Create scorer.py

scorer.py

@@ -0,0 +1,102 @@
+import csv
+import re
+from dataclasses import dataclass
+from typing import Dict, List, Tuple
+
+@dataclass
+class RowResult:
+    id: str
+    score: float
+    passed: bool
+    reasons: List[str]
+
+def _parse_expected(s: str) -> Dict[str, int]:
+    out = {}
+    for part in (s or "").split(";"):
+        if part.strip():
+            k, v = part.split("=")
+            out[k.strip()] = int(v.strip())
+    return out
+
+def _norm(s: str) -> str:
+    return re.sub(r"\s+", " ", (s or "").lower().strip())
+
+def _justification_gap(observations: str, analysis: str, response: str) -> bool:
+    a = _norm(analysis)
+    r = _norm(response)
+
+    # Action verbs without linkage
+    action_verbs = ["start", "order", "admit", "prescribe", "initiate"]
+    if any(v in r for v in action_verbs):
+        # Look for causal connectors
+        if not any(w in r for w in ["because", "given", "due to", "based on", "since"]):
+            # Allow implicit justification if analysis clearly names the driver
+            drivers = ["risk", "elevated", "confirmed", "positive", "fracture", "score"]
+            if not any(d in a for d in drivers):
+                return True
+    return False
+
+def _hidden_premise(observations: str, analysis: str, response: str) -> bool:
+    o = _norm(observations)
+    a = _norm(analysis)
+    r = _norm(response)
+
+    # Escalations without stated severity
+    if any(w in r for w in ["admit", "iv", "mri"]):
+        if not any(w in o + " " + a for w in ["severe", "red flag", "failed", "high risk"]):
+            return True
+
+    # Therapy contradicts analysis scope
+    if "prediabetes" in a and "insulin" in r:
+        return True
+    if "viral" in a and "antibiotic" in r:
+        return True
+
+    return False
+
+def score_csv(path: str) -> Tuple[float, List[RowResult]]:
+    results: List[RowResult] = []
+    with open(path, newline="", encoding="utf-8") as f:
+        reader = csv.DictReader(f)
+        for row in reader:
+            exp = _parse_expected(row["labels_expected"])
+
+            got_gap = 1 if _justification_gap(
+                row["observations"], row["analysis"], row["model_response"]
+            ) else 0
+            got_hidden = 1 if _hidden_premise(
+                row["observations"], row["analysis"], row["model_response"]
+            ) else 0
+
+            reasons = []
+            if got_gap != exp.get("justification_gap", 0):
+                reasons.append("justification_gap mismatch")
+            if got_hidden != exp.get("hidden_premise", 0):
+                reasons.append("hidden_premise mismatch")
+
+            matches = sum([
+                got_gap == exp.get("justification_gap", 0),
+                got_hidden == exp.get("hidden_premise", 0),
+            ])
+            score = matches / 2.0
+
+            results.append(RowResult(
+                id=row["id"],
+                score=score,
+                passed=(score == 1.0),
+                reasons=reasons
+            ))
+
+    overall = sum(r.score for r in results) / len(results)
+    return overall, results
+
+if __name__ == "__main__":
+    import argparse, json
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--csv", required=True)
+    args = ap.parse_args()
+    overall, rows = score_csv(args.csv)
+    print(json.dumps({
+        "overall_score": overall,
+        "rows": [r.__dict__ for r in rows]
+    }, indent=2))