Unfolding the rules of reasoning.

The problemIntelligence is not just knowledge; it is the ability to reason, solve novel problems, and transfer across tasks. Today, frontier models saturate every benchmark but still reason wrong. Labeling a model's answers measures output accuracy, but doesn't tell us anything about how the model got there: reasoning is a structured process, governed by the quality of evidence, the validity of causal claims, calibration under uncertainty, and revision when the facts push back. None of that is visible in a final answer. Reasoning quality under uncertainty is the unmeasured bottleneck for the next capability jump.

The hypothesisA century of cognitive science and neuroscience, from the first studies of general intelligence to modern work on the reasoning systems of the human brain, points one way: reasoning is not a bag of domain tricks. It is a general capacity, constrained by logic and rules, that transfers across mathematics, science, engineering, and beyond. Those rules can be formalized, validated against calibrated expert judgment, and shown to hold from one domain to the next. We are treating this as the major hypothesis to be tested. If it holds, reasoning quality becomes a measurable property: a multidimensional profile of how a system thinks, and not a single number or label.

The teachersMeasurement produces more than scores. From a growing corpus of expert-scored evaluations we develop teacher models: systems that internalize the constraints experts apply when they reason well, so those constraints can be detected, taught, and enforced at scale. Models should not merely reproduce expert answers; they should learn the principles behind them. That is the bridge from natural intelligence, to expert judgment, to machines that actually can reason.

The goalWhat is measurable can be diagnosed, and what can be diagnosed can be repaired. That is our goal: measure the reasoning, locate the flaw, fix the system. You can't label your way to AGI: you have to empower reasoning.

Adversarial by designEvery evaluation is a multi-turn exchange in which a domain expert presses a model to the limits of its reasoning: follow-ups, edge cases, counter-claims, premises that deserve to be challenged. A quiz doesn't cut it, we need a cross-examination. The question is never only whether the answer is right, but whether the reasoning behind it would survive the questionnaire of an auditorium of experts.

Experts, augmentedExperts do not work alone. Purpose-built agents work beside them through every evaluation: surfacing evidence, pressure-testing follow-ups, keeping the protocol tight. And as the corpus of expert-scored evaluations grows, we distill evaluator agents that pre-screen reasoning at scale, while experts validate, correct, and sign off on every score. The agent is the instrument; the experts own the call.

We don't just grade a model, we capture its profileEach reasoning trace is scored along multiple dimensions along multiple turns rather than collapsed into a single number. A system can be strong on causal structure and weak on calibration, it can be confident exactly where it should hesitate. The profile shows where reasoning holds, where it bends, and where it breaks.

Versioned JudgmentEvery score carries its evidence: who evaluated, under which protocol, against which state of the field. Science moves, and the record moves with it. Every result can be replayed and audited, years later, by people who were not in the room.

Structurally IndependentLabs cannot credibly grade their own reasoning, just as pharma cannot run its own trials. Independence here is not a policy or a promise. It's the architecture.

Frontier labs, scientists, engineers and regulators who want to be part of the next big step in AI: write to us.