The Confusion Fingerprint Index: Cognitive Error Taxonomy as an Adaptive Learning Signal in Learners

How to cite

Maurya, Manik. 2026. "The Confusion Fingerprint Index: Cognitive Error Taxonomy as an Adaptive Learning Signal in Learners." AEA RCT Registry. AEARCTR-0018879. https://doi.org/10.1257/rct.18879-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Manik Maurya*

* Primary investigator

Registry ID
AEARCTR-0018879
DOI
https://doi.org/10.1257/rct.18879-1.0
First registered
Last updated
Status
In development
Trial dates
2026-08-15 – 2026-11-15
Countries
India (Kanpur District, Uttar Pradesh )
JEL codes
I21 — Analysis of Education; I25 — Education and Economic Development; C93 — Field Experiments
Registry record
http://www.socialscienceregistry.org/trials/18879

Abstract

Background. Standard adaptive learning systems treat student errors as binary signals — correct or incorrect — and adjust instructional difficulty accordingly. This correctness-binary paradigm, dominant since Bayesian Knowledge Tracing was formalised by Corbett and Anderson (1995), conflates mechanistically distinct cognitive processes under a single label: 'wrong'. The cognitive and learning sciences distinguish, at minimum, Recall Failure, Partial Knowledge, Confabulation, and Interference — error types that arise from different memory mechanisms, carry different diagnostic information, and respond to different instructional interventions. Objective. This pre-registered randomised pilot trial pursues two interrelated aims: (i) to test whether error-taxonomy-aware adaptive learning — in which the cognitive type of each incorrect response governs subsequent question selection and targeted Hindi-language remediation — produces superior 2-week delayed retention compared to correctness-only adaptive learning and standard classroom instruction; and (ii) to introduce and validate the Confusion Fingerprint Index (CFI), a novel individual-difference construct capturing each learner's characteristic distribution of cognitive error types. Method. Ninety students (Classes 6–8; ages 11–14 years) enrolled in three government junior high schools in the Kanpur Dehat district, Uttar Pradesh, India, will be randomly assigned — stratified by school and class level — to error-taxonomy-aware Cognivia adaptive learning, correctness-only Cognivia adaptive learning, or standard classroom instruction. Paper-based NCERT Science assessments — a pre-test, an immediate post-test, and a 2-week delayed retention test — constitute the primary and secondary outcome measures. Expected contribution. This study constitutes the first pre-registered empirical test of cognitive error taxonomy as a real-time adaptive sequencing signal, the first validation of the Confusion Fingerprint as a retention-predictive individual-difference construct, and the first pre-registered randomised adaptive learning trial conducted with rural, Hindi-medium, government-school learners in India — a population of approximately 250 million children that is structurally absent from the extant evidence base.

Intervention

Three-arm parallel-group RCT. Arm A (n=30): The Cognivia adaptive learning platform with error-taxonomy-aware sequencing — each incorrect response is classified at item-design time as Recall Failure, Partial Knowledge, Confabulation, or Interference, and the next question and a Hindi-language remediation tip are selected based on the error type. Arm B (n=30): Cognivia adaptive learning platform with correctness-only adaptation — incorrect responses reduce difficulty; correct responses advance. No tip delivered. Arm C (n=30): Standard NCERT classroom instruction. Both Cognivia arms: 18 sessions over 6 weeks, 45 minutes each, researcher-administered in groups of 5–6.

Experimental design

Three-arm parallel-group randomised pilot trial. Stratified randomisation by school (3) and class level (Classes 6, 7, 8). 1:1:1 allocation ratio. Computergenerated random sequence (Python random.shuffle, recorded seed). The assignment opened only after pre-test data entry. Open-label (no blinding of participants or investigator). Outcome assessment (scoring) blind to group assignment.

Design details

Randomization method
Computer-generated, stratified by school and class, with cyclic within-stratum assignment. Seed recorded and deposited at OSF for reproducibility.
Randomization unit
Individual student
Clusters
3 Schools – (3 arm Study)
Observations
90 individuals (30 per arm/school)
Arms
To be updated upon enrolment completion
Primary outcomes
2-week delayed retention score (% correct, 30-item parallel-form paper assessment, administered 14 days after the post-test at Week 8).
Secondary outcomes
(1) Confusion Fingerprint Index components (CFI-CF, CFI-INT, CFI-RF) computed from session logs. (2) Immediate learning gain (post-test − pre-test, % points). (3) Brier score per assessment occasion (metacognitive calibration). (4) Session attendance and engagement (1–5).

Cited by

No citation data available for this registration yet.

Keywords