Company
Policies & Governance
The principles that guide our research and reporting.
Real measurements
Report measured results rather than invented performance claims. Compare systems on the same task set, with the same evaluator, data split, and conditions. Published measurements should disclose the relevant task, evaluator, models, data split, and date.
Literal status
Distinguish hypotheses, active research, and measured results. Use status language such as “in development”, “planned”, and “released” literally. Do not imply model, dataset, or product releases that have not happened.
Trustworthy rewards
Use reinforcement learning only where a trustworthy, verifiable reward exists. Evaluation and failure analysis guide the next data and post-training step.
