Alibaba Cloud Community published a write-up on 3 November of integrating JEV into AgentScope. JEV, released by TypeSafe, belongs to a model class the company calls System One: instead of generating text, it receives a context and a typed question and returns options, scores and probabilities that program code can branch, rank and route on. The developer defines the question and its answer space in advance; the model judges the supplied material against that space. The post gives ticket routing, factual-grounding checks and risk grading as examples, and notes that generative models keep the work of analysis and content generation while JEV handles the discrete decisions between steps. Because the output is a probability, a workflow can express uncertainty as a branch: keep the current route, gather more evidence, or send the case to human review. The write-up states that calibration here means predicted probabilities match observed outcome frequencies, and warns explicitly that this does not make any single judgment correct, so thresholds still have to be validated on a team's own samples.
Our takeSplitting decision from generation gives the harness a typed contract with a number attached, which is what lets thresholds live in code instead of in prose. The caveat is calibration: matching frequencies on the data a model was built around says nothing about a new task's cutoff, so the number still needs a local labelled sample before it can gate anything. So what: the integration work is measurement work — define the question space, then measure where the probability actually separates the two cases you care about.