An AI recommendation to approve a loan can quote the right figures and still fail the lending policy. In our synthetic demo, a cached model response recommends approval even though the credit score is below the configured minimum. Its supplied evidence matches the record. The Validation Firewall blocks the recommendation because it fails the encoded policy. This replay makes no fresh model call. The demo covers selected checks, and routing is simulated. For a lending review, ask two separate questions: do the cited figures match the record, and does the recommendation follow the policy? Passing the first check cannot settle the second. #CreditRisk #ModelRiskManagement #LendingPolicy #AIGovernance