To play devils advocate, isn’t
any security approach fundamentally statistical because we exist in the real world, not the abstract world of security models, programming language specifications, and abstract machines? There’s always going to be a chance of a compiler bug, a runtime error, a programmer error, a security flaw in a processor, whatever.
Now, personally I’d still rather take the approach that at least attempts to get that probability to zero through deterministic methods than leave it up to model alignment. But it’s also not completely unthinkable to me that we eventually reach a place where the probability of a misaligned model is sufficiently low to be comparable to the probability of an error occurring in your security model.