PEARL: a 4B open model beats a 685B giant at turning plain English into working optimization code

By letting the model run its own solver, see the errors, and fix them in a loop, a 4B model outperforms the 685B DeepSeek-V3.2 at optimization modeling.