The release introduces Jev through the TypeSafe client and defines three experimental decision types: Noul for Boolean decisions, Choice for selecting among described alternatives, and Score for ordered rubrics. The release notes say these types return a decision together with probability or confidence-related evidence. ReAnchor Calibrates Decisions Against Your Metric. Its stated job is to fit Boolean thresholds, Score cuts, and Choice weights in Predict programs without rewriting instructions, criteria, or demonstrations.
The mechanism is local decision policy layered over a DSPy predictor. Thresholds, score cuts, and choice weights are applied locally, which the release says allows identical backend requests to reuse cached evidence. This matters now because DSPy calls version 3.4 its LM transition release and names 3.5 as the migration deadline. The implication is that teams can separate an underlying model’s evidence from the operational rule that turns that evidence into routing or escalation behavior. Evidence boundary: the release notes describe experimental APIs and do not establish that these outputs are statistically calibrated in every setting. Watch whether ReAnchor improves held-out operational metrics across real decision workflows, rather than only fitting thresholds on a development dataset.