README.md

---
license: apache-2.0
task_categories:
- text-classification
language:
- en
- de
- es
- fr
- ru
- zh
- ja
- ar
- hi
- bn
tags:
- ai-text-detection
- llm-attribution
- cloze-congruence
- multilingual
- model-provenance
- synthetic-text
size_categories:
- 1K<n<10K
---

# ClozeCongruence 3.0: Multilingual AI Text Forensics & Provenance Benchmark

Official reproducible benchmark datasets accompanying the research paper:
**"ClozeCongruence 3.0: Cross-Lingual Macro-Discourse Reconstruction and Dynamic Burstiness Gating for Zero-Shot AI Text Forensics across 10 Typologically Diverse Languages"** (Debdip Bandyopadhyay, 2026).

## Benchmark Splits

This benchmark spans 10 global languages across 6 typological writing scripts:
- **Latin:** English (en), German (de), Spanish (es), French (fr)
- **Cyrillic:** Russian (ru)
- **CJK:** Simplified Chinese (zh), Japanese (ja)
- **Arabic:** Modern Standard Arabic (ar)
- **Devanagari / Indic:** Hindi (hi), Bengali (bn)

### Included Datasets

1. `data/multilingual_longform_benchmark.json`:
   - 70 scholarly documents (42-45 paragraphs each; 3,120 paragraphs, 138,008 words).
   - Evaluated across 7 author classes: Human Academic, OpenAI GPT-4o, Anthropic Claude 3.5 Sonnet, Google Gemini 3.7 Flash, DeepSeek-R1, Alibaba Qwen 2.5 72B, and Meta Llama 3.3 70B.

2. `data/real_world_downloaded_human_benchmark.json`:
   - 10 authentic academic treatises downloaded from Wikipedia across all 10 target languages (447 paragraphs, 19,161 words) establishing real-world 0.00% False Positive Rate (FPR).

3. `data/multilingual_mixed_human_ai_benchmark.json`:
   - 30 long-form documents across 3 realistic human-AI collaboration regimes (50/50 interleaved, 25/75, and 75/25).

## Key Empirical Metrics

- **Mean 10-Language AUROC:** 98.92%
- **Native Human False Positive Rate (FPR):** 0.00% (Zero false accusations)
- **Top-1 LLM Attribution Accuracy:** 99.71%
- **Cryptographic Provenance:** ISO/IEC 27037 Tamper-Evident SHA-256 and Ed25519 digital certificates.

## Citation

```bibtex
@article{bandyopadhyay2026cloze3,
  title={ClozeCongruence 3.0: Cross-Lingual Macro-Discourse Reconstruction and Dynamic Burstiness Gating for Zero-Shot AI Text Forensics across 10 Typologically Diverse Languages},
  author={Bandyopadhyay, Debdip},
  journal={Transactions on Machine Learning Research},
  year={2026}
}
```