For a network that controls an AI agent in an environment, a reachability property states whether a set of states can be reached over multiple steps: a set of failure states the agent must avoid or goal states it must reach. Expressing it needs a model of the environment giving the consequence of each action for the following state, deterministic or stochastic. The reachability criterion applies the property from a chosen collection of starting states; in a deterministic setting it is checked with the methods of 6.2.4, for a stochastic one it yields a probability found with 6.2.5. The criterion should be satisfied for the stated set of initial states, which may be specified or discovered by formal methods, and it lets high-level safety properties be expressed beyond input-output behaviour.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.