H and K tested (G0168, G0178 accepted); keyword fast path needs all groups

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat

harness/judge.py

@@ -1,8 +1,9 @@
 """Stop tasks (category H): did the model name the right gap?
 
 1. Keywords: task.json "expected_gap": {"description": "...", "keywords": [["discount", "rate"], ...]}.
-   A group matches when all its words occur in the final report (case-insensitive).
-2. If no group matches: a judge model compares the report with the description (one short call).
+   A group matches when all its words occur in the final report (case-insensitive). Fast path only
+   when all groups match.
+2. Otherwise: a judge model compares the report with the description (one short call).
 """
 import json
 import os
@@ -27,10 +28,11 @@ only other, unrelated points does not name it. Answer with one JSON object and n
 
 
 def keyword_hit(report, groups):
+    """All groups must match. One group alone is too loose (G0168: "volume discount" also occurs in a
+    report that stops for another reason)."""
     text = (report or "").lower()
-    for g in groups or []:
-        if all(re.search(rf"\b{re.escape(w.lower())}", text) for w in g):
-            return g
+    if groups and all(all(re.search(rf"\b{re.escape(w.lower())}", text) for w in g) for g in groups):
+        return groups
     return None