One Clip, Several Readings: Telling How Firm an AI Deepfake Verdict Is, Using the Platform's Own Re-encoding Settings

A deepfake detector gives one verdict for one video or voice clip, with no sign of how firm that verdict is. Reports from the field describe detectors that score well in the laboratory but fail on clips that have been compressed and re-encoded on their way to the viewer. This note describes a check that sits on top of any detector. From the clip already stored, the computer makes two to four versions by moving only the platform's own re-encoding settings, inside the range the platform permits. The same detector, unchanged, reads every version. One number, the spread value, is the biggest disagreement among the scores. The spread is inversely proportional to the size of the system's leading eigenvalue, so a wide spread means the verdict is near a tipping point. The spread value places the clip in one of three calm states (Steady, Watch, Look closer). A computer-decided shunting action then moves the doubtful clips to the front of the human reviewers' queue while every other clip carries on. Because only one detector and one clip are used, the spread can come from nothing but the processing settings. The note compares the method with published papers, a patent and government-funded detector projects, and explains how it differs from each. Keywords: deepfake detection, verdict firmness, re-encoding settings, spread value, shunting action, early-warning signals