How do we check that evaluation answers have not leaked into training material?

**Are different review identifiers enough to separate training and evaluation?**

No. Two reviews can expose the same underlying account history or repeat the same decision rationale under different identifiers. A customer may effectively encounter the evaluation answer during training even when exact record matching finds nothing.

**What should we inspect?**

Check shared entities, copied narrative passages, repeated policy-specific answer patterns and references to solution files. Inspect retrieval indexes as well as the visible dataset. An evaluator-only rationale stored in a searchable document collection can undermine an otherwise careful split.

**Should every shared policy be removed?**

Not necessarily. If the task measures applying a known policy to new cases, sharing the policy is intentional. If it measures adaptation to unfamiliar rules, the policy family may also need separation. State which capability the split is intended to test.

**What changes when we combine releases?**

Resellers should reassess the combined corpus rather than relying solely on each component's original split labels. A case held out in one product may closely match training material in another. Ask for grouping metadata that supports checks across the products customers are likely to use together.