RoBERTa sequence classifier (human vs. AI-generated) fine-tuned on academic prose. It is the default hf detector backend of the automated_humanizer pipeline, where it gates a humanize → detect → validate rewrite loop.

This detector was trained on the same data family as the project's humanizer, so its confidence is an internal, biased estimate — useful as an optimization signal inside the loop, not as independent evidence of authorship. The project README recommends an external detector (e.g. GPTZero) for final sign-off, and test_detectors.py in the repo reproduces why.

Check the human-class label index once per deployment with the project's calibrate_detector.py.