Worthify Journal

How we build and test AI.

Research and engineering notes from Worthify on AI training methods, evaluation results, and the design choices behind our work.

Featured article

How We Trained Decision-1

Decision-1 scores 2–16 supplied choices. See the training method, held-out results, and two government tasks to test: cybersecurity evidence review and document labeling.

Illustration of records entering a choice-scoring system, with a separate route to review
Decision-1 scores the choices you provide; your application handles review and action.