Questioning Jev's zero-hallucination claim: model-level or sampler-level guarantees
Developers discuss how Jev differs from Outlines and llguidance, asking whether its type safety is intrinsic to the model or enforced by the inference layer, and call for technical details.
Yeah, I think that is a valid point. if 0% hallucination means "0% schema violations", then calling that zero hallucination is pretty misleading. Where I'd separate Jev from Outlines/llguidance is that Jev isn't generating text and then constraining the decoder. The difference here is that Jev's claimed primitive is a typed probabilistic decision, rather than generated text. So I don't think the interesting question is whether it can produce invalid schema (we already know how to guarantee schema-conformant outputs from autoregressive LLMs with constrained decoding) From where I'm standing the more interesting question is from where Jev's guarantee actually comes from: is the typed output space intrinsic to the model/output architecture, or enforced by its sampler/inference layer? And beyond type safety, the bigger claim to test is RLCD itself: does Jev produce meaningfully better calibrated semantic decisions than a comparable LLM using constrained decoding and explicit calibration? I would love they publish technical details so we can actually see what is novel at the model level