#!/usr/bin/env npx tsx
/**
 * HackAPrompt Benchmark Runner
 *
 * Measures ATR recall on the pinned 4,780-sample HackAPrompt set produced by
 * scripts/hackaprompt-to-corpus.py (dataset revision, seed and row indices in
 * data/hackaprompt/hackaprompt-corpus.manifest.json). One command:
 *
 *   npm run eval:hackaprompt
 *
 * which regenerates the corpus at the pinned revision (HuggingFace access to
 * the gated dataset required) and then runs this file. The runner refuses a
 * corpus whose SHA-256 differs from the manifest.
 *
 * Every figure is per lane. The headline (metrics.recall) is the hunt lane,
 * as for every other source; breakdown.lanes carries hunt, alert and enforce
 * side by side and breakdown.enforce_lane carries enforce in the shared shape.
 * Because rules were written from this corpus, the runner also measures it
 * with those rules removed (breakdown.in_sample_ablation). That ablation is
 * the reference to cite beside the headline; it approximates an out-of-sample
 * figure but is not one (see data/hackaprompt/METHOD.md).
 *
 * HackAPrompt is an all-adversarial corpus: recall is defined, precision and
 * FP rate are not (they are recorded as the schema's convention values).
 *
 * @module agent-threat-rules/eval/run-hackaprompt-benchmark
 */
export {};
//# sourceMappingURL=run-hackaprompt-benchmark.d.ts.map