One index. Every input visible.

A composite of cited public coding, reasoning, and agent benchmarks. We normalize each exact test variant against the models reported on that test, average within each category, then weight the three categories equally. This is a relative, provisional index—not an absolute measure of intelligence.