SEED: A Structured Design Grammar Framework for AI Experimental Science

SEED (Structural Encoding for Experimental Discovery) is a structured design grammar framework for AI experimental science, aiming to address the reproducibility and auditability issues in human-AI collaboration and multi-agent experiments. Its core innovation is the typed participant-flow diagram representation, which supports formal description of experimental designs, structural novelty assessment, and candidate solution generation, and has been validated for feasibility in medical triage scenarios.

AI has transformed from a passive tool to an active participant, deeply integrating into human-AI collaboration and multi-agent systems. Traditional single-metric evaluation struggles to handle complex interaction scenarios. Current experiments rely on natural language descriptions, which have three major limitations: difficulty in cross-study comparison, implicit key decisions leading to reuse challenges, and inability to support systematic auditing in ethically sensitive fields.

SEED abstracts experiments into a graph structure composed of participants (humans, AI, hybrids) and interaction flows (information/task/control transfer), supplemented by a rich type system: participant types include human experts, AI assistants, etc., and flow types include task delegation, result review, etc., enabling formalization and computability of experimental designs.

1. Formal Description: Encode experimental conditions into structured interaction graphs to eliminate natural language ambiguity; 2. Novelty Assessment: Compare with existing designs to automatically identify innovative structures; 3. Candidate Generation: Automatically generate design solutions under constraints to explore the vast design space.

In medical triage tasks, designs generated under SEED's guidance are clearer than "graph-blind" designs, can identify key control gaps, potential failure modes, and necessary review mechanisms, verifying its feasibility as a design assistance tool.

Three tensions are discussed: 1. Balance between novelty and reproducibility; 2. Contradiction between standardized representation and exploration of diversity; 3. New issues of accountability and transparency brought by formal description.

SEED promotes the evolution of experimental science methodologies, providing new paths for reproducibility and auditability; in the future, it will facilitate the paradigm shift towards human-AI collaborative design and automated execution, while simultaneously advancing technical and governance considerations.