/**
 * Darwin — Multi-Critic Evaluator
 *
 * Runs 3 specialized critics in parallel and takes the median score.
 * More robust than a single critic — reduces bias and random variance.
 *
 * Agent-aware: Different agents get different evaluation criteria.
 *
 * Investigator Critics:
 *   A: Facts & Sources — accuracy, citations, primary documents
 *   B: Honesty & Courage — intellectual bravery, clear positions, uncomfortable truths
 *   C: Completeness & Structure — full investigation, proper format, both sides covered
 *
 * Writer Critics:
 *   A: Task Compliance & Accuracy — did the writer follow the brief? Correct claims?
 *   B: Persuasion & Voice — tone, engagement, conviction, audience awareness
 *   C: Substance & Originality — depth, concrete value, fresh angles
 */
/** Critic prompt definition */
export interface CriticPromptDef {
    name: string;
    prompt: string;
}
/** Score result from a single critic */
export interface CriticScore {
    critic: string;
    score: number;
    report: string;
}
/** Combined result from multi-critic evaluation */
export interface MultiCriticResult {
    /** Median score across all critics */
    medianScore: number;
    /** Individual critic results */
    critics: CriticScore[];
    /** Combined report text */
    combinedReport: string;
}
/** Function that runs a critic and returns its output */
export type RunCriticFn = (systemPrompt: string, task: string, criticName: string) => Promise<string>;
/** Options for {@link runMultiCritic} (v0.7.0). */
export interface RunMultiCriticOptions {
    /**
     * v0.7.0 — Strip markdown from the agent output BEFORE handing it to the
     * critics, so they score CONTENT, not FORMAT.
     *
     * Why: LLM-as-judge research consistently documents a FORMAT/STYLE bias —
     * judges tend to prefer well-formatted (markdown) answers over identical
     * plain prose. When evolving prompts produce outputs in different formats,
     * an un-normalised judge measures formatting, not quality, and the optimizer
     * drifts toward "add more bold/headers" instead of "be more correct".
     * Normalising both candidates' outputs the same way removes that confound.
     * (Position bias, by contrast, is reported to be negligible on current
     * frontier judge models, so this normalises format only — not order.)
     *
     * Off by default (byte-for-byte v0.6.0 behaviour). Turn ON for agents whose
     * deliverable is prose; leave OFF when the format itself is the deliverable
     * (e.g. an agent that must "produce a markdown table").
     */
    normalizeForJudging?: boolean;
    /**
     * v0.12.0 — Bring your own judges: an explicit critic-prompt set for THIS
     * call, bypassing the built-in {@link getCriticPrompts} name lookup.
     *
     * Why: the built-in `AGENT_CRITIC_MAP` covers a handful of generic agent
     * archetypes (investigator / writer / research / critic / analyst / …).
     * Any fleet with domain agents beyond those archetypes previously had to
     * FORK this file to register its own judges — unknown names silently fall
     * back to `INVESTIGATOR_PROMPTS`, which mis-scores domain output (e.g. a
     * game-simulation turn judged as an investigative report). With this
     * option the caller keeps its critic sets in its own codebase and passes
     * the right set per call.
     *
     * Semantics: any count ≥ 1 works (the median handles even counts). Entries
     * that are not a `{ name, prompt }` pair of non-empty strings are dropped —
     * a config-loaded judge list with holes degrades instead of crashing. An
     * empty array, a non-array, or an array with no usable entry falls back to
     * the built-in lookup, so misconfigured callers get v0.11 behaviour instead
     * of judging with zero critics.
     *
     * Judge contract: each prompt must instruct the critic to emit
     * `===SCORE=== N` (or an `X/10` figure) — outputs without either count as
     * a failed critic, and if ALL critics fail the result is `medianScore: 0`.
     */
    criticPrompts?: CriticPromptDef[];
    /**
     * v0.12.0 — Override the output label used in the evaluation preamble
     * ("Evaluate the following {label} for the task …"). Pairs with
     * {@link RunMultiCriticOptions.criticPrompts} for agents outside the
     * built-in `AGENT_OUTPUT_LABELS` map, whose label would otherwise be the
     * generic "output". Ignored when empty/whitespace-only.
     */
    outputLabel?: string;
}
/**
 * v0.7.0 — Strip markdown formatting to plain prose for style-bias-free
 * judging. Preserves the words and line structure; removes only formatting
 * tokens (headers, bold/italic, code fences + inline code, links→text,
 * images→alt, list bullets, blockquotes, table pipes, horizontal rules,
 * strikethrough). Pure, deterministic.
 */
export declare function stripMarkdownForJudging(text: string): string;
/**
 * Get the right critic prompts for an agent.
 * Falls back to investigator prompts for unknown agents (backward-compatible).
 */
export declare function getCriticPrompts(agentName: string): CriticPromptDef[];
/** @deprecated Use getCriticPrompts(agentName) instead. Kept for backward compatibility. */
export declare const CRITIC_PROMPTS: CriticPromptDef[];
/**
 * Run multiple specialized critics and return the median score.
 * Critics are selected based on the agent being evaluated.
 *
 * @param agentOutput - The agent's output to evaluate
 * @param task - The original task description
 * @param runCritic - Function to run a critic (injected, uses Claude CLI)
 * @param agentName - Name of the agent being evaluated (determines which critic set to use)
 */
export declare function runMultiCritic(agentOutput: string, task: string, runCritic: RunCriticFn, agentName?: string, options?: RunMultiCriticOptions): Promise<MultiCriticResult>;
//# sourceMappingURL=multi-critic.d.ts.map