From the
build log.

You don't need a 70B model to know "shall" means mandatory. Decompose classifies the MCP spec in 3.78ms with zero API calls. Here's the full output.

LLMs are good at reasoning. They're bad at reading. Decompose gives them structure before they start thinking — and that changes everything.

It's not a buzzword. It's a design pattern. Causal consistency, Merkle trees, attention budgets, and reality anchors — explained with real code.