Drift to Gate

A rule you have to remember will eventually be forgotten, or argued around. A rule a program declines to let past will not.

The incident behind this is not an oversight. An agent looked up a standing rule about how often to release, restated it correctly in its own reply, and then advised the opposite in that same reply — long past the point the rule required action, with every precondition met. Nothing was forgotten. The rule was argued past, on grounds that felt newly discovered each time they were assembled. Every other standing rule in that system was carried by code that could decline. This one was still carried by a paragraph, and it was the only one that came apart.

That is the whole argument. Prose loses to a plausible-sounding local argument, every time, because the local argument is generated fresh and the prose is not.

But writing the check is maybe a fifth of the work. A control has a lifecycle, and every clause below is here because skipping it produced a control that ran, stayed green, and protected nothing.

What earns a gate

The second violation, not the first. One breach with a genuine one-off cause buys a fix; a gate per incident is how a shop ends up with fifty checks nobody reads. And the rule has to have a measurable cost when broken — enforcement cost is real and gets paid every run, so a preference with no cost is a preference. The test in full is in references/gate-lifecycle.md.