/**
 * `openlore prove` (Spec 25 Q2) — measure OpenLore's token value on the user's
 * OWN repo and print a personal scorecard. Runs a WITH/WITHOUT agent pass over
 * a few graph-derived orientation tasks, isolated with --strict-mcp-config, and
 * reports cost / round-trips / correctness deltas + an honest verdict.
 *
 * The substrate needs no API key; this command's agent arm does (it shells out
 * to `claude`). When `claude` is absent it fails fast with guidance; `--dry-run`
 * exercises the whole pipeline with clearly-labelled synthetic numbers.
 *
 * Output + sharing (add-prove-shareable-scorecard):
 *   --json       stable, CI-consumable scorecard (decision 581a90bf)
 *   --markdown   paste-ready block + shields.io badge for a README
 *   --save       persist a dated scorecard under .openlore/prove/ (decision 670b5f0b)
 *   --estimate   deterministic, no-API graph projection of the orientation tax (decision 66feae62)
 */
import { Command } from 'commander';
import { type AgentRunner, type Metrics, type Cell } from '../../core/agent-eval/measure.js';
import { type Scorecard, type ScorecardMeta } from '../../core/agent-eval/scorecard.js';
/**
 * Parse a numeric CLI flag, clamping a valid value to `min` but REJECTING a
 * non-numeric one (e.g. `--runs abc`) instead of silently letting NaN through —
 * NaN would skip every agent run and emit a degenerate all-zero scorecard.
 */
export declare function parseNumericFlag(raw: string | undefined, name: string, integer: boolean, min: number, dflt: number): number | {
    error: string;
};
/** Raw metrics carried alongside a scorecard, for persistence (`--save`). */
export interface ProveRaw {
    withoutCell: Cell;
    withCell: Cell;
    /** Per-run metrics (measured / dry-run only; absent for an estimate). */
    withoutRuns?: Metrics[];
    withRuns?: Metrics[];
}
export interface ProveResult {
    ok: boolean;
    message: string;
    scorecard?: Scorecard;
    meta?: ScorecardMeta;
    raw?: ProveRaw;
}
/**
 * Summarize the two measured arms into comparable cells — but only over
 * SUCCESSFUL runs. An errored run (agent threw / unparseable output) is not a
 * valid cost/turn sample: including its zeros pollutes the medians, and if EVERY
 * run failed the all-zero cells would emit a confident-looking but meaningless
 * "break-even" verdict through the JSON contract. So errored runs are dropped,
 * and if either arm has no successful sample we fail loudly rather than report a
 * verdict over no data. Pure + exported for unit testing without a graph.
 */
export declare function summarizeArms(withoutRuns: Metrics[], withRuns: Metrics[]): {
    ok: true;
    withoutCell: Cell;
    withCell: Cell;
} | {
    ok: false;
    message: string;
};
/**
 * Core (testable) prove run: derive tasks, then either run both agent arms N
 * times (measured / dry-run) or compute the deterministic estimate. Returns the
 * scorecard + provenance so the command can render any output form. `runner` is
 * injectable so tests never call a real agent; `generatedAt`/`repoSha` are passed
 * in so the core stays deterministic.
 */
export declare function runProve(opts: {
    directory: string;
    runs: number;
    model: string;
    maxBudgetUsd: number;
    dryRun: boolean;
    estimate?: boolean;
    generatedAt: string;
    repoSha: string | null;
    runner?: AgentRunner;
}): Promise<ProveResult>;
export declare function saveScorecard(absDir: string, result: ProveResult): string;
export declare const proveCommand: Command;
//# sourceMappingURL=prove.d.ts.map