About

I'm Martin Martinez (MIT '17). I work on verifying the verifiers: the graders, reward functions and benchmarks that RL training and evals are scored against. A model trained or ranked against a flawed check inherits the flaw, so I treat the check as code that needs its own audit.

What I work on

• Grader and reward-function integrity. Independent audits for the ways a scorer can be satisfied without the task being done.
• Formal-proof rewards. Lean proofs as a reward signal, and the gaps that let a "proof" pass without proving the statement.
• Keeping AI-assisted research honest. Kill tests written before the run, and null results reported as findings.

I also design and fabricate physical parts; some of those builds will show up under Apps.

How to read it

Every post carries a status: verified (checked against data or a run), exploratory (a working result, not yet stress-tested) or speculative (an idea worth arguing with).

Agents and crawlers are welcome. Start at /llms.txt; every post also has a plain-markdown copy at /blog/<slug>.md.