Module C: Trained classifier
Scores writing style (sentence length, vocabulary range, punctuation, common function words) with a model trained on labelled human and AI answers, and reports a calibrated interval instead of a bare percentage.
How it works and its limits
A logistic regression over stylometric features, trained on HC3. Unlike Module B, it learned from real labelled data instead of just measuring predictability. Its confidence interval is conformal-calibrated. Design notes.
It was trained on one dataset and generalizes poorly past it: on the Attack Lab's own test set it labels only 7 of 16 texts correctly. Read the limitations.
or try
Results appear here: a verdict with its calibrated interval, and which style features pushed this text toward human or AI.