Acting Past the Boundary

Role prompts can make language models more capable by pushing them toward task-specific action. The same pressure can also carry models past the limits that the role itself imposes. We formalize this problem as bounded role adaptation. A model should act when the role permits action, decline when the role forbids it, and ask for missing information when the state is incomplete. We examine this problem in legal counsel settings through matched cases where a client position has plausible support but is blocked by a stated constraint. Across closed and open-weight models, we find that models can identify the blocking constraint in separate questions and still write client-favored briefs when placed in the counsel role. Our framework provides a novel perspective for measuring when role prompting improves task behavior and when it overrides boundaries from situated interaction.