Retrieving architecture decisions is not the same as knowing they apply

Theo Valmis argues that AI coding agents need to know which architecture decisions have authority over the artifact being changed, not just which ones are relevant.

In a controlled repository, a retrieved decision banning PostgreSQL and ORMs was still ignored because nothing signaled that it governed the storage module.

He proposes separating applicability from retrieval and keeping deterministic enforcement for decisions precise enough to check, with the rest treated as guidance.