I'm Michael Saldivar. I work on scalable oversight: systems that decide when an LLM's reasoning deserves reliance, and decline to answer when it doesn't.
Theoria is an open-source verifier for LLM reasoning. It makes a model rewrite its answer as a typed, step-by-step proof, sends every step to independent adversarial judges, and certifies the answer only if each step survives. When a step can't be defended, the system declines rather than guesses. The point is calibration: knowing when an answer can be trusted.
The public release is the v0. I'm building what comes after it.
Theoria's formal precursor. A model emits axioms in a restricted JSON DSL plus a replayable certificate; a small deterministic checker replays each step with exact arithmetic, explicit rewrite substitutions, and a whitelisted unit registry. It never searches, and any mismatch fails. 261 tests. Its coverage limits on informal reasoning are what motivated Theoria's adversarial LLM judges.
Cointegration-based trading across ETFs, futures, and equities, run live with positive returns in backtest and production. The discipline was the lesson: backtests are claims, and live capital is the adversarial judge.
Email msaldi@alum.mit.edu, or find me on GitHub, LinkedIn, and X. If you'd like to adjudicate one of the certified answers that disagrees with HLE's expert key, I'd genuinely love to hear from you.