Pure-JAX linear and hyperplane decision trees (port of LLNL's hyperplanetree). Piecewise-linear surrogate modeling with first-class uncertainty quantification (linear-propagation, quadratic-difference, Mondrian leaf conformal), batched active learning (ask/tell + run sugar with campaign persistence), and MIP-friendly export (solver-neutral JSON, hand-built ONNX, optional discopt adapter for direct MINLP embedding without Pyomo).
2026-07-04