Axiom

Michael Sargis · 2026 · Research project

An autonomous scientist that treats its own ideas as the first thing to distrust.

The idea

Most research agents are very good at making a result sound convincing. I wanted one that would spend its entire compute budget trying to make the result go away.

Axiom proposes a hypothesis, writes a test, and hands the result to a separate model whose only job is to find a less exciting explanation. A claim survives only when both the experiment and the attempted refutation are reproducible.

Results

Our first large run reproduced 2,400 published experiments over a weekend. The useful output was the failure log. It catalogued missing seeds, undocumented filters, and conclusions that depended on one surprisingly load-bearing line of code.

What I built

The experiment scheduler, the adversarial review loop, and a record of every hypothesis, command, and failed run. Each conclusion links back to the exact experiment that supports it.

Back to projects