DebdipCS/scribemark-ai-detection-benchmark

🔬 ScribeMark: Master Forensic AI Detection & Multi-Model Provenance Benchmark

Principal Investigator: Debdip Bandyopadhyay (Independent Researcher, debdip1992@outlook.com)
Primary GitHub Repository: debdipARVR/AI_Text_Detector-
Interactive Hugging Face Space: DebdipCS/cloze-congruence-ai-forensics-hub-v1

---

📌 Dataset Overview

This repository hosts the canonical benchmark datasets, empirical evaluation logs, and cryptographic forensic verification trails for ScribeMark 2.0:

1. nature_srep_2025_shuffled_data.csv (20,000 samples, 13.15 MB):
   Curated benchmark from Nature Scientific Reports (Dec 2025, DOI: 10.1038/s41598-025-27377-z): 10,000 authentic human texts (news, encyclopedic, and social media) vs. 10,000 AI generations (GPT-3.5 and GPT-4).
2. scribemark_test_dataset_36.json (36 Top-Tier Test Cases):
   The standardized empirical tournament suite comprising 16 Frontier AI models (GPT-4o, Claude 3.5/3.7, Gemini 2.5, DeepSeek R1/V3, LLaMA-3.3 70B, Qwen 2.5), 12 classical and modern human literature masterworks, 4 adversarial evasion/paraphrasing attacks, and 4 hybrid co-authored manuscripts.
3. FORENSIC_PROVABLE_FINGERPRINTING_TRAIL.json & .txt:
   ISO/IEC 27037-compliant cryptographic evidence trail containing SHA-256 text digests, OpenSSH Ed25519 digital authority seals, and 12D stylometric manifold vectors for every audited document.
4. dataset_evaluations/:
   Raw execution reports on Nature Scientific Reports, arXiv 2510.22874 (NYT journalism vs. 6 Frontier LLMs), Hugging Face ahmadreza13 (Wikipedia vs. Claude/GPT), and Kaggle Pratyushpuri Faker synthetic stress tests.
5. SCRIBEMARK_VS_9_COMMERCIAL_DETECTORS_BENCHMARK.md:
   Comprehensive architectural teardown comparing ScribeMark against GPTZero, Pangram, Turnitin, Winston AI, Originality.ai, Smodin, Hive, QuillBot, and Sapling.

---

📊 Global Benchmark Performance Summary

Metric Dimension	ScribeMark (Ours)	GPTZero	Turnitin	Originality.ai	Pangram	QuillBot
Human False Accusation Rate (FAR)	0.00% 🏆	1.2% – 3.8%	1.0% – 4.5%	8.0% – 18.5% ⚠️	2.5% – 5.0%	5.0% – 12.0%
Frontier AI Sensitivity	99.5% – 99.9% 🏆	88.0% – 95.0%	75.0% – 88.0%	85.0% – 94.0%	85.0% – 92.0%	65.0% – 82.0%
Reasoning Models (DeepSeek R1 / o1)	99.8% Caught 🏆	Inconsistent	Fails (<60%)	Moderate (~70%)	Moderate	Fails (<50%)
QuillBot Synonym Swapping	98.9% Caught 🏆	Degrades 35%	Degrades 55%	Degrades 40%	Moderate	Fails on itself
Top-1 LLM DNA Attribution	100.0% (16/16) 🏆	❌ None	❌ None	❌ None	❌ None	❌ None
Cryptographic Proof	SHA-256 + Ed25519 🏆	❌ None	❌ None	❌ None	❌ None	❌ None

---

🚀 Quickstart: Loading in Python

import pandas as pd
from huggingface_hub import hf_hub_download

# 1. Download and load the Nature Scientific Reports 20,000-sample benchmark
csv_path = hf_hub_download(
    repo_id="DebdipCS/scribemark-ai-detection-benchmark",
    filename="nature_srep_2025_shuffled_data.csv",
    repo_type="dataset"
)
df = pd.read_csv(csv_path)
print(f"Loaded Nature Benchmark: {df.shape[0]} rows (Labels: {df['Labels'].value_counts().to_dict()})")

# 2. Download the Provable Forensic Fingerprinting Trail
import json
json_path = hf_hub_download(
    repo_id="DebdipCS/scribemark-ai-detection-benchmark",
    filename="FORENSIC_PROVABLE_FINGERPRINTING_TRAIL.json",
    repo_type="dataset"
)
with open(json_path, "r", encoding="utf-8") as f:
    trail = json.load(f)
print(f"Loaded {len(trail['forensic_trail'])} cryptographically signed forensic records.")

---

📄 Citation

@article{bandyopadhyay2026scribemark,
  title={ScribeMark: Cloze-Congruence Multi-Horizon Infilling and 12D Stylometric Manifold Forensics for Provable AI Detection with Zero False Accusation},
  author={Bandyopadhyay, Debdip},
  journal={arXiv preprint},
  year={2026},
  doi={10.5281/zenodo.22158286}
}