| 124 | |
| 125 | |
| 126 | def _attach_delta(result: dict[str, Any], baseline: dict[str, Any]) -> dict[str, Any]: |
| 127 | enriched = dict(result) |
| 128 | enriched["delta_vs_baseline"] = { |
| 129 | "tier_match_accuracy": result["tier_match_accuracy"] - baseline["tier_match_accuracy"], |
| 130 | "pass_rate": result["pass_rate"] - baseline["pass_rate"], |
| 131 | "exact_match": result["exact_match"] - baseline["exact_match"], |
| 132 | "pass_count": result["pass_count"] - baseline["pass_count"], |
| 133 | } |
| 134 | return enriched |
| 135 | |
| 136 | |
| 137 | def _print_group(title: str, results: list[dict[str, Any]]) -> None: |