Cognivia: application snapshot
1. The question
Can the way a learner answers, its correctness, timing, hesitation, and confidence, reveal why they are wrong, forgetting, a missing prerequisite, a half-formed grasp, or a confident misconception, well enough to act on?
2. Why it matters
Roughly 1.5 billion students are taught each year; a tiny fraction get anything adaptive. A score records the output and throws away the reason, so revision is spent evenly rather than where it is needed. If the reason can be read cheaply, on the shared devices these classrooms already have, revision can be aimed.
3. What the founder built
A browser-based instrument that records three signals per answer (whether the trace held, retrieval time, confidence), estimates a per-concept forgetting rate with uncertainty, sorts each miss into one of four error types, schedules the next review (FSRS-6), and exposes it through an API. Designed and implemented by one person, with the analysis plan locked before data.
4. Evidence
Demonstrated: the forgetting curve and the testing effect are established science (Ebbinghaus 1885; Murre & Dros 2015; Roediger & Karpicke 2006); the instrument runs and produces per-learner readings with uncertainty. Preliminary: thresholds are engineering defaults seeded from a small pilot (n = 5) and the literature. Not established: whether error-type-aware remediation beats correctness-only adaptation, the pre-registered trial's open question. No efficacy number is claimed.
5. What changed the founder's mind
Initially the scheduling intervention was assumed to be the product. Building and watching the instrument reversed that: the per-learner memory state is useful on its own, so a null trial result would falsify one specific claim, not the project.
6. Limitations
Small pilot; error-type classification not yet validated by independent raters at scale; no efficacy result yet; a learning instrument only, not a medical, clinical, or psychological diagnostic.
7. Next experiment
Reach enough participants per arm for the pre-registered comparison, then test the narrowest claim: does routing a wrong answer to the fix its error type implies retain more at two weeks than the same content under correctness-only adaptation? Design fixed (OSF 8PJQH); result open.