Use of Generative AI in HydraR

In accordance with the rOpenSci AI Policy and JOSS Generative AI Statement, this document discloses the use of Large Language Models (LLMs) and agentic workflows during the development of HydraR.

🤖 AI-Aided Development

A significant portion of the HydraR codebase was developed using Antigravity, an agentic AI coding assistant. The development process followed a rigorous “Human-in-the-loop” pattern:

1. Architecture & Design: The core R6 architecture, state management patterns, and Git worktree isolation strategies were designed and authored by the human authors (Ignatius Pang and Aidan Tay) during a focused 4–5 day architectural sprint and then rigorously tested manually.
2. Implementation: Antigravity was used to implement specific logic blocks, boilerplate R6 methods, unit tests (testthat), and documentation (roxygen2).
3. Verification: Every line of code generated by Antigravity was manually reviewed, modified, and tested by the human authors. Build checks (devtools::check()) and local integration tests were run after every significant AI contribution to ensure correctness and adherence to APAF Bioinformatics standards.

📍 Agentic Roles in HydraR

HydraR is not just a package built with AI; it is a framework for AI. It handles agents as first-class citizens:

• Logic Nodes (AgentLogicNode): These are human-defined logic units that can include hardcoded functions or LLM-driven prompts.
• LLM Nodes (AgentLLMNode): Specifically designed for asynchronous, state-aware LLM interactions.
• Auditors: Specialized agents or functions used to validate the output of other nodes, ensuring a high degree of autonomous quality control.

🛡️ Affidavits

• All AI-generated material has been carefully reviewed by the human authors.
• No AI-generated code has been included without manual verification and testing.
• The authors take full responsibility for the correctness and security of the codebase.

---