/**
 * prompt-grader — independent LLM-as-judge grader.
 *
 * Evaluates a single trajectory against user-provided rubric criteria.
 * Uses a structured JSON response from the LLM with per-criterion scores
 * and reasoning, then maps to the standard GraderResult interface.
 */
import type { Grader, GraderInput, GraderMetadata, GraderResult, GraderComparisonInput, GraderComparisonResult } from "../types.js";
import type { LlmClient, PromptGraderConfig, ResolvedRubric } from "./types.js";
import { type ResolvedLlmEvidence } from "./evidence.js";
import { type JudgeEvidenceFile } from "./judge-workspace.js";
import type { Trajectory } from "../../trajectory/types.js";
/** Config schema for the `prompt` grader in eval.yaml. All fields optional. */
export type Config = PromptGraderConfig;
/**
 * Resolve the rubric for a stimulus, falling back to the default when it is
 * missing *or empty*. An empty rubric would leave the judge with no criteria,
 * yet the response schema requires at least one entry — which invites tool-call
 * failures — so treat `[]` the same as absent.
 *
 * The `source` is reported as `rubric_source` in grader metadata, since the
 * default criteria read plausibly enough that a fallback otherwise goes unseen.
 */
export declare function resolveRubric(stimulus: GraderInput["stimulus"]): ResolvedRubric;
export declare class PromptGrader implements Grader {
    metadata: GraderMetadata;
    private client;
    constructor(client: LlmClient);
    /**
     * `prompt "did it use bicep"` — the rubric distinguishes instances. Double
     * quotes signal prose (a judge question) rather than a literal to match.
     * Optional: the grader falls back to the stimulus rubric, in which case
     * there is nothing to discriminate on and the bare type is right.
     */
    defaultName(config: Record<string, unknown>): string;
    grade(input: GraderInput): Promise<GraderResult>;
    /**
     * Comparison mode: judge a baseline trajectory head-to-head against a
     * treatment for the same stimulus, against the same rubric. Uses
     * position-swap debiasing (forward + reverse) and maps the verdict onto the
     * signed, treatment-relative {@link ComparisonMagnitude} scale.
     *
     * v1 supports only `baseline-relative` topology; `n-way` is reserved.
     */
    compare(input: GraderComparisonInput): Promise<GraderComparisonResult>;
    private compareOnce;
}
export declare function buildUserMessage(input: GraderInput, rubric: string[], customPrompt?: string, resolvedEvidence?: ResolvedLlmEvidence, nonce?: string): string;
export declare function buildWorkspaceSystemMessage(systemMessage: string, files: readonly JudgeEvidenceFile[], workingDirectory: string): string;
export declare function buildComparisonUserMessage(trajectoryA: Trajectory, trajectoryB: Trajectory, taskPrompt: string | undefined, rubric: string[], customPrompt?: string, nonce?: string): string;
//# sourceMappingURL=prompt-grader.d.ts.map