Start with structural checks
Confirm that required files can be read, fields have documented types, and episode references resolve. Check array lengths and image availability against the declared schema. These checks identify broken delivery structure, but they do not establish that values represent the correct sensor or task. Retain the schema version and source information needed for deeper review.
Inspect changes over time
Look for unexplained discontinuities, constant fields, invalid values, and gaps in timestamps. Interpret them with source context rather than assuming every unusual value is an error. A stationary robot can legitimately produce little change, while a frozen measurement during visible motion requires investigation. Compare multiple related records when the project includes them, without inventing missing signals.
Relate records to task evidence
Review a complete episode with media, actions, and measured state side by side. Check whether the documented movement and gripper behavior are plausible given the observations. This can expose swapped streams or incorrect episode references. It remains a review of available evidence, not proof that a recorded command was executed exactly or that every physical state is directly observable.
Record issues and dispositions
Separate repairable export problems, source uncertainty, and task-level exclusions. Keep affected episode identifiers and the reason for each decision. Recheck changed records through the intended loader rather than validating only a replacement filename. Ask Fiund about source documentation and representative sample review before treating a broad dataset description as evidence of sensor integrity.
Review checklist
- Validate references and schema
- Inspect continuity with source context
- Compare records to observed behavior
- Track issues by episode and version