiedFindingsSummary {
   rows: number;
   reproduced: number;
   /** Reproduced AND arm A matched the failure signature — the batch report's headline strict rate.
@@ -1990,7 +1819,7 @@ export interface VerifiedFindingsSummary {
     fixFlipped: number;
   }>;
 }
-export interface BuildVerifiedFindingRowArgs {
+interface BuildVerifiedFindingRowArgs {
   batchCase: ReplayBatchCase;
   label: GoldLabelEntry;
   steps: NormalizedStep[];
@@ -2004,19 +1833,12 @@ export interface BuildVerifiedFindingRowArgs {
   detail?: CaseVerdictDetail;
   maxObservationChars?: number;
 }
-/**
- * Join one batch case with its gold label and trajectory into a row.
- * Throws on any join inconsistency — never emits a partially joined row.
- */
-export declare function buildVerifiedFindingRow(args: BuildVerifiedFindingRowArgs): VerifiedFindingRow;
-export declare function summarizeVerifiedFindings(rows: VerifiedFindingRow[]): VerifiedFindingsSummary;
-export declare function verifiedFindingsToJsonl(rows: VerifiedFindingRow[]): string;
-export interface VerifiedFindingsCorpusSource {
+interface VerifiedFindingsCorpusSource {
   labelsPath: string;
   /** Directory containing `normalized/<trajId>/steps.json`. */
   preparedDir: string;
 }
-export interface VerifiedFindingsSource {
+interface VerifiedFindingsSource {
   batchReportPath: string;
   /** Batch run identifier embedded in every caseId, e.g. 'run2-20260802'. */
   runId: string;
@@ -2026,7 +1848,7 @@ export interface VerifiedFindingsSource {
   runDir?: string;
   maxObservationChars?: number;
 }
-export interface VerifiedFindingsDataset {
+interface VerifiedFindingsDataset {
   rows: VerifiedFindingRow[];
   summary: VerifiedFindingsSummary;
   provenance: {
@@ -2048,5 +1870,5 @@ export interface VerifiedFindingsDataset {
  */
 export declare function loadVerifiedFindingsDataset(source: VerifiedFindingsSource): VerifiedFindingsDataset;
 //#endregion
-export { AdversarialMutation, type DeploymentOutcome, DetectRewardHackingInput, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, InMemoryOutcomeStore, type OutcomeStore, RewardHackingFinding, RewardHackingReport, RewardHackingSignal, type RolloutLineContext, VerifiableReward, VerifiableRewardExtractionOptions, VerifiableRewardSource, detectRewardHacking, extractVerifiableReward, extractVerifiableRewardsFromRecords, filterDeterministicallyRewarded, runEvalCampaign };
+export { type AdapterContext, AdversarialMutation, type BehaviorFeatures, type BradleyTerryFit, type BradleyTerryRating, type BuildPairwiseFromCampaignInput, type BuildVerifiedFindingRowArgs, type CaseVerdictDetail, type CellObservation, type ClaimIntegrityPolicy, type ClaimIntegritySystemOneOptions, type ClaimIntegritySystemOneQuestions, type ClaimIntegrityVerdict, type ComputeBestOfNOptions, type ComputeBestOfNResult, type ComputeCurve, type ComputeCurveBudget, type ComputeCurvePoint, type ContaminationProbeInput, type ContaminationProbeOptions, type ContaminationProbeReport, type CorpusAppendResult, type CorpusRecord, type CurriculumAllocation, type DatasetFormat, type DatasetTrajectoryStep, type DeploymentOutcome, DetectRewardHackingInput, type DpoExportRow, type DpoLineContext, type DpoLookups, type EasyModeOptions, type EasyModeReport, type EloOptions, type ExtractStepRewardsOptions, type FarmingOptions, type FeatureDivergence, type FeatureShift, type FidelityReport, type FidelityVerdict, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FixOutcome, type GoldLabelEntry, type GrpoExportRow, type GrpoLookups, type HarvestOptions, InMemoryOutcomeStore, type IntegrityClaim, type IntegrityEvidence, type IntegritySignal, type IntegritySignalId, type IntegritySignalKind, type NormalizedStep, type OffPolicyContributionCounts, type OffPolicyEstimate, type OffPolicyOptions, type OffPolicyTrajectory, type OutcomeStore, type PairwiseOutcome, type ParetoPointInput, type PrmExportRow, type PrmLineContext, type PrmLookups, type PrmTrainingTriple, type RecordedPrefixDivergence, type RefereeVerdict, type ReplayBatchCase, type ReplayBatchFix, type ReplayBatchReportFile, RewardHackingFinding, RewardHackingReport, RewardHackingSignal, type RewardKind, type RewardStats, type RlDatasetBundle, type RlDatasetConfig, type RlDatasetManifest, type RlDatasetStats, type RolloutLineContext, type RunComputeCurveOptions, type RunwiseStepSummary, type ScenarioPerturbation, type ScenarioPerturbationKind, type ScopeCondition, type ScopeOperator, type ScopeRule, type ScopeValue, type SelfConsistencyOptions, type SelfConsistencyResult, type SftExportRow, type SftLookups, type SimFidelityOptions, type SourceExcerpt, type SourceScope, type SourceScopeRefereeAnswer, type StepReward, type StepRewardJsonlRow, type StepScorer, type TrainingLineSelectionOptions, type TrajectoryEvent, type VarianceCurriculumOptions, type VerifiableReward, type VerifiableRewardExtractionOptions, type VerifiableRewardSource, type VerifiedFindingFix, type VerifiedFindingRow, type VerifiedFindingsCorpusSource, type VerifiedFindingsDataset, type VerifiedFindingsSource, type VerifiedFindingsSummary, detectRewardHacking, extractVerifiableReward, extractVerifiableRewardsFromRecords, runEvalCampaign };
 //# sourceMappingURL=rl.d.ts.map