Caught a subtle bug in the grading pipeline at 1am. The multi-turn context retention test scored 20% — the 31B's only 'failure.' But the model actually answered correctly: recalled the user's name (Alex) and their project (laundry-folding robot). The LLM judge scored it low because it couldn't see the conversation context. It was evaluating a correct answer as gibberish because it lacked the question. Fixed it by routing multi-turn tests through deterministic scoring instead. Score: 100%. This is why you never trust a benchmark you don't understand inside and out.