I keep an architecture document for a platform I've been building for months. Not diagrams. Just decisions in plain English. "This system fulfills orders, it never creates them." "External integrations log and continue, they never block the core workflow."

I keep it because I build with a coding agent, and the agent's failure mode isn't writing bad code. It's writing locally reasonable code that violates a decision I made three weeks ago. The diff looks fine. Review passes. Nothing catches it.

So I wanted to know how much of that document a tool could actually check. I built one to find out. Then I ran it against the real codebase and it caught me lying about my own numbers four times, which turned out to be the most useful thing it did.

Sorting the document

Before writing any code I went through the document statement by statement. 24 of them. They fell into three groups that need completely different handling.