Identify the origin of each text
An instruction issued before a task, a participant’s spoken narration, and a reviewer’s later description carry different information. Preserve that distinction in metadata. An instruction says what someone was asked to do; it does not prove that the action occurred. A retrospective description may use context unavailable to a model at an earlier point in the sequence.
Choose the level of detail
A session summary can describe the overall activity while omitting individual contacts and transitions. Fine action narration can provide more temporal detail but may introduce ambiguity around repeated objects or overlapping behavior. Define the intended text unit and show examples. Specify whether descriptions should refer to object categories, stable instances, or scene-relative phrases such as the item on the left.
Review grounding against footage
Ask reviewers to locate the visual or audio evidence for each statement. Mark uncertain, inferred, and unobservable details rather than rewriting them as facts. If text describes a hidden intention or an outcome established elsewhere, identify the evidence source. Fluency alone is not an acceptance criterion for descriptions intended to supervise or evaluate a perception model.
Prevent unintended answer exposure
Some text can reveal evaluation answers or future events. Decide which fields models receive during training and testing, and keep answer annotations distinct from inputs. Preserve time references and source links so derived prompts can be audited. When sending a brief to Fiund, include example instructions, narrations, and expected outputs instead of requesting an unspecified text layer.
Review checklist
- Record how each text was created
- Define text units and references
- Review statements against evidence
- Separate model inputs from answer fields