<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> Draftpaper-loop addresses the challenge of creating auditable, traceable research paper drafts by providing a local-first, evidence-driven loop engine. It targets researchers and technical users, enabling repeatable scientific workflows and structuring accumulated knowledge for faster field entry and cross-disciplinary insight.

How It Works

The system employs a deterministic outer loop that orchestrates research and writing through explicit, rerunnable stages. It cycles through observing project state, deciding the next action, executing commands, verifying outputs against automated gates, and persisting artifacts. This hybrid approach combines deterministic handling of scientific contracts (e.g., citation existence, result binding) with agent-assisted open-ended tasks like literature interpretation. Key components include defined Goals for each stage, Context management via stable project files (not chat history), a controlled CLI Tool surface, automated Evaluation gates, and clear Stop conditions. The core philosophy is evidence-first: auditable evidence is established before natural language prose is generated.

Quick Start & Requirements

Installation involves cloning the repository, setting up a Python virtual environment, and installing the CLI with plotting extras: pip install -e .[plotting]. Prerequisites include Python 3.x and Git. Optional full-text fetching requires installing third_party\paper-fetch-skill. A basic tutorial video is available on Bilibili.

Highlighted Details

• Local-first, single-paper project model with staged manifests for auditable workflows.
• Evidence-first approach prioritizes verifiable data and results before prose generation.
• Traceable citation evidence (citation_evidence.csv) ensures manuscript claims are auditable.
• Pluggable data acquisition planning and discipline-specific reviewer-engineering engines enhance domain adaptability.
• Metadata-only research-code mining facilitates curation of potential discipline modules without code copying.
• Hash-based stale detection and explicit artifact traceability ensure reproducibility.

Maintenance & Community

The project is primarily maintained by Jinray Xie, with contributions from Chen Wei. It actively welcomes reusable discipline-module contributions, such as data connectors, method templates, and reviewer rules, to expand its capabilities across various scientific fields.

Licensing & Compatibility

Draftpaper-loop is source-available under a non-commercial license (DPL schema family). Commercial use, SaaS deployment, or integration into commercial products requires separate written authorization from the developer. This license preserves commercial authorization requirements and is not compatible with standard open-source licenses like Apache-2.0.

Limitations & Caveats

The system's user experience is currently strongest in domains like geography, environmental science, and remote sensing. Broader applicability relies on the accumulation of discipline-specific data connectors, method templates, and reviewer rules. Commercial adoption is strictly prohibited without explicit licensing.