Candidate formal methods are classified on four axes before one is chosen. Complete verifiers give an exact answer, a proof or a concrete counterexample, but struggle with high-accuracy networks on hard datasets; incomplete verifiers use abstraction to scale but can fail to prove a property that holds. Deterministic verifiers prove a property for every input in the region, whereas networks that output a distribution need parameters of that distribution bounded or probabilistic guarantees. Glass-box verifiers need the architecture and learned parameters, not the training data; where the model is inaccessible, closed-box verifiers that only run the model are used, at some loss of precision. Most verifiers assume ideal real arithmetic, so their guarantees do not transfer to floating-point execution unless a verifier sound with respect to the computer arithmetic is used.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.