A security analytics pilot uses machine learning to block accounts automatically. Which control best addresses unreliable model decisions?

1. Remove all event records to avoid training bias
2. Evaluate only inference speed before enabling enforcement
3. Grant the model unrestricted authority because it learns continuously
4. Validate outputs and define monitored escalation and override procedures