A team is working in Microsoft Foundry to test and compare large language model (LLM)...

The team cannot rely on live user traffic for evaluation because it would make results non-reproducible and could expose users to untested prompts. Synthetic data generation within Microsoft Foundry allows the team to produce diverse, representative input examples that simulate real user queries without any live traffic risk, giving a stable, consistent test set evaluated under identical conditions. Option B (content filters) and Option C (blocklists) are safety controls applied at inference time, not evaluation inputs. Option D (observability metrics) is a monitoring capability for production systems. With synthetic data, the team can construct edge cases, varied phrasings, and domain-specific scenarios that would take months to accumulate organically from real users, directly enabling the controlled evaluation the question requires without any dependency on live traffic.

Microsoft Learn Reference Topic: Evaluate generative AI apps with synthetic data – Microsoft Foundry prompt evaluation