# Kodezi Chronos ## The World's First Debugging-First Language Model for Repository-Scale, Memory-Driven Code Understanding [](https://arxiv.org/abs/2507.12482) [](https://chronos.so) [](LICENSE) [](paper/chronos-research.md) [](benchmarks/multi-random-retrieval/)

67.3% Autonomous Debugging Success 89% Human Preference 7.8 Average Fix Iterations 40% Time Reduction

MRR Benchmark Results

Performance by Bug Category

Repository Scale Performance

Key Innovations in Chronos

1. Debugging-First Architecture

• Trained on 42.5M real debugging examples (not code completion)
• Specialized for root cause analysis and multi-file patches
• 78.4% root cause accuracy vs 15.8% best baseline

2. Persistent Debug Memory (PDM)

• Repository-specific learning from debugging sessions
• Improves from 35% 65% success rate over time
• Cross-session pattern recognition

3. Adaptive Graph-Guided Retrieval (AGR)

• O(k log d) complexity with dynamic k-hop expansion
• 92% precision, 85% recall on multi-file context
• Handles unlimited repository scale intelligently

4. Output-Optimized Design

• Optimized for ~3K output tokens (fixes, tests, docs)
• 47.2% output entropy density vs 12.8% for completion models
• Designed for complex patch generation

5. Autonomous Debugging Loop

• Average 7.8 iterations to successful fix
• Propose test analyze refine cycles
• 67.3% fully autonomous success rate

Architecture Overview

Seven-Layer System Design

1. Multi-Source Input Layer: Processes code, logs, traces, tests, docs simultaneously
2. Adaptive Retrieval Engine (AGR): Dynamic k-hop graph traversal (92% precision)
3. Debug-Tuned LLM Core: 42.5M debugging examples, not code completion
4. Orchestration Controller: Autonomous debugging loop management
5. Persistent Debug Memory: Repository-specific learning (35% 65% improvement)
6. Execution Sandbox: Isolated test validation environment
7. Explainability Layer: Human-readable root cause analysis

See architecture documentation for detailed specifications.

Multi-Random Retrieval (MRR) Benchmark

What is MRR?

MRR simulates real-world debugging complexity by: - Spatial Distribution: Bug context scattered across 10-50 files - Temporal Dispersion: Relevant information from 3-12 months of history - Obfuscation Levels: Low/medium/high code complexity - 5,000 Scenarios: Comprehensive evaluation across languages and bug types

MRR Results

| Metric | Chronos | GPT-4+RAG | Claude-3+VectorDB | Gemini-1.5+Graph | |:-------|:-------:|:---------:|:-----------------:|:----------------:| | Precision@10 | 89.2% | 42.3% | 48.1% | 51.7% | | Recall@10 | 84.7% | 31.7% | 36.2% | 41.8% | | Fix Accuracy | 67.3% | 8.9% | 11.2% | 14.6% | | Context Efficiency | 0.71 | 0.23 | 0.28 | 0.31 |

Full benchmark available in benchmarks/multi-random-retrieval/

Getting Started

Running the MRR Benchmark

```bash

Clone the repository

git clone https://github.com/kodezi/chronos-research.git cd chronos-research

Install dependencies

pip install -r requirements.txt

Run MRR benchmark on your model

python benchmarks/runmrrbenchmark2025.py \ --model yourmodel \ --scenarios 100 # Start with subset

Analyze results

python benchmarks/analyze_results.py ```

Model Access

** The Chronos model is not included in this repository**

Chronos will be available via Kodezi OS: - Q4 2025: Enterprise beta - Q1 2026: General availability - Join waitlist: chronos.so

Repository Contents

chronos-research/ benchmarks/ # MRR Benchmark Suite multi-random-retrieval/ # 5,000 scenario benchmark evaluation_metrics/ # Metrics implementation run_mrr_benchmark_2025.py # Main benchmark runner reference_implementations/ # Algorithm references (NOT the model) algorithms/ # AGR, PDM implementations NOTICE.md # Proprietary model notice paper/ # Research paper chronos-research-2025.md # Full paper (arXiv:2507.12482) results/ # Performance data raw_data/ # 5,000 scenario results case_studies/ # Debugging examples figures/ # Paper visualizations paper_figures/ # 11 paper figures docs/ # Documentation MODEL_ACCESS.md # How to access Chronos LEADERBOARD.md # Performance rankings

Research Highlights

Training Dataset

• 42.5M debugging examples (not code completion)
• 15M GitHub issues with fixes
• 8M stack traces with resolutions
  
• 3M CI/CD debugging logs
• 2.5M production sessions
• 14M curated from Defects4J, SWE-bench, BugsInPy

AGR Performance

• k=1 hop: 58.2% success
• k=2 hops: 72.4% success
• k=adaptive: 87.1% success
• Flat retrieval: 23.4% success

PDM Learning Curve

• Initial: 35% success rate
• After 100 sessions: 52% success
• After 500 sessions: 65% success
• 7.3x token efficiency gain

Detailed Performance Analysis

Language-Specific Performance

Debugging Cycle Efficiency

Context Window Efficiency

Ablation Studies

Documentation

Contributing

We welcome contributions to the evaluation framework and benchmarks!

```bash

Fork and clone

git clone https://github.com/[your-username]/chronos-research cd chronos-research

Create feature branch

git checkout -b feature/your-contribution

Make changes and test

python -m pytest tests/

Submit PR

git push origin feature/your-contribution ```

See CONTRIBUTING.md for detailed guidelines.

Citation

If you use this research in your work, please cite:

bibtex @article{khan2025chronos, title={Kodezi Chronos: A Debugging-First Language Model for Repository-Scale, Memory-Driven Code Understanding}, author={Khan, Ishraq and Chowdary, Assad and Haseeb, Sharoz and Patel, Urvish}, journal={arXiv preprint arXiv:2507.12482}, year={2025}, url={https://arxiv.org/abs/2507.12482} }

About Kodezi

Kodezi is building the future of autonomous software maintenance. Our mission is to empower developers with AI that truly understands code at scale.

Contact & Community

License

This research repository is licensed under the MIT License - see LICENSE for details.

** Important**: The Kodezi Chronos model itself is proprietary technology and is not included in this repository. Model waitlist access is available at chronos.so.