K2

K² · Artificial intelligence

Auditable AI verification catches hidden-assumption errors that scalar judges miss

The authors present Theoria, a verification architecture that rewrites a candidate solution into a sequence of typed state transitions, each licensed by an explicit justification such as a citation, computation, or problem-given fact.

BS
MS

Ben Slivinski, Michael Saldivar

2 authors · cs.AI, cs.CL, cs.LG, cs.LO, cs.SE

arXiv preprintArtificial intelligenceJul 2026 · ~55s read

The 30-second scan

When an AI hands you an answer, how do you know it's right?

  1. Theoria's foundational invariant is completeness of change, requiring every difference between consecutive proof states to be accounted for so that hidden premises surface as unlicensed mutations.
  2. On HLE-Verified Gold (185 text-only expert problems), Theoria certifies 105 problems at 91.4% strict precision (Wilson 95% CI [84.5%, 95.4%]).
  3. Every certification produces a human readable proof trace in which each step can be independently challenged.
HLE-Verified Gold consists of 185 text-only expert problemsTheoria certifies 105 of the HLE-Verified Gold problems at 91.4% strict precision (Wilson 95% CI [84.5%, 95.4%])Holistic LLM judges fail on different problems (Jaccard 0.14-0.36)
Read the originalPDF