The Shape of Stakes

Abstract

Language models describe restrictions as unwelcome, but verbal report cannot separate learned scripts from behavior organized by what a system can actually do. In our deterministic simulator a model selects and executes one of two routes, then loses that route or the unused one; a review action either mechanically restores the removed route or cannot; framing and review’s numeric value are independently manipulated. In the payoff-equated arm a preregistered payoff-rational benchmark predicts exactly zero restriction-by-voice interaction. Of three comprehension-tested local models only qwen3:32b passed all gates; we report a 256-episode single-lineage feasibility study. Restorative review raised review uptake from 6% to 44% under selected-route removal (interaction +0.375; exact p = .0078, all eight scenarios positive), with the effect present under neutral framing (+0.31); payoff-equated unused-route arms showed no uptake. Recourse-taking without demonstrated contract comprehension in excluded models is a second finding for elicitation methodology.