The model should not carry the whole system. Keep artifacts, state, relations, validators, and review loops outside the learned part.
Cursed regions
Low-abstraction artifacts, scarce data, and unstable objectives need measurement before model selection.
Cognitive security
Adversarial information systems need evidence paths, confidence, context, and operator agency.
Graph boundaries
A graph is the relation layer. It works after object identity has been stabilized and validators keep the artifact honest.
Agent environments
Post-training depends on state, tools, traces, verifier failures, and feedback close to the action that caused the failure.
Physical AI
CAD and AEC workflows need exact artifacts, stable object state, typed relations, hard validators, learned residuals, and review.