This repository provides an open-source implementation of the Kosmos AI scientist architecture, addressing critical gaps left in the original paper. It targets researchers and developers aiming for autonomous scientific discovery, offering a foundation for experimentation and extension of the Kosmos framework.

How It Works

The project implements solutions for six foundational gaps identified in the Kosmos paper, leveraging patterns from the K-Dense ecosystem. Key innovations include a hierarchical 3-tier context compression mechanism (20:1 ratio) for handling large document sets, a hybrid 4-layer state manager (JSON, knowledge graph, vector store, citation tracking), and a Plan Creator/Plan Reviewer orchestration for strategic task generation. Agent integration uses a skill loader with domain-specific prompts, while code execution is managed via a secure, Docker-based Jupyter sandbox with pooling and resource limits. Discovery validation employs an 8-dimension ScholarEval quality framework.

Quick Start & Requirements

• Installation: Clone the repository, cd into it, and run pip install -e .. Copy .env.example to .env and configure API keys.
• Prerequisites: Python 3.11+, Anthropic or OpenAI API key, Docker (for sandboxed code execution). Optional: Neo4j, Redis.
• Documentation: See GETTING_STARTED.md for detailed examples. Configuration is via environment variables (.env.example).

Highlighted Details

• All 6 critical implementation gaps from the Kosmos paper are addressed with working code.
• Leverages K-Dense ecosystem patterns for context compression, orchestration, agent skills, and validation.
• Features a Docker-based Jupyter sandbox for secure, efficient code execution with container pooling and dependency management.
• Includes a hybrid state manager supporting JSON artifacts, optional Neo4j knowledge graphs, and vector stores.
• Implements the ScholarEval 8-dimension framework for rigorous discovery validation.

Maintenance & Community

The project was last updated on November 25, 2025, with version 0.2.0-alpha. Contributions are welcomed, particularly in Docker sandbox testing, integration test updates, R language support, and performance benchmarking. No specific community channels (e.g., Discord, Slack) are listed.

Licensing & Compatibility

The project is released under the MIT License, permitting commercial use and integration into closed-source projects.

Limitations & Caveats

Code execution relies on Docker; without it, mock implementations are used. The arxiv package has Python 3.11+ compatibility issues, limiting literature search features. The implementation is Python-only, lacking R language support mentioned in the paper. Full autonomous research loops (20 cycles, 10 tasks each) have not been end-to-end validated, and the system may converge early or require manual intervention. Significant API costs are incurred for extensive research runs without optimization. The project is considered experimental, not production-ready, and has accumulated technical debt. It implements the architecture but does not reproduce the paper's claimed scientific discoveries or accuracy metrics.