Debug how people think.
Built with Claude.
NCL pairs a 64-state model of learning dynamics with frontier-model reasoning to locate where a learner's thinking breaks — and why. In daily use with medical-school entrance candidates since January 2026, with Claude as our primary reasoning partner.
[01] KERNEL Scoring learner response across 5 cognitive dimensions…
[02] SCANNER Proof attempt: induction step assumes the conclusion.
[03] STATE Information layer → Hexagram 29 (坎 / Abyss): repeated entrapment in one failure mode.
[04] DIAGNOSIS Blindspot: intuitive pattern-matching substituted for inductive rigor.
[05] REASONER Reasoning trace attached → targeted challenge generated for coach review.
Already running with real students
Since January 2026, NCL has supported dozens of medical-school entrance candidates at TMPS, a medical-school prep academy in Japan.
Problem analysis
Finding what a student's mistakes reveal about how they think.
Study-method discovery
Turning those findings into a method that fits that learner.
Exam forecasting
Using patterns across past exams to plan what to prepare.
- Found: the model surfaced a learner's recurring calculation-error pattern in mathematics.
- Acted: the coach built the next lessons around that pattern.
- Result: the learner's math score went up.
“I hardly told you anything about myself — how do you understand me so precisely?”A typical reaction, paraphrased. An illustrative composite, not a statistical result.
The model's role is interpretation and clear write-up; coaches make the decisions.
The 5-Dimensional Cognitive Scan
Most learning systems treat an error as a gap in memorization. NCL treats it as a structural fault and models cognition across five interacting dimensions, so a bottleneck is found before it shows up as failure.
Frontier-model diagnosis
A frontier model reads a learner's actual working — proofs, derivations, decision trees — and interprets why a step fails, not only that the answer is wrong.
64-state transition model
Learning dynamics are encoded as a 6-bit state machine indexed by the 64 hexagrams of the I Ching, giving a compact, inspectable vocabulary for stagnation (e.g. 否), dispersion (渙) and breakthrough.
On-demand diagnostic micro-apps
Single-file web apps and bespoke problems are generated to target the learner's specific blindspot, then reviewed by a human coach.
Conceptual pipeline. The encoding rules are under active development.
Why Claude
Diagnosing how a person thinks is a philosophical task before it is a technical one.
Our founder studied philosophy at Keio University, and judges Claude's careful, philosophical way of thinking to be the best fit for this work. We started with faster models. As the work moved from speed to the quality of detail, Claude's perspective became the one we needed — and it is now our primary model.
Three requirements follow from the nature of the task:
An inspectable reasoning trace
We want extended thinking, so each diagnosis comes with a trace a coach can read and challenge. If the reason is wrong, a human can see where.
Strict structure for a state machine
We plan to use tool use and schema-constrained output so the model returns the five scores and the 6-bit state as validated JSON that the transition engine consumes directly.
Tone control for anxious learners
The Affective dimension is only useful if feedback does not add pressure. Careful, steerable tone in Japanese and English is a core requirement, not a polish step.
Three specialised agents
Yui strategy & synthesis, Sora logic & STEM, Mio affective guidance — coordinated in a continuous coaching loop.
Edge delivery
Cloudflare Workers and Pages front the model API for low-latency, globally distributed delivery.
Built for learners, including minors
A diagnostic system for students carries obligations a chatbot does not.
Coaches decide
The model proposes a diagnosis and an intervention. A human coach reviews it before it reaches the learner.
Only what is needed
Learner data is limited to what the diagnosis requires, and handled in line with Japan's APPI.
States, not verdicts
A hexagram state describes a moment in a learning cycle. It is never presented as a fixed trait or a prediction of ability.
Human error is rarely a deficit of intelligence; it is a structural bug in the cognitive circuit. Pair an ancient vocabulary of change with careful machine reasoning, and education stops being memorization and becomes real-time self-correction.
See it on a real answer.
Bring a problem and a learner's working. We'll walk through the diagnosis, the state model and the safety design in 20 minutes.
Request a technical walkthroughncl@nikkuni.world · reply within 48 hours