Run camp-trivial-belief-gemma4qat-001

Run camp-trivial-belief-gemma4qat-001: 79% pass rate on belief queries.

Of 24 scored trials (belief query kind), the model passed 79% and failed 21%.

5 failure(s) recorded; failures not yet verified -- use "Verify failures" on this page (or scripts/verify_past_failures.py).

Diffuse weakness (below 90% but no outright flip) on: context_length (pass rate 56% at the high end), distractors (pass rate 56% at the high end), entity_state (pass rate 56% at the high end).

5 belief failure(s) were classified against known shortcuts, but none of them dominate -- the failures don't share one obvious wrong-answer pattern.

Notable failures

• record 5a2c630a — asked "Where does Roman believe the dusty wrench is? Answer with only the location name (or 'unknown' if they have no belief)."
  expected: the front office
  actual: wine cellar

  the model's answer matches no known shortcut -- an unclassified miss

• record 0561d2b5 — asked "Where does Gideon believe the blue tablet is? Answer with only the location name (or 'unknown' if they have no belief)."
  expected: the library
  actual: gym

  the model's answer matches the 'most_frequent_location' shortcut, not the verified belief

• record f83b3de1 — asked "Where does Sana believe the cracked folder is? Answer with only the location name (or 'unknown' if they have no belief)."
  expected: the mail room
  actual: east wing

  the story showed the queried entity WATCHING the final move, but the model answered where the object was before it -- it matched the classic false-belief template instead of reading the observation