Screening User-Supplied Video for Sexual Content Before Processing: A Two-Classifier, Human-Reviewed Gate and 50 Days of Production Data

Services that take a user's video and transform it (clipping, captioning, reframing, dubbing) are expected by payment processors and by law to keep sexually explicit material out. Published evaluations of NSFW detectors mostly use curated still images. Few report what happens when such a detector runs on the frames of ordinary user video in production. We describe a pre-processing content gate in a short-form video clipping service. It screens every job at up to three points before any output exists. It scores sampled frames with an open-weights image classifier, counts a flag only when a second, independently trained classifier agrees, and sends anything short of a clear block to a human. It fails closed when frames cannot be read, and it stores scores and timestamps, never images. Over 50 days the gate ran 3,369 screens on 1,599 jobs and scored 88,822 frames. Setting aside our own explicit test video, 177 jobs drew a flag under at least one rule. Human review confirmed 2 as violations and cleared 128 as false positives, so about 1.5% of reviewed flags were correct. The second classifier overturned the primary's flag in 126 screens, with a clear margin, and none of the jobs it allowed has been found to be a violation. Recall on unflagged traffic was not measured. On this traffic the dominant problem is false positives on ordinary video, including children's animation, sport and music. A single high-scoring frame is noise. A platform's own signals (views, age restriction) are useful evidence but are not a substitute for review. We publish the design, the aggregate data, our mistakes, and recommendations for teams building similar systems. Aggregate data (four CSV files, no personal data) is included.