Axiom Earthlab

Empiraxis is a human-supervised research prototype that explores how AI agents can coordinate reproducible scientific engineering, with evidence and deterministic verification deciding which conclusions are justified. We demonstrate it on the 2015 and 2019 Southeast Asian haze and Singapore PM2.5. The problem: language models can write a confident scientific answer that no experiment supports. In Empiraxis, Meta open-weight models plan the study, propose methods and review evidence. They never compute results. Allowlisted tools acquire data and train models. Independent code scores every candidate against simple baselines such as persistence. A hash-chained trace records each model call, tool call and checkpoint hash. A person makes every final decision. Allowed outcomes include “train”, “keep the baseline”, “abstain” and “request better data”. Empiraxis keeps negative results. In our multisource pilot, satellite fire counts and reanalysis wind did not give a supported gain over PM2.5 history. In our initial five-case evaluation, a council of different models did not outperform the simpler baselines, and Empiraxis records that finding instead of hiding it. Every run can be replayed offline without a model or GPU. This is a research prototype: not a forecast, not health advice and not production automation.