Distillation into a smaller model trained to reproduce the answers of a larger one on these standards.

Retrieval, because the knowledge changes often and updating a document updates the answers immediately.

Fine-tuning on the standards, so the model internalizes them and can answer without additional lookup.

Continued pre-training on the engineering corpus so the model develops genuine domain understanding.