main experiments

experiments

Questions put to many models at once, with the method, the controls and every score shown. Each set has its own page.

loading…

Checks on the results

Verification of the best candidates

done17 Oct 2026 · 10 items · 3 verifier models · 30 reviews

The 5 best candidates of each set, re-derived from scratch by Claude Opus 5.5, GPT-6.1 and Fable 5.1, with counterexample searches and proof attempts. None was found wrong; most are variants of known results.

standout: “in a stationary reversible hidden Markov chain, matching the 2- and 3-term sum laws forces the whole output to be independent”: correct and new as far as all three verifiers know; proof checked by the agent.

open →

Re-scoring with different judge panels

done17 Oct 2026 · both sets · no new model calls

How much the ranking depends on who judges: without the most lenient judge, with only the strictest ones, and leaving out one judge at a time. Two items per set stay in the top 10 under every panel.

open →