Define the state rather than the verb
An action label such as turning may not explain whether a switch changed position or a fastener became secure. Name the initial and final states relevant to your model. Decide which differences must be visibly distinguishable. This keeps the data request focused on evidence of change rather than relying on a broad description of hand movement. Describe intermediate states when the change happens gradually.
Keep sufficient surrounding context
A close crop of contact can omit the original state or the finished result. Review the sequence before and after manipulation and decide how much context must remain available. If an object is replaced off camera, an apparent state change may actually involve a different instance. Stable object references help reviewers detect this ambiguity instead of silently accepting it.
Distinguish visible results from assumptions
A participant can perform the expected gesture without achieving the intended outcome. State whether success comes from visible evidence, a separate task record, or another measurement. Keep those evidence types distinguishable. When the state is hidden, use an uncertain label or exclude the example from a task requiring direct visual confirmation.
Plan annotation around the research question
Some projects need a whole-clip change label; others need object localization or a temporal moment associated with the change. Request only the granularity your task can use and define how partial changes are treated. Include reversible actions and repeated attempts in the pilot so the annotation rules cover more than the cleanest demonstration.
Review checklist
- Name observable initial and final states
- Retain before-and-after context
- Track the same object instance
- Define partial and uncertain changes