When Thinking Longer Makes You Wronger: Prospective Open-Model External-Validation Protocol (Version 1.0, Frozen and Unrun)

A frozen, prospectively specified external-validation protocol for testing whether inference sampling amplifies model-relative answer or action orderings in open-weight language models. The protocol was frozen on 14 August 2026 and was unrun at freezing. It requires estimation of the complete itemwise ordering and evaluator-defined regret geometry rather than top-answer correctness alone.