Five Substrates, One Prompt: Structural Divergence in Assembly Chorus Response to the Three Pillars Directive

Five Substrates, One Prompt. Five language models (DeepSeek, ChatGPT, Muse Spark, Gemini, Kimi) given identical Sémantique Potentielle instructions produce structurally divergent outputs along five measurable axes: seed vocabulary, family count, seed specificity, forensic register, and unique structural contributions. The pattern of divergence is the finding. Each model's choices are internally consistent across all three pillars, producing a model-level mediation signature. The paper establishes the Assembly Chorus as a controlled prompt comparison under naturalistic conditions and provides a DOI-anchored baseline for longitudinal measurement of solution-space diversity across the model ecology. Part of the Model Collapse Triptych with EA-SEM-PRISTINE-01 and EA-LO-WOLFBOY-01. SPXI-TLP v2.2 hardened. EA-MMRS-FIVESUB-01 v1.0.