Linda-Pro 1.0: a local ensemble detector of AI-generated English text

Technical report on Linda-Pro 1.1: a local ensemble detector of AI-generated English text (stylometry + two fine-tuned DeBERTa models) retrained on real outputs of commercial AI humanizers. On HumanizerBench (not used for training) the share of humanized texts flagged rose from 31% to 70%; on a held-out October 2026 benchmark cycle from 47% to 82%, with human-text false positives not higher overall (TOEFL essays 8.8% to 4.4%) and a small rise on school ELL essays (0.7% to 1.1%). Caveats and limitations (older small-model texts, monthly changing humanizers, not an independent audit) are reported in full. Version 1.0 report: https://doi.org/10.5281/zenodo.23072494 . Code and evaluation script: https://github.com/Lendarixon/Linda . Self-published, not peer reviewed.