Limiting AI system claims to what has actually been tested and evaluated, avoiding overstatement of capabilities.