This paper addresses text-transformation tasks that are conceptually simple to describe but hard to implement with traditional rule-based code, while avoiding the repeated cost and latency of querying a large remote model at inference time. The approach compiles a natural-language specification by having teacher models generate task-specific training examples, which are then used to train a compact local neural adapter that becomes a reusable function that can be stored, versioned, and composed like ordinary software. On a challenging benchmark subset, the method achieves 83.6% semantic accuracy against a faster competing compiler, at the cost of roughly one minute of additional compilation time.
