Distinct mathematical misconception rules can produce the same incorrect answer, making diagnosis from a single response ambiguous. We formalize this algebraic collision and derive a conditional Bayes error bound; internally generated reasoning does not add student-specific information beyond its inputs under an explicit conditional-independence assumption. We propose Mondrian Adaptive Conformal Misconception Sets (MACMS) and Symbolic-Verified Counterfactual Micro-Probing (Sym-EIG). MACMS has curriculum-group marginal coverage under exchangeability, not a per-student correctness guarantee. Sym-EIG searches for probes whose outputs separate modeled candidate rules; identification additionally depends on the candidate set and response model. An audit of a local public Eedi graph finds 2,118 of 2,258 linked constructs associated with at least two misconception IDs; construct-level associations alone do not establish same-answer collisions. Other project performance tables require recomputation because their script assigns numerical results directly. No diagnostic performance or classroom learning benefit is claimed here. A classroom study is proposed for future validation.
翻译:暂无翻译