Code Origin Detector

A research-grade toolkit for estimating whether a source file was written by a human developer or produced by an AI assistant. The system combines language-aware heuristics, handcrafted features, and statistical models to output probabilistic verdicts with supporting rationale.

Status: early prototype. Ships with a heuristic baseline, stylometry/program-analysis feature scaffolding, a Typer CLI, and a redesigned React dashboard for demonstrations.

Table of Contents

• Screenshots
• Key Capabilities
• Getting Started
  • Backend CLI
  • Frontend Dashboard
  • Frontend Experience Highlights
• Testing and Quality
• Data Roadmap
• Responsible Use
• Citations and References

Screenshots

The Code Origin Detector provides an intuitive web interface for analyzing code authorship with real-time feedback and detailed explanations.

Key Capabilities

• AST and stylometry feature extraction for Python and JavaScript with a pluggable pipeline for additional languages.
• Interpretable heuristic signals (naming entropy, whitespace density, comment coverage, idiom usage) that double as model features and explanations.
• Statistical model hooks (logistic regression, random forest) with calibrated probabilities and optional stacking.
• Command-line interface for single files, directories, and benchmark manifests with JSON or pretty outputs.
• Dataset utilities for reproducible collection, hashing-based deduplication, and manifest-driven experiments.

README.md
pyproject.toml               # Python package metadata
requirements.txt             # Runtime dependencies for the CLI
.github/workflows/           # Continuous integration workflows
docs/                        # Design notes, prompt catalog, and reports
examples/                    # Sample walkthrough inputs and outputs
frontend/                    # React + Vite dashboard implementation
  public/                    # Static assets served by Vite
  src/                       # UI components, hooks, and utilities
  package.json               # Frontend tooling and scripts
notebooks/                   # Exploration and modelling notebooks
scripts/                     # Data collection and preprocessing jobs
src/                         # Python package (CLI, heuristics, features, models)
  detector/                  # CLI entrypoint and core detection logic
tests/                       # Python unit and integration suites

Getting Started

Backend CLI

python -m pip install -r requirements.txt
python -m pip install -e .
code-origin-detector predict ./path/to/project --include "*.py,*.js" --output-format pretty

The default heuristic_v1 model relies on interpretable rules. To use trained scikit-learn baselines, save a calibrated estimator with helpers in src/detector/models/train.py and point the CLI at the exported Joblib artifact (for example --model artifacts/rf_v1.joblib).

Frontend Dashboard

cd frontend
npm install
npm run dev

Open the Vite URL (default http://localhost:5173) to explore the detector flow, revised copy, and accessibility patterns. The UI runs client-side heuristics only; wire it to APIs as they become available.

Frontend Experience Highlights

• AnalyzerPanel includes filename editing, drag-and-drop uploads with validation, curated example tiles, model profile selection, and heuristic overlay toggles.
• ResultPanel surfaces verdict badges, a confidence meter, heuristics (expandable), limitations callouts, copy/download actions, and a selectable history timeline.
• Tailwind design tokens in tailwind.config.js extend brand/neutral palettes, card shadows, container widths, and motion-safe focus styles for consistency.
• Analysis settings (model, heuristics, explanation density) and runtime metadata are persisted per prediction for clearer provenance.

Testing and Quality

• Python: pytest, pytest --cov, ruff check, and mypy (install via pip install -e .[ci]).
• Frontend: npm run lint, npm test, npm run build, and npm run format.

Data Roadmap

1. Human corpus: Sample pre-2020 commits from vetted OSS repositories (license and language filters). Remove generated or minified artifacts, cap per-repo contributions, and tag as human-authored.
2. AI corpus: Generate program variants via scripted prompts in docs/prompts.md, recording task descriptions, temperatures, and rewrite strategies for reproducibility.
3. Splits and dedup: Apply SHA-256 hashing plus 20-token shingles to remove duplicates. Split train/validation/test by repository and maintain a temporal hold-out (post-2024 code).

Metadata schemas live in data/metadata/schema.json for consistent ingestion across tooling.

Responsible Use

The detector produces advisory signals, not definitive judgments. False positives can occur for human-written code (especially boilerplate or generated scaffolding), and AI-generated code can resemble expert human work. Treat probabilities as guidance to focus manual review, not as an automated gate or policy decision.

Citations and References

Core Technologies

Backend

• Python (≥3.10) - https://www.python.org/
• Typer (≥0.9) - Command-line interface framework - https://typer.tiangolo.com/
• Rich (≥13) - Terminal output formatting - https://rich.readthedocs.io/
• scikit-learn (≥1.3) - Machine learning library - https://scikit-learn.org/
  • Pedregosa et al., "Scikit-learn: Machine Learning in Python", JMLR 12, pp. 2825-2830, 2011
• NumPy (≥1.23) - Numerical computing - https://numpy.org/
• pandas (≥2.0) - Data analysis and manipulation - https://pandas.pydata.org/
• SHAP (≥0.44) - Model interpretability - https://shap.readthedocs.io/
  • Lundberg & Lee, "A Unified Approach to Interpreting Model Predictions", NeurIPS 2017
• tree-sitter (≥0.20) - Parser generator for syntax trees - https://tree-sitter.github.io/
• Radon (≥6.0) - Code complexity metrics - https://radon.readthedocs.io/

Frontend

• React (^19.1.1) - UI framework - https://react.dev/
• Vite (^7.1.7) - Build tool and development server - https://vite.dev/
• TypeScript (~5.8.3) - Type-safe JavaScript - https://www.typescriptlang.org/
• Tailwind CSS (^3.4.17) - Utility-first CSS framework - https://tailwindcss.com/

Related Research

For research on code authorship attribution and AI-generated code detection, refer to:

• Burrows et al., "Stylometry and Code Authorship Attribution", Digital Investigations 2016
• Caliskan et al., "De-anonymizing Programmers via Code Stylometry", USENIX Security 2015
• Solaiman et al., "Release Strategies and the Social Impacts of Language Models", arXiv 2019

Citation

If you use Code Origin Detector in your research, please cite:

@software{code_origin_detector,
  title = {Code Origin Detector},
  author = {Code Origin Detector Team},
  year = {2025},
  version = {0.1.0},
  url = {https://github.com/rasikasrimal/code-origin-detector}
}

For the full citation metadata, see CITATION.cff.