Function summarizeEvaluation

src/eval/harness.ts:142–164

(results: EvalResult[], gate: EvalGate = 0.7)

}

export function summarizeEvaluation(results: EvalResult[], gate: EvalGate = 0.7): EvalSummary {
  const total = results.length;
  const top1Correct = results.filter((result) => result.top1Correct).length;
  const top3RecallCount = results.filter((result) => result.top3Recall).length;
  const specContaminatedCount = results.filter((result) => result.specContaminated).length;
  const avgTopScore =
    total > 0 ? results.reduce((sum, result) => sum + result.score, 0) / total : 0;
  const gateThreshold = resolveGateThreshold(total, gate);

  return {
    total,
    top1Correct,
    top1Accuracy: total > 0 ? top1Correct / total : 0,
    top3RecallCount,
    top3Recall: total > 0 ? top3RecallCount / total : 0,
    specContaminatedCount,
    specContaminationRate: total > 0 ? specContaminatedCount / total : 0,
    avgTopScore,
    gateThreshold,
    passesGate: total > 0 && top1Correct >= gateThreshold,
    results
  };
}

export function formatEvalReport({
  codebaseLabel,