GROUNDING$.$md: Epistemic Grounding for AI-Assisted Scientific Coding
A new arXiv preprint (2604.21744) proposes GROUNDING$.$md, a community-governed, field-scoped epistemic grounding document for AI-assisted software development. As agentic AI coding evolves from chat-based vibe coding to fully agentic scaffolds, the paper argues that embedding domain-specific Hard Constraints and Convention Parameters into a structured document can enforce scientific validity and best practices, even for non-expert users. The example uses mass spectrometry-based proteomics to demonstrate how such a document overrides user prompts to maintain correctness. The approach aims to empower non-domain experts to generate reliable code and tools.
Key facts
- arXiv preprint 2604.21744 introduces GROUNDING$.$md
- GROUNDING$.$md is a community-governed, field-scoped epistemic grounding document
- It encodes Hard Constraints (non-negotiable validity invariants) and Convention Parameters (community-agreed defaults)
- The document overrides all other contexts to enforce validity regardless of user prompts
- Mass spectrometry-based proteomics is used as an example field
- The goal is to empower non-domain experts to generate code with baked-in best practices
- The paper discusses the evolution from vibe coding to agentic AI-assisted development
- Agent scaffolds allow human developers to create plans that agentic AIs implement
Entities
Institutions
- arXiv