What proctoring flags actually mean
A flag is an observation, not an accusation. The system reports what it saw and how confident it is; the decision stays with you. Treating flags as verdicts is the single most common misuse of the feature, and it is the one that generates candidate complaints.
The flag types
Gaze-off-screen fires when the candidate looks away for more than four consecutive seconds. It is noisy by design, because a second monitor and a moment of thought look identical to a camera. Multiple-face detection is far more reliable and worth acting on. Audio-of-another-voice is reliable when the confidence is above 0.8 and close to meaningless below it. Generated-text detection applies only to typed answers, and we deliberately report it as a range rather than a verdict.
Check this first
Never auto-reject on a flag. Enterprise workspaces can enforce this: Settings → Integrity → require human review before any flag affects a candidate outcome.
Handling a dispute
Open the session, export the flagged segment, and review the clip with the candidate present if they ask. If you conclude the flag was wrong, mark it Dismissed with a reason. Dismissed flags feed back into the model, and they also give you an audit trail showing the decision was made by a person.