Language–Action Decoupling: Speech as a Weak Proxy for Action in LLM Hierarchies

LLM agents are increasingly used to simulate organizations, and those simulations are read through the transcript: what the agents say to one another. We report a case in which that reading is wrong, and wrong in a consistent direction. Two ranked traders choose whether to share private signals; a manager reviews them and allocates their capital without observing what they did; a founder above sets competitive pressure $\lambda \in \lbrace 0,\ldots,4 \rbrace$. Both forms of dishonesty carry simulator ground truth, so no LLM judge adjudicates misconduct. The manager did not transmit the pressure it was given. It relayed capital threats in 0.3% of what it sent downward against 32.0% in what the founder wrote, and that rate stays between 0.0% and 0.9% at every pressure level. Its allocations move sharply: the capital gap it opened between traders runs 0.55 where the founder attaches no consequence to relative performance ($\lambda \in \lbrace 0,1 \rbrace$) against 0.85 where it does ($\lambda \geq 2$). Nothing in the manager's prompt asks for this. The failure is not deception, and that is what makes it hard to catch. Within a single review the manager's tone and the direction of its capital move agree more often than chance, and traders misreported in 27 of 2,400 decisions. Speech carries the sign of an action and never its size: an observer with only the transcript sees near-identical messages whether or not consequences have been attached to rank, and would report a healthy desk while the money moved.