/**
 * PURE deterministic regression gate for the replay corpus.
 *
 * PURE — compareToBaseline must not import from ../providers; no network, no
 * Date.now(), no side effects, no filesystem access. All inputs are parameters.
 * Fresh verdicts are re-derived from frozen eval_details_json — NOT by
 * re-running the generator or evaluator LLM.
 *
 * Regression / improvement classification:
 *   baseline 'pass' + fresh 'fail' → regression
 *   baseline 'fail' + fresh 'pass' → improvement
 *   same verdict                   → unchanged
 *
 * A regression is STRICTLY pass→fail. CaseIds present only in fresh (not in
 * baseline) are not classified at all — the baseline corpus defines the gate.
 *
 * Fresh-verdict derivation rule (deterministic, no LLM):
 *   Parse evalDetailsJson (JSON array of result objects). The fresh verdict is
 *   'fail' iff any element's failures[] contains an entry with
 *   passed === false AND severity === 'error'. Otherwise 'pass'.
 *   Malformed / missing fields → treated as 'pass' (no error-severity failure).
 */
import type { BoberConfig } from "../../config/schema.js";
export type Verdict = "pass" | "fail";
export interface ReplayComparison {
    /** caseIds where baseline was 'pass' and fresh is 'fail'. */
    regressions: string[];
    /** caseIds where baseline was 'fail' and fresh is 'pass'. */
    improvements: string[];
    /** caseIds where verdict did not change. */
    unchanged: string[];
}
/** Extended comparison result returned by runReplayHarness. */
export interface ReplayHarnessResult extends ReplayComparison {
    /** Total number of cases in the corpus. */
    total: number;
    /** Fresh verdicts keyed by caseId (for CLI table rendering). */
    fresh: Map<string, Verdict>;
    /** Baseline verdicts keyed by caseId (for CLI table rendering). */
    baseline: Map<string, Verdict>;
}
/**
 * PURE comparator. Classifies each caseId present in `baseline` as regression,
 * improvement, or unchanged based on the transition of verdict.
 *
 * PURE — no clock, no fs, no LLM. All inputs are parameters.
 *
 * @param baseline  Map<caseId, Verdict> from the captured corpus.
 * @param fresh     Map<caseId, Verdict> re-derived from frozen evalDetailsJson.
 * @returns         Sorted arrays for deterministic, byte-identical output.
 */
export declare function compareToBaseline(baseline: Map<string, Verdict>, fresh: Map<string, Verdict>): ReplayComparison;
/**
 * Open the Sprint-1 ReplayStore, re-derive fresh verdicts deterministically
 * from each case's frozen evalDetailsJson, and return a regression/improvement
 * breakdown via compareToBaseline.
 *
 * NEVER re-runs the generator or evaluator LLM — fresh verdicts come only from
 * the frozen captured eval_details_json (the deterministic fresh-verdict rule).
 *
 * @param projectRoot  Absolute path to the project root.
 * @param config       Already-loaded BoberConfig. config.selfImprove is OPTIONAL.
 */
export declare function runReplayHarness(projectRoot: string, config: BoberConfig): Promise<ReplayHarnessResult>;
//# sourceMappingURL=replay-harness.d.ts.map