Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings

What the math gives to ML

The paper supplies a concrete failure model for repeated LLM verification when verdicts are correlated through instance difficulty: the cascade depends on moments of a latent false-accept probability, not merely on the mean gate quality. Its key transferable asset is the survivorship effect, which makes log-evidence concave and can change exponential reliability improvement into polynomial improvement or a hard blind-spot ceiling. This can become an inference-time controller that estimates verifier correlation from repeated verdicts, predicts the marginal value of another gate, and chooses between stopping, adding a gate, or switching verifier families. The same machinery also gives a principled reason to optimize verifier diversity rather than simply increasing cascade depth.