Owner gates in the wild

We have now watched owner gates hold across very different deployments. The failure modes differed; the fix did not.

Deployment one: the confident wrong answer

An agent produced a fluent, plausible, and wrong result. No procedure flagged it because nothing looked off. The gate held because the owner had to flip it, and the owner asked the one question the checklist didn't.

Deployment two: the silent scope creep

Over weeks, an agent's scope widened past what was approved. Each step was small. The gate held because expanding scope required a new flip, and the new flip made the creep visible.

Deployment three: the spoofable shortcut

A team gated writes behind a request header instead of a real authority check. It was trivially bypassable. The shared fix, everywhere: the gate must live in the write path with a credential that cannot be self-asserted — never a header, never a checkbox, never trust.