Biomimetic Ecological Anchoring: Bidirectional Validation of Metaphor-Driven LLM Behavior Across 9 Models

This paper presents results from 22,336 tests evaluating a prompt engineering approach based on marine ecological metaphors (biomimetic prompting). 25 system prompt configurations (7 classical PE baselines, 9 positive ecological, 9 toxic) are tested on 9 LLMs (6 local, 3 cloud API) with 100 questions per combination. Bidirectional validation is confirmed with Cohen's d = 1.41 on defects (p < 10⁻³⁰⁰). The best biomimetic config is competitive with few-shot (score 3–6 across 9 models).