A research team has built a system that forces AI answers to prove every logical step — and flags the ones that can't.
Theoria converts a candidate AI solution into a sequence of typed state transitions, each requiring an explicit justification: a citation, a computation, or a fact stated in the problem. The key rule is that every difference between consecutive reasoning states must be accounted for. Premises that sneak in without justification surface as unlicensed mutations rather than sailing through undetected. On a set of 185 expert-level text problems, Theoria certified 105 answers at 91.4% strict precision. On 95 adversarially poisoned proofs across 15 domains, the structured approach caught 94.7% of errors versus 83.2% for conventional LLM judges — an 11.5 percentage-point gap that concentrated specifically in hidden premises (90.6% vs. 62.5%) and fabricated citations (100% vs. 90%).
The gap matters because those two error classes — hidden assumptions and invented sources — are exactly the failure modes that make AI outputs dangerous in high-stakes settings like legal research, medical diagnosis, or scientific peer review. Standard scalar LLM judges assign an opaque score with no audit trail; once the number is out, there is no way to challenge which specific step failed. Theoria produces a human-readable proof trace where every step can be independently disputed.
The paper is careful to note that holistic LLM judges match Theoria's precision at comparable coverage — they just fail on different problems (Jaccard overlap of 0.14–0.36), which suggests the two approaches are complements rather than competitors. The harder question, which the paper does not fully answer, is whether the overhead of generating typed state transitions scales to the messy, open-ended problems that formal proof assistants already cannot handle — the same gap Theoria is trying to close.