This started as a document, and documents get ignored.

The rulebook said things like: a feature is done when a command exits zero and every row of the evidence matrix is filled in, not when someone says it is done. It named the gates. It said which ones a person had to close. Written down, in one file, where anyone could read it.

The problem with a rulebook is that following it is voluntary. On a slow week everything gets checked. On the week before a deadline, with three agent sessions running at once, the screenshot gate becomes a glance and the architecture check becomes a feeling. Nobody decides to skip it. It just quietly stops happening, and you find out later.

So we built the thing that makes it non-optional. Same gates, same evidence, except now a task cannot move while a gate is red and the command line has no approve command for an agent to reach for. Quality stopped depending on who remembered.

We run it on our own work first. That is not a marketing line about dogfooding, it is the only honest way to find out which gates are load-bearing and which ones just look responsible on a slide.