Look beyond identical files
The same interaction can appear in a full session, a trimmed clip, a stabilized copy, and a preview. These files may have different encodings and names while preserving the same event. Ask for source relationships and transformation records. Exact file checks are useful for integrity, but they do not identify every semantically overlapping example. Include thumbnails and extracted image sequences in the source inventory.
Choose the grouping unit
Decide whether training and evaluation must be separated by session, participant, location, or task setup. The correct grouping follows the generalization question. Neighboring segments from one activity may share objects and context even without overlapping frames. Document the grouping rule before sampling so related examples do not enter different partitions through later processing.
Review near repeats intentionally
Repeated demonstrations can be valuable for studying variation, but they should remain distinguishable from accidental duplication. Record attempts and resets, and check whether apparent variety comes mostly from different crops of the same source. If repeated setup conditions are intentional, report them as repeated conditions rather than implying every clip offers a new environment.
Maintain the split as data changes
New annotations, extra clips, and revised files can reintroduce source overlap after the original split is complete. Preserve partition assignments at the source level and check additions against them. Keep a versioned manifest that explains grouping decisions. Evaluation independence should be supported by the collection’s lineage, not inferred from a folder labeled test.
Review checklist
- Link transformed versions to sources
- Choose the intended grouping unit
- Distinguish repeated attempts from duplicates
- Recheck additions against the split