Work
All are in development.
We are fine-tuning and evaluating open models for UK law, public administration, regulated professions and everyday language. The work focuses on distinctions that general models often blur: jurisdiction, institutional roles, formal guidance, local terminology and professional conventions.
Evaluation will use real task types: explaining a rule without importing US assumptions, handling official documents, identifying when guidance is out of date, and knowing when to defer.
For some organisations, sending prompts or documents to a US-hosted service is not acceptable. We are designing deployment options that keep processing and storage under agreed UK control.
Each arrangement must state data location, retention, access, logging, model changes, incident handling and exit. Those are requirements, not supplementary paperwork.
British speech models are often generic or theatrical. We are researching models that represent regional accents, vocabulary, rhythm and social context more accurately.
The work includes consent, provenance and evaluation by speakers from the communities represented. A convincing voice is not enough if the method is careless.
Publication
We will publish methods, evaluations and limitations when there are results worth examining.