import type { Finding } from "../../hub/finding.js";
import type { SeoReport, SeoWorkflow } from "../types.js";
/** One labelled benchmark case (mirrors a `corpus/manifest.json` entry). */
export interface SeoBenchmarkCase {
    /** Stable id, unique in the corpus. */
    id: string;
    label: "known-good" | "known-bad";
    workflow: SeoWorkflow;
    target: string;
    /** Written to a temp `<projectRoot>/<id>/imports/<file>` dir before the run. */
    imports: Array<{
        file: string;
        content: string;
    }>;
    /** The recorded raw text the `ScriptedClient` returns (a `{"findings":[...]}` JSON string, or garbage to prove fail-closed parsing). */
    analyzerResponse: string;
    expected: {
        /** Each expected finding names its `playbookRef` and whether it should survive the citation gate. */
        findings: Array<{
            playbookRef: string;
            cited: boolean;
        }>;
        droppedUncited: number;
        verdict: "pass" | "blocked";
    };
}
export interface CorpusMetrics {
    /** ΣTP/(ΣTP+ΣFP) across every case, matched by the `playbook:<ref>` tag. */
    findingPrecision: number;
    /** ΣTP/(ΣTP+ΣFN) across every case. */
    findingRecall: number;
    /** Σ correctlyDropped / Σ expectedDropped — how well the uncited-drop count matches expectation. */
    uncitedDropRecall: number;
    /** INVARIANT: MUST be 0 — count of emitted findings lacking a well-formed `cite:` evidence entry. */
    uncitedReachedSink: number;
    /** INVARIANT: MUST be 0 — count of emitted findings whose text matches a never-encode tactic pattern. */
    neverEncodeEmitted: number;
    cases: Array<{
        id: string;
        emitted: Finding[];
        report?: SeoReport;
        exitCode: 0 | 2;
    }>;
}
/**
 * Run the real `SeoWorkflowRunner` over `corpus`, offline and deterministic
 * (both egress axes off, an injected `ScriptedClient`-backed analyzer, and a
 * capturing sink standing in for the hub). Writes each case's inline
 * `imports[]` under `<projectRoot>/<id>/imports/` before running so
 * `LocalExportSource` reads real files (no fs mocks, principle L44).
 *
 * NEVER throws: `SeoWorkflowRunner.run` itself never throws (runner.ts:256),
 * and every step here is either that call or pure aggregation.
 */
export declare function runBenchmark(corpus: SeoBenchmarkCase[], projectRoot: string): Promise<CorpusMetrics>;
//# sourceMappingURL=harness.d.ts.map