meo Benchmark Domains

meo scores models across eleven reasoning domains, each targeting a distinct failure mode — from base-rate bias and theory of mind to long arithmetic and tape-machine simulation. A model’s per-domain profile shows where it genuinely reasons versus where it pattern-matches.

For several domains (long arithmetic, regex & automata, tape-machine, unsatisfiable constraints, state tracking) every item is generated together with its programmatic ground-truth answer. That yields effectively unlimited fresh, self-verifying problems that cannot be leaked or memorized.

Because a single aggregate score hides where a model is strong or weak. The leaders differ sharply by domain, so per-domain rankings let you match a model to the kind of reasoning your workload actually needs.