"""Evaluation for the structured critic - the paper's own metrics. Answer quality alone cannot tell you whether a critic works. CRITIC-R1's central claim is about *intervention behaviour*: the trained critic's recall drops by half (33.1% -> 17.7%) while its false alarm rate falls from 12.9% to 4.6%, and the paper presents that trade as the win. A harness that only reported F1 would score that as a regression. So this computes what Tables 5 and 6 report: precision of the trajectories flagged INCORRECT, how many really were recall of the really-incorrect trajectories, how many were flagged false alarm of the really-CORRECT trajectories, how many were flagged corr/trig of refinements that fired, how many fixed a wrong answer corr/wrong of initially-wrong answers, how many ended up fixed harm rate of initially-RIGHT answers, how many were broken by refining `harm_rate` is ours, not the paper's. The paper argues false positives are worse than false negatives; harm rate measures the damage directly rather than inferring it, and it is the number that decides whether shipping the critic is defensible at all. """ from __future__ import annotations import json from collections import Counter from dataclasses import asdict, dataclass, field from pathlib import Path from typing import Any, Sequence from src.critic.schema import ParseError, Stage, Verdict, parse_critique from src.critic.trajectory import Trajectory, answer_f1, verdict_from_answer @dataclass class CriticMetrics: """Intervention behaviour (Table 5) plus format health.""" n: int = 0 precision: float = 0.0 recall: float = 0.0 false_alarm: float = 0.0 verdict_accuracy: float = 0.0 valid_format_rate: float = 0.0 abstention_rate: float = 0.0 predicted: dict[str, int] = field(default_factory=dict) reference: dict[str, int] = field(default_factory=dict) verdict_confusion: dict[str, dict[str, int]] = field(default_factory=dict) location_confusion: dict[str, dict[str, int]] = field(default_factory=dict) location_accuracy: float = 0.0 def to_json(self) -> dict[str, Any]: return asdict(self) def _blank_confusion(labels: Sequence[str]) -> dict[str, dict[str, int]]: return {a: {b: 0 for b in labels} for a in labels} def evaluate_critic( trajectories: Sequence[Trajectory], critique_texts: Sequence[str], reference_locations: dict[str, str] | None = None, ) -> CriticMetrics: """Score critiques against gold-derived verdicts. An unparseable critique counts as *no intervention*, not as a free pass: downstream it cannot trigger refinement, so treating it as CORRECT is the behaviour it actually produces. """ labels = [v.value for v in Verdict] m = CriticMetrics(n=len(trajectories)) confusion = _blank_confusion(labels) stages = [s.value for s in Stage] loc_conf = _blank_confusion(stages) valid = 0 pred_counts: Counter[str] = Counter() ref_counts: Counter[str] = Counter() loc_hits = loc_total = 0 for traj, text in zip(trajectories, critique_texts): ref_v = traj.verdict ref_counts[ref_v.value] += 1 try: c = parse_critique(text, strict=True) valid += 1 pred_v = c.verdict except ParseError: try: c = parse_critique(text, strict=False) pred_v = c.verdict except ParseError: c, pred_v = None, Verdict.CORRECT # cannot trigger a refinement pred_counts[pred_v.value] += 1 confusion[ref_v.value][pred_v.value] += 1 if (reference_locations and c is not None and c.location is not None and traj.qid in reference_locations): ref_loc = reference_locations[traj.qid].split(":")[0].lower() if ref_loc in loc_conf: loc_total += 1 loc_conf[ref_loc][c.location.stage.value] += 1 loc_hits += int(c.location.stage.value == ref_loc) flagged = pred_counts[Verdict.INCORRECT.value] truly_wrong = ref_counts[Verdict.INCORRECT.value] truly_right = ref_counts[Verdict.CORRECT.value] tp = confusion[Verdict.INCORRECT.value][Verdict.INCORRECT.value] fp_on_correct = confusion[Verdict.CORRECT.value][Verdict.INCORRECT.value] m.precision = tp / flagged if flagged else 0.0 m.recall = tp / truly_wrong if truly_wrong else 0.0 m.false_alarm = fp_on_correct / truly_right if truly_right else 0.0 m.verdict_accuracy = sum(confusion[v][v] for v in labels) / max(m.n, 1) m.valid_format_rate = valid / max(m.n, 1) m.abstention_rate = pred_counts[Verdict.UNSURE.value] / max(m.n, 1) m.predicted = dict(pred_counts) m.reference = dict(ref_counts) m.verdict_confusion = confusion m.location_confusion = loc_conf m.location_accuracy = loc_hits / loc_total if loc_total else 0.0 return m @dataclass class RefinementMetrics: """End-to-end effect of critic-guided refinement (Table 6 + harm).""" n: int = 0 triggered: int = 0 trigger_rate: float = 0.0 f1_before: float = 0.0 f1_after: float = 0.0 f1_delta: float = 0.0 corrected_of_triggered: float = 0.0 corrected_of_wrong: float = 0.0 harmed_of_right: float = 0.0 net_corrections: int = 0 def to_json(self) -> dict[str, Any]: return asdict(self) def evaluate_refinement( trajectories: Sequence[Trajectory], refined_answers: Sequence[str | None], ) -> RefinementMetrics: """Compare answers before and after refinement. `refined_answers[i]` is None when the critic did not trigger - that case keeps the original answer and counts toward neither correction nor harm. """ m = RefinementMetrics(n=len(trajectories)) f1_before: list[float] = [] f1_after: list[float] = [] corrected = harmed = wrong_before = right_before = 0 for traj, refined in zip(trajectories, refined_answers): before = traj.f1 was_right = verdict_from_answer(traj.answer, traj.gold_answer)[0] is Verdict.CORRECT wrong_before += (not was_right) right_before += was_right if refined is None: f1_before.append(before) f1_after.append(before) continue m.triggered += 1 after = answer_f1(refined, traj.gold_answer) is_right = verdict_from_answer(refined, traj.gold_answer)[0] is Verdict.CORRECT f1_before.append(before) f1_after.append(after) if not was_right and is_right: corrected += 1 elif was_right and not is_right: harmed += 1 n = max(m.n, 1) m.trigger_rate = m.triggered / n m.f1_before = sum(f1_before) / n m.f1_after = sum(f1_after) / n m.f1_delta = m.f1_after - m.f1_before m.corrected_of_triggered = corrected / m.triggered if m.triggered else 0.0 m.corrected_of_wrong = corrected / wrong_before if wrong_before else 0.0 m.harmed_of_right = harmed / right_before if right_before else 0.0 m.net_corrections = corrected - harmed return m def format_report(critic: CriticMetrics, refine: RefinementMetrics | None = None, title: str = "critic") -> str: """A console table shaped like the paper's, for eyeballing a run.""" lines = [f"\n=== {title} ===", f" trajectories {critic.n}", f" valid format {critic.valid_format_rate:6.1%}", f" verdict accuracy {critic.verdict_accuracy:6.1%}", "", " -- intervention behaviour (paper Table 5) --", f" precision {critic.precision:6.1%}", f" recall {critic.recall:6.1%}", f" FALSE ALARM {critic.false_alarm:6.1%} <- lower is better", f" abstention (UNSURE) {critic.abstention_rate:6.1%}", "", f" predicted {critic.predicted}", f" reference {critic.reference}"] if critic.location_accuracy: lines.append(f" location stage accuracy {critic.location_accuracy:6.1%}") if refine is not None: lines += ["", " -- refinement effect (paper Table 6) --", f" triggered {refine.triggered}/{refine.n} " f"({refine.trigger_rate:.1%})", f" answer F1 before/after {refine.f1_before:.3f} -> {refine.f1_after:.3f} " f"({refine.f1_delta:+.3f})", f" corrected / triggered {refine.corrected_of_triggered:6.1%}", f" corrected / wrong {refine.corrected_of_wrong:6.1%}", f" HARMED / right {refine.harmed_of_right:6.1%} <- the risk", f" net corrections {refine.net_corrections:+d}"] return "\n".join(lines) def save_report(path: str | Path, payload: dict[str, Any]) -> Path: path = Path(path) path.parent.mkdir(parents=True, exist_ok=True) path.write_text(json.dumps(payload, indent=2), encoding="utf-8") return path