Pangram 4 reports a 0.0041% false positive rate on AI-text detection.

The July 29 technical report claims AUROC 0.9916 with a 0.0041% false positive rate and 0.3396% false negative rate, plus better out-of-distribution generalization and adversarial robustness than Pangram 3, with fine-grained discrimination of edits and mixed AI-human co-writing rather than binary authorship. The FPR is the number to scrutinize, because it's the figure that determines whether detection can be used punitively at scale. Two orders of magnitude below typical classifier claims deserves independent replication before anyone builds policy on it. It arrives the same week a NeurIPS reviewer reported receiving entirely LLM-generated rebuttals.