LastSolve

LastSolve wraps any expensive black-box solver — PDE simulators, legacy code, subprocess binaries — learns its behavior from roughly ten strategic calls, and serves the remaining queries in microseconds with validated error bounds. Hence the name: the last few solves you’ll ever pay for.

A single @accelerate decorator gives you transparent surrogate caching. Under it: split-conformal error certification with finite-sample guarantees, parameter identification with Cramér–Rao bounds, bifurcation detection, and blind subspace auditing. The Φ₁ “dial” — Resona’s effective-rank measurement — flags when a cheap surrogate simply doesn’t exist, so the tool knows when not to trust itself.

Validated against 35 nonlinear PDE families (Burgers, KdV, chaos, solitons): ~570× faster queries at machine precision, saturating theoretical inversion bounds.

Pure NumPy + SciPy + Resona. No ML frameworks, no GPU, no training loops.

Source: github.com/dimaq12/lastsolve