Privacy

The image analysis endpoint reads the uploaded image into backend memory, writes a temporary file for analysis, and removes that temporary file in its cleanup step. Video uploads are staged in a backend temporary file; FFmpeg-extracted frames are also temporary. Both are deleted in cleanup paths. The server does not retain the uploaded video or extracted frames. The original video preview is held in browser memory during the current tab's result journey; refreshing or leaving the page may make that local preview unavailable.

The backend stores investigation results, including verdict, evidence summaries, trace, explanations, provider metadata, and a timestamp, in a local SQLite database. User reports store the result token, category, comment, and timestamp. No retention period or automatic deletion policy is configured in this checkout.

The insights endpoint returns aggregate analysis counts and verdict distribution. The backend contacts the configured model provider for a status check and may use it for tool selection or explanation when enabled. The optional VISION_MODEL tool can send the uploaded image itself to that configured provider if it is enabled and called. A configured language model may receive structured video metadata, sampled-frame scores, timeline timestamps, and the deterministic verdict to draft an explanation; it is not given the uploaded video by the video investigator. Configured Hugging Face classifiers run through the local Transformers pipeline; their weights may be downloaded from the model registry. Review provider terms and model sources before enabling them. No analytics or advertising tracker integration was found in the application source inspected for this release.

Provider credentials are read by the backend from environment configuration and are not intentionally exposed to browser code. Results communicate uncertainty and do not prove authorship or authenticity. Actual retention and access controls depend on the deployment.