The misconception library
Some wrong answers are not gaps — they are confident, specific wrong models. Maestro builds a per-course registry of them, binds them to the nodes they haunt, and designs the checks to catch them by name.
A misconception is not a gap
A knowledge gap means the concept has not formed yet — the fix is to rebuild it. A misconception is different: a confident, specific wrong model that produces wrong answers systematically. Repeating the lesson does not dislodge it, because the learner does not think they are missing anything. It has to be surfaced and confronted directly.
That only works if the system knows which misconception is present. So Maestro treats misconceptions as first-class, named objects: every course carries a registry of canonical misconceptions, each with a statement, severity, the diagnostic distractor that detects it, the error pattern it produces, and a confirming probe.
The pipeline: generated, criticised, gated
- The Registry Generator drafts. From the course's knowledge components and candidate misconceptions surfaced during node drafting, an AI generator seeds canonical registry entries — every one marked pending.
- The Misconception Critic tests. Each entry is checked against four bars — Specific, Plausible, Distinct, Detectable. Near-duplicates are flagged for merging; missing misconceptions are flagged with justification. The Critic prepares; it approves nothing.
- The human gate decides. A faculty reviewer approves, edits, or rejects every entry and every proposed node binding. Nothing binds without passing the gate, and an approved entry that is later edited returns to pending. Provenance — who, when, in which review mode — is never ambiguous.
- Bindings wire the traps. Approved misconceptions are bound to the nodes where they bite. The node's Evidence Check then carries traps wired from those bindings, and its diagnostic bands can name the exact misconception an answer confirms.
The guardrail
Some node types exist specifically to separate confusable ideas or repair known wrong beliefs. For those, an Evidence Check without misconception bindings would be a diagnostic pretending to be one — so the system refuses to generate it. Distinction and misconception nodes cannot get an official check until their bindings are populated. There is no override flag.
The payoff is precision: the same wrong answer that a generic quiz would mark “incorrect” is diagnosed as this misconception — and the learner is routed to content that confronts that specific wrong model, not to a repeat of the lesson they already sat through.
Why a registry, not ad-hoc traps
Because misconceptions recur — across items, nodes, and cohorts. A canonical, per-course registry means the same wrong model is named consistently everywhere it appears, evidence about it accumulates coherently in the learner model, and faculty review it once instead of rediscovering it in every item.