S3R: registered head-to-head of the best-value rule and RouteLLM routers, replayed on saved answers (registration)

Question Does a published, learned router (RouteLLM, routellm 0.2.0) choose between the cheap model and the strong model better than the registered best-value rule, when both are scored by the paper's utility (right minus wrong minus cost over V) on the same held-out test questions? “Registered before any router score on a test item was computed. Answers were bought in earlier registered studies and are replayed; no new model answers are bought. Files frozen by sha256 (FROZEN_HASHES.txt).”