Formal verification techniques depend on architecture: piecewise linear networks can be recast as linear classifiers or solved as satisfiability or mixed-integer problems; binarized networks can be encoded exactly as Boolean formulas; recurrent networks can be approximated by a learned finite-state shadow model and model-checked, or treated by abstract interpretation; transformer self-attention layers are decomposed into sub-layers whose bounds are propagated. The organisation also accounts for the input data type, since methods are limited by scalability with input size and by how perturbations can be modelled: image variations can often be stated on pixel values or through masks, time series need per-instance handling, natural language variations may be hard to formalise, graph structure perturbations are discrete, and heterogeneous tabular data inherits all these limits.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.