Set out the constraint honestly: disconnected environments cannot run frontier models, so the question is what is lost.
Identify the tasks where narrowness helps rather than hurts: triage, correlation, entity resolution, plan construction.
Cover the tasks where it does not, and say so plainly.
Describe the Lancaster research partnership and the grant funded work, at the level of detail the agreement permits.
Close on the practical result: what runs on what hardware, inside which boundary.