ProgramAsWeights compiles specs into local neural programs

Core idea is packaging task-specific behavior into compact program weights instead of shipping a prompted full model. The browser-local path is the most compelling practical angle because it makes local inference real rather than aspirational. The benchmark gap looks strong, but it is still a project claim and the real test is whether it generalizes beyond fuzzy-function benchmarks. The main tradeoff is compile-time dependence on a hosted 4B compiler, which favors consumers of compiled programs over users who want to train locally.