| --- |
| language: |
| - en |
| - ru |
| license: mit |
| tags: |
| - ai-text-detection |
| - reproducibility |
| - bilingual |
| - adversarial-robustness |
| - calibration |
| --- |
| |
| # ContentOS — Reproducible Bilingual AI-Text-Detection Ensemble |
|
|
| **Pre-print v1.0 (2026-04-27)** |
|
|
| This repository contains the open pre-print and supporting artifacts for |
| ContentOS, a reproducible English+Russian AI-text-detection ensemble. |
|
|
| ## Authors and affiliation |
|
|
| - **Gregory Shevchenko** — author, Humanswith.ai (founder) |
| - **Humanswith.ai team** — methodology, calibration, evaluation infrastructure |
|
|
| ContentOS is a Humanswith.ai product. This preprint is published under |
| the author's personal HuggingFace account; the supporting code repository |
| is maintained under the organization account (see "Code repository" below). |
|
|
| - Author profile: https://huggingface.co/gshevchenko |
| - Organization: https://humanswith.ai |
| - Contact for collaboration: open a Discussion on this dataset |
|
|
| ## Code repository |
|
|
| Public benchmark + evaluation scripts: |
|
|
| **https://github.com/humanswith-ai/contentos-benchmark** |
|
|
| The repo includes regression test suite (8 pinned baselines, 0.05s), |
| streaming-CSV eval scripts (partial-tolerant), per-genre AUROC |
| analyzer, and the calibration JSON shape for v1.11 production state. |
|
|
| ## Headline numbers (v1.11 production, 2026-04-29 measurement) |
|
|
| | Metric | EN | RU | |
| |---|---|---| |
| | OOD AUROC (176-sample expanded smoke) | **0.864** | **0.846** | |
| | Wrong-rate | 4% | 9% | |
| | p50 latency (EN ensemble) | **1.2 s** | — | |
| | Adversarial AUROC (n=300, OOD-paired) | **0.998** | — | |
|
|
| Earlier v1.0 paper reported 0.802 / 0.847 on the original 44-text |
| smoke battery; the 4× expanded battery with class balance per |
| (lang, genre) cell stabilized numbers upward. Per-genre details in |
| the [companion repo](https://github.com/humanswith-ai/contentos-benchmark). |
|
|
| ## Files |
|
|
| - `paper.pdf` — full pre-print (~6,000 words, 9 sections + 5 appendices) |
| - `paper.html` — self-contained HTML version with embedded figures |
| - `paper.md` — source markdown |
| - `figures/` — 4 figures (PNG + SVG) |
| - `REPRODUCIBILITY.md` — open methodology, how to reproduce in 90 minutes |
|
|
| ## Reproducibility |
|
|
| The full methodology and calibration corpus description are documented in |
| `REPRODUCIBILITY.md`, which is sufficient for independent re-implementation |
| of the ensemble. |
|
|
| A public mirror with the evaluation scripts (`eval_ensemble_corpus.py`, |
| 8 pinned regression tests, atomic-swap deploy with 30-second rollback) |
| will be released within ~2 weeks following the v1.12 RU recalibration |
| chain. Target reproduction infrastructure: Hetzner CX43 (8 vCPU, no GPU, |
| ~€14/month) or equivalent. |
|
|
| For early access before the public mirror, please open a discussion on |
| this dataset. |
|
|
| ## Cite as |
|
|
| ```bibtex |
| @misc{contentos2026, |
| title={ContentOS: A Reproducible Bilingual AI-Text-Detection Ensemble with Adversarial Robustness Evaluation}, |
| author={Humanswith.ai team}, |
| year={2026}, |
| url={https://huggingface.co/datasets/gshevchenko/contentos-preprint}, |
| } |
| ``` |
|
|
| ## License |
|
|
| MIT for code, methodology, and corpus aggregation. Underlying data sources retain their original licenses (HC3, AINL-Eval-2025, ai-text-detection-pile). |
|
|
