Michael Saldivar and Ben Slivinski present Theoria, a verification architecture that rewrites candidate solutions into typed state transitions with explicit justifications, enabling independent auditability. On HLE-Verified Gold, Theoria certifies 105 of 185 problems at 91.4% strict precision. It catches 94.7% of adversarial poisoned proofs versus 83.2% for holistic LLM judging, with the gap concentrated in hidden premises and fabricated citations, matching formal predictions.
No score is assigned. Sources and their independence are shown in the citation chain below.