How can resellers avoid evaluation leakage when distributing related variants?

**Is changing record identifiers enough to create an independent evaluation case?**

No. Two cases can use different identifiers and still share the same message structure, dependency pattern and decision path. If one appears in training and the other in evaluation, a model may recognise the template rather than solve the workflow.

**What should a reseller inspect?**

Ask for case-family identifiers and a description of how variants are derived. Useful grouping dimensions include the underlying dependency graph, policy branch, exception mechanism and source-message template. These make structural overlap visible even when surface wording differs.

**How should partitions be supplied?**

A release can include recommended grouped splits and a separate manifest describing their construction. Resellers combining releases should check whether families recur across products and preserve the grouping metadata through repackaging.

An evaluation sample should also contain combinations that require the intended reasoning, not merely familiar phrasing. Downstream buyers can compare results on familiar structures and held-out structures, with the distinction clearly documented. This does not establish universal generalisation, but it helps prevent a catalogue's apparent breadth from disguising repeated versions of the same underlying task.