Writing evaluations for agent behaviour

An evaluation for agent behaviour is a versioned set of tasks, expected outcomes, refusal conditions, and scoring rules that constrain a non-human builder.