Start with the temporal question
A clip showing an object on a table may support recognition while failing to show how it arrived there. Describe whether the model needs an instant, an action, a complete activity, or a longer session. The shortest convenient export is not automatically the shortest useful example. Give concrete examples of the reasoning your input must enable.
Define completeness operationally
Specify the beginning and ending conditions that make a sequence complete. A cooking activity might require setup and cleanup, while a tool-use task may need only the manipulation and visible outcome. Decide how pauses, retries, interruptions, and failed attempts are treated. Keep incomplete examples flagged rather than allowing their duration to imply that the required activity is fully present.
Plan clipping and source links
If short clips are derived from longer recordings, retain source identifiers and time offsets. Explain whether clips may overlap and how related clips are grouped during evaluation. Multiple segments from one session should not silently look like independent recordings. Preserve the unclipped source when the project requires later inspection of omitted context and the agreed source permissions allow that workflow.
Inspect full examples during review
A highlight reel cannot demonstrate continuity across ordinary and difficult moments. Request complete sample units, including transitions and quiet stretches. Check whether the proposed duration distribution supports the task rather than only an average length. Fiund can review archive and custom-collection options, but completeness, availability, sample access, and commercial scope are established for the particular proposed material.
Review checklist
- Define the model’s required temporal evidence.
- Write start and finish conditions for completeness.
- Keep source identifiers and clip offsets.
- Review full sample units rather than highlights.