import type { AgencyConfig } from "../../config/config.js";
import type { EvalRunGrading } from "../../eval/runTypes.js";
import { type ScoreDraft } from "../../runDirectory/mutations.js";
import { type GraderSource } from "./gradeRun.js";
import type { Scorecard } from "./scorecard.js";
export type GradeSuiteResult = {
    grading: EvalRunGrading;
    scorecard: Scorecard;
    passId: string | null;
    /** LLM spend of this pass's judge calls, in USD. */
    judgeCostUsd: number;
};
/**
 * Score a run directory and record the verdict as one grading pass: one
 * `score` annotation per grader per trace, all sharing a fresh pass id, the
 * last one marking the pass complete. Reads one snapshot, grades it, converts
 * the grades to score drafts, and hands them to `recordGradingPass` — the run
 * directory owns how they land. Shared by every caller that grades a
 * directory (`eval grade`, the optimizer's per-candidate grading).
 *
 * `record: false` grades without writing (the optimizer's throwaway candidate
 * directories, and dry runs). `defaultGoal` is `eval grade --goal`: the goal
 * for tests that recorded none.
 */
export declare function gradeSuite(runDir: string, graders: GraderSource, config: AgencyConfig, options?: {
    record?: boolean;
    defaultGoal?: string;
}): Promise<GradeSuiteResult>;
/** One draft per grade that actually ran. A gate-failed or ungraded trace
 *  contributes the rows it has (its zero is the fold's business: a failed
 *  gate is a `pass: false` row, an ungraded trace has none). A judge's row
 *  records the goal it scored against, so a re-grade with a different
 *  `--goal` stays tellable apart from the first. */
export declare function scoreDrafts(scorecard: Scorecard): ScoreDraft[];