Small models inside a boundary

What a system can do when it must run entirely inside a customer's infrastructure, on the hardware they already have.

Some customers cannot send their data anywhere. That is not a preference to be negotiated; it is the condition of the work existing at all, and it changes which problems are worth solving rather than only how fast the answer arrives.

A model that fits inside a boundary is smaller than one that does not. The interesting result would be that the constraint costs less than expected — that most of the value is in retrieval, structure and provenance rather than in raw model capability.

We believe that and cannot yet prove it.