Evaluate and release

The evaluate and release focus area provides best practices for making a responsible release decision, given the release criteria from Release criteria, the evaluation datasets from Dataset planning, and the AI system design from System planning. The release decision is a binary (yes or no) decision that aggregates the results of testing the AI system against each individual release criterion. A well-designed decision factors in the compounded uncertainty in the individual criterion decisions, that is have we met each criterion with the required confidence (for example, are we at least 95% confident that the release criteria have been satisfied?). If a specific release criterion is not met, release may still be possible if the residual risk is disclosed through Monitoring transparency mechanisms. Otherwise, you should return to System planning to review options for baking and filtering mitigations.