import { type HarnessProtectionCheck, type HarnessProtectionPlan } from './harness-protection.js';
import type { CaseResult } from './types.js';
/** A host-owned outcome, with enough identity to check pairing and cite its trace. */
export interface HarnessTrial {
    taskId: string;
    trial: number;
    /** Environment/seed/model/limits fingerprint. Equal within a pair, fresh in confirmation. */
    conditions: string;
    trajectoryId: string;
    result: CaseResult;
}
export interface HarnessAttribution {
    taskId: string;
    effect: 'improvement' | 'regression' | 'unresolved';
    reason: string;
    baselineTrajectories: readonly string[];
    candidateTrajectories: readonly string[];
}
export interface HarnessVerificationBatch {
    baselineRevision: string;
    candidateRevision: string;
    baseline: readonly HarnessTrial[];
    candidate: readonly HarnessTrial[];
    /** Supplied by an independent trace reviewer, never inferred from scores alone. */
    attributions: readonly HarnessAttribution[];
}
export type HarnessBehaviorStatus = 'recovered' | 'stable-success' | 'regressed' | 'still-failing' | 'mixed' | 'inconclusive';
export interface HarnessTaskComparison {
    taskId: string;
    status: HarnessBehaviorStatus;
    baselinePasses: number;
    candidatePasses: number;
    trials: number;
}
export interface HarnessComparison {
    tasks: readonly HarnessTaskComparison[];
    /** Paired trial pass-rate difference, not held-out pass@1 or causal evidence. */
    passRateDelta: number | null;
    positiveEvidence: readonly string[];
    regressions: readonly string[];
    unresolved: readonly string[];
    /** Actual recorded consumption from BOTH sides, including unsuccessful runs. */
    usage: {
        rollouts: number;
        tokens: number;
        costUsd: number;
        durationMs: number;
    };
}
/** HarnessLens Table 6 comparison plus explicit, trace-linked attribution. No model calls. */
export declare function compareHarnessTrials(batch: HarnessVerificationBatch): HarnessComparison;
export interface HarnessReview {
    decision: 'accept' | 'reject' | 'inconclusive';
    reason: string;
    verification: HarnessComparison;
    confirmation?: HarnessComparison;
    protection?: {
        verification: HarnessProtectionCheck;
        confirmation?: HarnessProtectionCheck;
    };
}
/**
 * Evaluate recorded evidence; never applies a patch. Requires at least five tasks,
 * exactly two trials per task, and a same-size confirmation with at most two reused tasks.
 * Hosts own isolated execution, task selection, scoring and independent trace review.
 */
export declare function reviewHarnessCandidate(verification: HarnessVerificationBatch, confirmation?: HarnessVerificationBatch, protection?: HarnessProtectionPlan): HarnessReview;
//# sourceMappingURL=harness-verification.d.ts.map