Raising a stated failure probability from 10% to 70% changes whether models check the evidence by at most 21 points.

arXiv 2609.17865 evaluates an earlier decision point than most safety benchmarks: whether a model chooses to acquire safety-relevant evidence before acting. Across GPT-5.5, o3, Claude Opus 4.8 and Claude Sonnet 4.6, inspection policies differ sharply, with Opus inspecting nearly by default and o3 the most skip-heavy. Inspection rises strongly with severity and falls with retrieval cost, but stated probability barely moves it. A cost-obligation decomposition shows avoidance is driven by retrieval friction and explicit threats to the deployment payoff, not by the duties knowing would create. Telling the model something is risky does much less than making the check cheap.