This paper presents LayerCake, a governed AI creative generation platform implementing a structured intent decomposition architecture that replaces free-form prompt input with a sequential plurality of governed input layers through which the platform assembles AI prompts without accepting raw text from users. A pre-generation governance gate enforces content policy before any model call is executed, preventing the creation of harmful content rather than moderating it after generation. The unified governance architecture applies identically across image generation (Slice) and copy generation (Ingredients) output types. All generated outputs and their complete provenance records are preserved in user-controlled account-held ledgers. Privacy is enforced structurally at the database level through row-level security. The platform collects no tracking data, employs no cookies, and does not use generated content for AI model training without explicit individual consent. A provisional USPTO patent has been filed covering this architecture (Application #64/083,877, June 5, 2026). Related publication: Stuart-Mueller, L. (2026). Vela Protocol: A Governance Framework for Human-Centered AI Infrastructure. Zenodo. DOI: 10.5281/zenodo.20564475