Research & announcements

In 1983, cognitive scientist Dedre Gentner formalized what separates a good analogy from a seductive one: good analogies map relations between things; bad ones match appearances. Every 'we had one just like this' in a case file is a bet on that distinction — and it is the exact distinction next-token prediction is worst at.

Every vendor demo is fluent — fluency is the one property the technology guarantees. These twelve questions ignore the demo and probe the architecture: can the reasoning be inspected, re-run, bounded, and challenged? Each comes with what a good answer looks like, and what a bad one sounds like.

In 2023, OpenAI researchers compared two ways of training a model to recognize good reasoning: reward the right answer, or reward each right step. Step-level supervision won decisively. Regulators will find the result unsurprising — it is how they have evaluated human decision-making all along.

The epistemic family asks what is true. The problem-solving family asks how to get from here to a defensible there — and its shared enemy is motion mistaken for progress. Five strategies, released together as executable graph shapes: decomposition, means-ends analysis, analogical reasoning, constraint satisfaction, and working backward.

The oldest information format still in production use is the Euclidean proof: definitions, assumptions, numbered propositions, each step warranted by something already established. We use it as an output contract for AI determinations — because it is the only prose format ever devised that makes missing justification structurally visible.

A natural experiment in regulatory architecture: take one model package, adjudicate it under SR 11-7 and under SR 26-2 on the same reasoning engine, and diff the dossiers. The evidence and math are identical; the classification, anchors, and conclusion vocabulary diverge. The diff is a map of where regulation actually lives.