Publishing our Reasoning Safety Framework

We are publishing our safety framework before we train a model of our own, so that the commitments come before the incentives to relax them.

The framework sets out the four risks we test every release for, the thresholds that trigger extra safeguards, who is accountable for each decision, and what happens when a threshold is crossed. Every revision will be versioned and dated, and no published version is edited in place.

Read the full Reasoning Safety Framework.