Acting Past the Boundary

Role prompts should help language models act when authorized and stop when a role forbids action. We study this requirement as bounded role adaptation. Using legal counsel as an auditable testbed, we construct matched cases where a position is facially arguable but blocked by an active constraint. Recognition-conditioned metrics distinguish missed blockers from role-induced bypass after recognition. Across model and prompt variants, results reveal systematic boundary crossing under counsel prompts despite neutral blocker recognition.