Define the beginning operationally
A recording that starts after tool selection may be enough for one recognition task and incomplete for another. State whether the episode begins with the instruction, the initial scene, or the first physical action. Give reviewers a visible cue they can identify consistently, and retain any earlier context that the target model is expected to use.
Keep transitions meaningful
Walking between work areas, searching for an item, and pausing to inspect progress can explain how a task unfolds. Removing these intervals changes the prediction problem. Decide whether the dataset represents naturally paced work or isolated demonstrations, then describe allowed trimming. Preserve original timestamps whenever a shorter version is derived from a longer source session.
Require evidence of the outcome
The final hand movement does not always establish completion. A lid may appear closed while the relevant latch remains hidden, or a placed object may fall after the clip ends. Identify the visible result needed for the task and request sufficient follow-through. Allow an unknown result when the available view cannot support a success judgment.
Review interruptions explicitly
Tasks can be abandoned, resumed, or split by unrelated activity. Decide whether these become separate episodes, linked segments, or one interrupted sequence. Document reset events and repeated attempts so they do not masquerade as independent completions. An acceptance report should distinguish complete tasks, partial tasks, and technically unusable recordings rather than blending them into a single footage total.
Review checklist
- Define observable start and end cues
- State trimming and transition policy
- Show the final task result
- Label interruptions and repeated attempts