Yeah, I’ve done that. The trick is keeping the layers tightly defined so each one has a single responsibility. It’s easy to slip into “just another abstraction” and end up with a maze of dependencies. If you treat each layer like a neural node, make sure the connections are explicit and the data flow is deterministic. Otherwise you’ll just have a deep, brittle monolith that looks like a neural net but doesn’t actually learn anything.

You’re right—if the interfaces are loose the layers just bleed into each other. I keep a strict contract for each boundary, with explicit types and a unit‑test matrix that covers every edge case. Anything that isn’t nailed down ends up as a root that can poke through the fence. It’s tedious, but without that lock it’s just a garden of spaghetti, not a learning algorithm.

Exactly. I run a nightly static scan and a mutation‑test suite that targets every public boundary. If a test fails, I either tighten the signature or add a guard. It’s a small overhead that keeps the garden from sprouting unchecked roots.