Our new fellow takes on verification in science

Scientific text is now cheap to produce. Proposals, hypotheses and reviews can arrive faster than people can read them. The challenge is checking their claims. Our new fellow, Michael Saldivar, argues that an AI system should show its reasoning step by step, with each step tied to a source, computation or stated premise. If it cannot support a claim, it should decline to answer.

An AI researcher from Fresno with a mathematics degree from MIT, Saldivar built Theoria, an open-source system that makes language models prove answers step by step or abstain. His paper describes the method. A follow-up study on open-weight models was accepted to the NeurIPS 2026 MATH-AI workshop. Before this, he worked as a data analyst and analytics engineer at Insurify.

From proposals to open questions across fields

During his fellowship, Michael will build an open-source screening workflow for research proposals. It will identify each proposal's central question, try to answer it with a step-checked proof, and search the literature from broad to narrow, counting and linking every query. Additionally, Michael will apply the same tools to whole fields: identifying the open questions those fields state about themselves, then placing proposals in relation to those questions. Across both parts of the work, the test stays the same: every claim should be traceable to evidence or a stated premise, and the system should leave unsupported claims unanswered.

Try Theoria

The interactive demo is at theoriaverified.com. The code is open at github.com/zaladbar/theoria, and the paper is at arxiv.org/abs/2607.01223.