Limitations

• AI detectors can make mistakes, and newer generation methods may be difficult to detect.
• The backend can run configured Hugging Face image-classification models. No detector model IDs or labels are pre-verified by this repository; the operator must inspect and configure them.
• Model outputs are experimental evidence, not proof, and scores are not calibrated probabilities.
• Compression, resizing, editing, and screenshots can change image signals or remove metadata.
• Missing provenance does not prove that a file was AI-generated or that it is authentic.
• Video verdicts use configured frame classifiers sampled across the clip and deterministic score-distribution/neighbor-stability checks; this does not establish identity consistency or prove manipulation.
• Face detection/tracking, audio authenticity, audio/video synchronization, and deepfake-specific temporal models are not implemented.
• FFmpeg/ffprobe must be installed on the backend PATH. Codec support depends on that FFmpeg build, and a configured size/duration/resolution/FPS budget still applies.
• Optional language-model use is limited to explaining structured video evidence; it does not select video tools or change the deterministic verdict. No verified Nemotron model ID is supplied.
• Only checks recorded in the result trace should be considered completed.

Not yet evaluated. This checkout has no measured benchmark results or usable calibration system. Do not interpret raw detector scores as percentages, accuracy, or certainty.

The current provenance reader reports EXIF and C2PA availability only; it does not report a verified generator identity. The result page therefore does not identify a generator.