The IMEX reaction-diffusion loop moves out of TypeScript and into .m files under models/, one per model. numbl parses and lowers them in the browser; each element-wise line becomes a WebGPU compute kernel. The MATLAB is editable on the page, so changing the reaction changes the pattern live. How it fits together (src/mgpu/): - compile.ts specializes init() and step() for the current grid's argument types, via the same specializeUserFunction entry point numbl's own JIT uses, then runs numbl's inline pass. Lowering emits one statement per operator (ANF); the inline pass folds those back into per-line expression trees, so `uuv = u .* u .* v` is one kernel rather than three. - wgsl.ts emits one kernel per element-wise statement, the WebGPU counterpart of numbl's C-side fused emitter. Anything it cannot express is refused at compile time with a position in the file. - plan.ts builds every pipeline, buffer and bind group once. Because numbl fixes types and shapes at lowering time, the op sequence is static, so a timestep is pure synchronous command recording: one submit per batch, and the only await in the loop is the single readback per rendered frame. - externals.ts gives numbl the type rules for synth/analys through a .mtoc2.js workspace file, its sanctioned JS-builtin extension point; the backend maps those calls onto the existing transform pipelines. Argument binding goes through the MATLAB signatures, so a model file declares what it needs and an unknown parameter name is a compile error rather than a silently undefined variable. Tunable scalars are lowered without exact values so they live in a uniform buffer: editing the MATLAB recompiles, moving a slider does not. src/solver/ stays as an independent implementation of the same scheme and becomes the test oracle. test/mgpuChecks.ts runs both from the same seeded perturbation through the same fp32 transforms: the three models agree to 5.5e-8 - 3.2e-7 relative L2 after ten steps, and match exactly at init. It also asserts the per-model kernel count, which catches fusion silently not happening - a regression that leaves results correct but makes every operator its own dispatch. WebGPU is now required; there is no CPU fallback in the app. The editor gets MATLAB syntax highlighting, by overlaying a highlighted <pre> behind a transparent textarea.