AI Text Cascade Detect

<p align="center"> <img src="docs/assets/dtd-lrc-banner.svg" alt="DTD-LRC: uncertainty-gated evidence escalation for open-world machine-generated text detection" width="100%" /> </p> <p align="center"> <a href="https://huggingface.co/spaces/YohanChow/DTD-LRC-AI-Text-Detector"><img alt="Live demo" src="https://img.shields.io/badge/Live_Demo-Hugging_Face-FFD21E?logo=huggingface&logoColor=000"></a> <a href="https://youtu.be/Z0NXF_Ghasg"><img alt="Demo video" src="https://img.shields.io/badge/Demo_Video-YouTube-FF0000?logo=youtube&logoColor=white"></a> <a href="paper/DTD-LRC_UncertaiNLP_2026/"><img alt="Submission status" src="https://img.shields.io/badge/Manuscript-Submitted_to_UncertaiNLP_2026-orange"></a> <a href="LICENSE"><img alt="License" src="https://img.shields.io/badge/License-MIT-2ea44f"></a> <img alt="Python" src="https://img.shields.io/badge/Python-3.10%2B-3776AB?logo=python&logoColor=white"> </p> <p align="center"> <strong>Fast when confident. Deeper when uncertain.</strong><br/> A sole-author research artifact for uncertainty-aware, evidence-assisted AI-text detection. </p>

Overview

DTD-LRC is a two-stage cascade for open-world machine-generated text detection:

• DTD performs fast, lightweight first-stage detection from lexical, stylistic, syntactic, punctuation, and repetition features.
• MS-LRC is activated only inside an uncertainty band and examines cross-family, cross-scale language-model evidence.
• The system returns not only a decision, but also uncertainty and diagnostic evidence intended for human review.

The workshop manuscript, “DTD-LRC: Uncertainty-Gated Evidence Escalation for Open-World Machine-Generated Text Detection,” has been submitted to UncertaiNLP 2026 @ EMNLP through OpenReview. Submission does not imply acceptance.

Research contribution

The project studies a practical question:

Can a detector preserve low-cost inference for clear cases while escalating only ambiguous samples to deeper, more interpretable evidence?

DTD-LRC contributes:

1. Uncertainty-gated routing rather than always-on expensive inference.
2. Cross-family and cross-scale evidence through NLL-per-byte response patterns.
3. Ladder response curves and a family-scale matrix for structured Stage-2 evidence.
4. A reviewer-facing artifact with a live demo, API, precomputed evidence cards, full local pipeline, manuscript files, and explicit claim boundaries.

System architecture

flowchart LR
    A[Input text] --> B[Stage 1: DTD]
    B --> C{AI probability inside<br/>uncertainty band [0.41, 0.61]?}
    C -- No --> D[Return Stage-1 decision]
    C -- Yes --> E[Stage 2: MS-LRC]
    E --> F[NLL / byte across<br/>model families and scales]
    F --> G[Ladder response curves<br/>and family-scale matrix]
    G --> H[Escalated decision<br/>with interpretable evidence]

Headline validation results

| Component | Metric | Value | |---|---:|---:| | Stage 1 — DTD | AUC | 0.9899 | | Stage 1 — DTD | F1 | 0.9600 | | Stage 1 — DTD | Precision / Recall | 0.96 / 0.96 | | Stage 1 — DTD | Average latency | 38 ms | | Routing gate | Uncertainty band | [0.41, 0.61] | | Stage 2 — MS-LRC smoke evaluation | AUC | 0.9622 | | Stage 2 — MS-LRC smoke evaluation | Accuracy | 0.9333 | | Stage 2 — MS-LRC smoke evaluation | F1 | 0.9375 | | Cascade smoke evaluation | Stage-2 usage | 31 / 120 |

In the 120-example cascade smoke evaluation, 74.2% of samples were resolved at Stage 1, while 25.8% were escalated.

These are internal artifact-validation results. They are not universal performance guarantees, and external comparisons require matched datasets, thresholds, preprocessing, and evaluation protocols. See Reproducibility and Claim Boundaries.

Artifact links

| Artifact | Link | |---|---| | Interactive demo | Hugging Face Space | | Direct application | Deployed app | | Demonstration video | YouTube | | Manuscript files | paper/DTD-LRC_UncertaiNLP_2026/ | | Artifact manifest | ARTIFACT_LINKS.md |

Quick start

Lightweight reviewer mode

This mode runs interactive DTD inference and displays precomputed MS-LRC evidence without loading transformer models during startup.

git clone https://github.com/haveanicedaymydear/AI-Text-Cascade-Detect.git
cd AI-Text-Cascade-Detect
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py

Windows PowerShell:

.\.venv\Scripts\Activate.ps1

Open:

http://127.0.0.1:5000

Full local cascade

pip install -r requirements-full.txt
python src/cascade_demo_app.py

The full MS-LRC path requires transformer model downloads or cache and more memory than the lightweight artifact.

API

Health check:

curl http://127.0.0.1:5000/api/health

Detection request:

curl -X POST http://127.0.0.1:5000/api/detect \
  -H "Content-Type: application/json" \
  -d '{"text":"The implementation of artificial intelligence systems requires careful evaluation of reliability."}'

Precomputed Stage-2 examples:

curl http://127.0.0.1:5000/api/mslrc_examples
curl http://127.0.0.1:5000/api/mslrc_examples?id=uncertain_sample

Repository map

.
├── app.py                          # Reviewer-facing Flask/API entry point
├── src/
│   ├── web_app.py                  # Original lightweight DTD application
│   └── cascade_demo_app.py         # Full local cascade
├── models/                         # Public model artifact(s)
├── examples/                       # Human, AI, uncertain, and MS-LRC examples
├── paper/DTD-LRC_UncertaiNLP_2026/ # Submitted manuscript files
├── templates/                      # Web interfaces
├── tests/                          # Project tests
├── docs/                           # Reproducibility notes and visual assets
├── requirements.txt                # Lightweight demo dependencies
└── requirements-full.txt           # Full local pipeline dependencies

Demo modes

| Mode | Command | Purpose | |---|---|---| | Reviewer artifact | python app.py | Interactive DTD + precomputed MS-LRC evidence | | Original DTD app | python src/web_app.py | Original trained DTD web/API application | | Full local cascade | python src/cascade_demo_app.py | Transformer-backed Stage-2 execution |

Limitations and responsible use

DTD-LRC is a probabilistic aid, not an authority. It should not be used as the sole basis for punitive academic, employment, authorship, moderation, or disciplinary decisions.

Known limitations include:

• false positives on short, formulaic, translated, heavily edited, creative, or stylistically unusual human writing;
• domain shift and generator shift;
• calibration drift under unseen distributions;
• reduced reliability on multilingual or code-mixed text unless separately evaluated;
• higher compute cost for full MS-LRC inference;
• vulnerability to future generators and deliberate evasion strategies.

Recommended use is evidence-assisted review: combine the output with contextual evidence, provenance, human judgment, and an appeal or correction process.

Citation

Citation metadata is available in CITATION.cff. Until a final archival publication exists, cite the repository and submitted manuscript as a research artifact:

@misc{zhou2026dtdlrc,
  author       = {Livan Zhou},
  title        = {DTD-LRC: Uncertainty-Gated Evidence Escalation for Open-World Machine-Generated Text Detection},
  year         = {2026},
  howpublished = {Sole-author manuscript submitted to UncertaiNLP 2026 @ EMNLP and open-source research artifact},
  url          = {https://github.com/haveanicedaymydear/AI-Text-Cascade-Detect}
}

Author

Livan Zhou
First Class Honours BSc in Computing Science
GitHub · Email

License

Released under the MIT License.