/**
 * Frozen snapshot of the demo evaluators, captured byte-for-byte from the
 * owner's deployed evaluators (graders + Blockly workflow task graphs).
 *
 * Source of truth for `respan demo`. Each entry is one user-facing evaluator: the
 * deployed workflow's task graph plus the grader record(s) it references. The
 * grader `type` is captured verbatim ("human" is the stored default) — the actual
 * LLM engine lives in each eval task's `generation_method`/`_blockly_evaluator_kind`,
 * so replaying the graph reproduces the verified scoring behavior exactly.
 *
 * This is a frozen snapshot: if the source evaluators change, this file does NOT
 * update until someone re-snapshots and ships a new build.
 *
 * Recreate recipe (see seed-evaluators.ts): create each grader -> map its
 * `originalId` to the new id -> rewrite every eval task's `config.evaluator_id`
 * via that map -> createWorkflow -> commit -> deploy. Task ids are kept verbatim
 * so the compute task's `state.<task_id>` input references stay valid.
 */
/** A grader record referenced by an evaluator's workflow. */
export interface DemoEvaluatorGrader {
    name: string;
    /** The grader's id inside the captured task graph. Remapped to the freshly
     * created grader id at seed time so the eval tasks point at the new record. */
    originalId: string;
    type: string;
    score_value_type: string;
    score_config: Record<string, unknown>;
    passing_conditions: Record<string, unknown>;
    llm_config: Record<string, unknown>;
}
/** One demo evaluator: the deployed workflow graph plus its grader(s). */
export interface DemoEvaluatorFixture {
    /** Workflow name (the user-facing evaluator name); used for dedup. */
    workflowName: string;
    description?: string;
    graders: DemoEvaluatorGrader[];
    /** Verbatim deployed workflow task graph. Only each eval task's
     * `config.evaluator_id` is rewritten at seed time. */
    tasks: Record<string, unknown>[];
}
export declare const DEMO_EVALUATORS: DemoEvaluatorFixture[];