An insurer's question-answering tool checks every answer with an Amazon Bedrock Guardrails contextual grounding check set to a threshold of 0.95. Reviewers find that many correct answers, fully supported by the policy documents, are being blocked, and the logs show they scored just under 0.95. What should the team do?