A closed-loop architecture for building domain models on the high-fidelity data your operation already discards. Capture and verification precede training; training is the last, smallest move. On-policy distillation, the synthetic-data question, reverse-KL, early-trajectory selection, and the as-built closeout as both reward and eval — a field note and whitepaper read against a real Division 9 takeoff.

A ten-frame visual field map of the Turbo Framework — the closed loop, the exhaust manifold, the verifier valve, the wastegate, and the twin-scroll turbine, each mechanism drawn against a real Division 9 takeoff. A swipeable deck made to teach the architecture, not just state it.

A field method for putting capable models inside the building: distilling a domain expert's judgment into a structured knowledge graph, adapting a reasoning-capable base with per-vertical LoRA, and serving it quantized on Apple Silicon. The hard part isn't the training loop.

An eleven-frame visual companion to the field note — distillation, LoRA, on-policy training, quantization, and evaluation, each mechanism drawn against a real Division 9 takeoff. A swipeable deck made to teach the method.