An engine that measures how well a body of evidence is actually understood — as a fraction, traced to sources, with the holes named — instead of a confident summary that can't tell you what it's guessing.
When you research anything, you carry a model in your head: what I know for sure, what I'm assuming, what I still need to find out. That model is the real work — and it's invisible, lossy, and trapped in one head.
atomic-reason makes it explicit. The wager underneath is falsifiable: that "understanding" reduces to one mechanical operator over a few primitives — the same operator whether the subject is a bill in Congress, a hospital chart, a legal matter, or a spacecraft's telemetry. Public policy is just the first instance we've built.
A goal sets the denominator — the questions that matter. Each is answered by assertions that must point at evidence, so the engine can't pretend to understand: a guess doesn't count toward the score. Fold the assertions for a question and it lands in one of five states —
Reject a claim and every view recomputes — the percentage, the timeline, the conflicts, the list of what to investigate next. Nothing is re-read or re-trained. You can watch exactly that happen in the live trial.
The two middle pieces — the reasoner and the ledger — are domain-agnostic and reused unchanged. Everything domain-specific is a thin, swappable instance: a vocabulary, some connectors, a goal. Swap them and the same engine runs somewhere new.
The engine runs today as a local research tool. If you have a body of evidence and a real question — a policy fight, a due-diligence file, a literature you need to actually understand — a workspace can be built for it: a goal, a corpus, and a living briefing you correct as you go.
Start with the live trial → to see the engine work end to end on a real question.