correlated-errors

Open-source toolkit for AI agent verification: measure how many independent votes your review panel (LLM judges, CI checks, human reviewers) is really worth, audit review gates for correlated errors, and keep reproducible evidence for every accept/reject.

A diverse 6-LLM judge committee (4 labs, 3 countries) has correlated errors: ~1.9 effective independent judges of 6 (mean pairwise error-correlation ρ≈0.42) on RAGTruth. Provider diversity does not buy error-independence. Verdicts + one-command recompute included.

Antagonistic consensus for LLM agents: theory, simulation, and empirical anchoring under correlated errors.

Measure error correlation in an LLM judge panel or voter ensemble, report the effective number of independent votes, and correct the majority vote for correlated voters — including the panel shape where it cannot. Closed-form, label-free, numpy only.