Invigil

Notes

Working notes on measuring reasoning without trusting anyone.

Written because we have to think this through anyway, and because a claim you publish is one someone can check.

15 Sep 2026

Why the model never writes the theorem.

Most formal-reasoning benchmarks check whether a model's file compiles. That leaves the model free to change the question. The single design decision that closes that hole, and why it matters more than anything else in the harness.