Papers

Work is filed by the question it answers. One study can sit in more than one paper

CB·1

Scope

Which jobs a compact model can do unaided, and which need a larger system or a person in the loop.

CB·2

Fine-tuning

Training data, adapters, and decoding: what changes behavior when compute is the limit.

CB·3

Evaluation

Measurement that stays out of the loop it measures, with uncertainty on every figure.

CB·4

On-device

Running capable models offline, where the data already lives.

CB·P1Fact-Faithful Editing with Ensemble Reranking: A Pre-Registered Study2026-10