Start HereArchitectural inferencev1.10.0
Why Usefulness Does Not Eliminate Composition Risk
Evidence levelArchitectural inference
A component can be useful, well-intentioned, and locally evaluated while still participating in unsafe composed behavior. Usefulness is not a safety proof.
Dual-use architecture property
The same router that selects the cheapest specialist can select a combination never tested directly. The same memory that improves personalization can preserve a retired behavior. The same evaluator that enables release discipline can become a monoculture if every gate shares one model family.
The operational question
The question is not “is modularity good or bad?” The question is “what exact combinations are permitted, evaluated, recorded, monitored, and rollback-capable?”