Knowledge.
Memory.
Evidence.

People should be able to choose the knowledge their model works with. We build compact models that take in the documents a person chooses and keep that knowledge available as they reason.

Our central questions are retrieval, attribution and separation: can the model recover a fact, associate it with the right entity, and keep one source distinct from another?

A plausible answer is only a starting point. We investigate ways to recover supporting facts, check an answer against them, and recognise when the available information is insufficient.

We evaluate exact facts, confusingly similar entities, questions that require several pieces of evidence, and questions with no answer in the source.

We study training methods that make better use of a limited compute budget, so that capable models can be trained and run without large infrastructure.

Promising results on small models are hypotheses to test at a larger scale. We compare matched runs and track quality, time and memory together.