Theoria Verifies AI Reasoning with Auditable Proof Traces

Ben Slivinski, Michael Saldivar· July 2, 2026 View original

Summary

Theoria is a verification architecture that bridges the gap between formal proof assistants and LLM judges by rewriting AI solutions into auditable sequences of justified state transitions. It ensures completeness of change, surfacing hidden premises and achieving high precision in verifying informal reasoning.

Trusting AI system answers, especially in complex reasoning tasks, remains a significant challenge. While formal proof assistants offer certainty, they lack broad applicability, and scalar LLM judges provide coverage but suffer from opacity and potential coherence issues. Theoria introduces a novel verification architecture designed to overcome these limitations by providing auditable and transparent reasoning.Theoria transforms a candidate AI solution into a series of typed state transitions, each explicitly justified by a citation, computation, or problem fact. A core principle is "completeness of change," meaning every difference between consecutive states must be accounted for, thereby exposing any hidden premises or unjustified mutations. This approach yields human-readable proof traces where each step can be independently challenged. Theoria demonstrated high precision on expert problems and proved significantly more effective at catching adversarial poisoned proofs, particularly those involving hidden premises and fabricated citations, compared to holistic LLM judges.

Why it matters

Professionals deploying AI in critical applications need verifiable and auditable reasoning processes to build trust and ensure compliance, moving beyond opaque "black box" AI decisions.

How to implement this in your domain

  1. 1Investigate integrating verification architectures like Theoria into AI systems requiring high trust and auditability.
  2. 2Develop internal standards for explicit justification and completeness of change in AI-generated reasoning.
  3. 3Pilot Theoria or similar frameworks for validating AI outputs in sensitive domains.
  4. 4Train AI development teams on designing systems that produce auditable proof traces.

Who benefits

BFSILegalHealthcareDefenseSoftware Development

Key takeaways

  • Theoria provides a verifiable and auditable architecture for AI reasoning.
  • It transforms AI solutions into explicit, justified state transitions.
  • The "completeness of change" invariant exposes hidden premises and unjustified steps.
  • Theoria significantly outperforms holistic LLM judges in detecting adversarial reasoning errors.

Original post by Ben Slivinski, Michael Saldivar

"arXiv:2607.01223v1 Announce Type: new Abstract: When should an AI system's answer be trusted? Formal proof assistants offer certainty but cannot reach most of the problem distribution; scalar LLM judges offer coverage but produce opaque scores that cannot be audited after the fac…"

View on X

Originally posted by Ben Slivinski, Michael Saldivar on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses