Releases, in the open.

The first stable release of the local-first reasoning agent: evidence intake from files and the web, a calibrated confidence meter on every answer, and belief updates that show their work.

An open-weights reasoning model fine-tuned for uncertainty-aware answers. Reports calibrated probabilities instead of confident guesses, and abstains when the evidence is thin.

A toolkit for turning heterogeneous sources — PDFs, tables, transcripts — into structured evidence a reasoning agent can weigh and cite.

A study of how reasoning agents keep their confidence honest when the world moves away from their training data — with open benchmarks and reproducible code.