A model that does not write. You hand it the current state and a fixed list of options, it hands back one of them with a probability. That is the whole thing. I have it picking animal behavior in a game a few hundred times a minute, which was never affordable with a chat model.
It's just a normal LLM where instead of running it with an autoregressive sampler, you take the residual stream before the LM head and do some kind of projection onto a smaller space. Not sure *exactly* how the jev people do it but structured generation techniques aren't new.
Jev is a groundbreaking, state-of-the-art deep-learning parity classifier that, after billions of parameters and heroic amounts of compute, can tell you with 99% certainty whether a number is even. Truly a paradigm shift in the field of “looking at the last bit.”