Why 'one wrong answer ≠ one unknown knowledge point'
The three-layer architecture of error attribution: rules → statistics → LLM.
What we're validating: Adaptive vs random question selection
Experiment grouping is written into the data model from day one.