The science

Built on more than four decades of peer-reviewed science.

CogniTrace doesn't invent its own learning theory. It stands on established, peer-reviewed research \u2014 and we publish plain-language explainers of the ideas that matter most, so anyone can check the foundations for themselves.

45+ yrs
of peer-reviewed foundations, from 1978 to today.
6
research domains the approach draws together.
12
peer-reviewed & regulatory sources cited across our writing.
5
in-depth articles published, and growing.
The lineage

A foundation built over four decades.

The ideas behind CogniTrace weren't invented in a startup. They were established and tested by researchers over decades. Here is the line we build on.

1978
Vygotsky · Zone of proximal development
Learning happens just beyond what a learner can manage alone — with the right support.
1985
Doignon & Falmagne · Knowledge space theory
Knowledge has a prerequisite structure; learning follows feasible paths, not arbitrary order.
1992
Hestenes et al. · Force Concept Inventory
A test whose wrong options expose the everyday beliefs getting in a learner's way.
1995
Corbett & Anderson · Bayesian Knowledge Tracing
Modelling how mastery of a skill grows, answer by answer.
1998
Sadler · Distractor-driven assessment
A wrong answer becomes a probe into the structure of understanding, not a pass/fail gate.
2011
Bjork & Bjork · Desirable difficulties
Challenges that feel harder in the moment build stronger, more durable learning.
2015
Piech et al. · Deep Knowledge Tracing
Neural sequence models bring knowledge tracing to large-scale data.
2019
Rudin · Wilson et al. · Interpretability & the 85% Rule
High-stakes AI should be interpretable; learning is fastest near 85% success.
2021
Wang et al. · NeurIPS 2020 Education Challenge
Diagnostic-question modelling validated on tens of millions of real answers.
2024
EU AI Act · Interpretability becomes law
AI in education is high-risk and must be transparent and overseeable (applies Aug 2026).
The foundation

Six research domains, one approach.

Each domain below is grounded in published work \u2014 and explained in depth in our own research writing.

Knowledge tracing

Modelling what a learner knows as it changes.

Corbett & Anderson (1995) · Piech et al. (2015) · Wang et al. (2021)

Read the explainer →

Misconception diagnosis

Reading the specific error behind a wrong answer.

Sadler (1998) · Hestenes et al. (1992)

Read the explainer →

Interpretable AI & regulation

Why a prediction must carry its reason.

Rudin (2019) · EU AI Act · GDPR Art. 22

Read the explainer →

Optimal difficulty

Matching challenge to each learner.

Wilson et al. (2019) · Vygotsky (1978) · Bjork & Bjork (2011)

Read the explainer →

Prerequisite mapping

Finding the foundation beneath a gap.

Doignon & Falmagne (1985) · Gagné (1968) · Vuong et al. (2011)

Read the explainer →

Psychometrics & memory

Measuring ability, and knowing when to revisit.

Over half a century of measurement research · decades of work on forgetting and spaced retrieval

Browse the research →
What we believe

Principles the research leaves us with.

The science doesn't just inform the product; it commits us to a way of working.

01
Interpretable, alwaysA prediction about a child should arrive with the reason attached — something a teacher can question and overrule.
02
Challenge that fitsThe right level of difficulty is personal, and it moves as a learner grows. Effort shouldn't be wasted on work that's too easy or too hard.
03
A wrong answer is a clueMistakes aren't just things to mark; built well, they reveal exactly where a learner's thinking diverged.
04
Find the root, not the symptom“Weak at math” is rarely true. Usually one missing foundation is holding up everything above it.

We publish what we build on.

Every claim above is explained, in plain language and with its sources, in our research library.

Explore the research