2. Design Principles (P1–P5)

• P1. Intent/meaning separation (three layers). The AI declares meaning (intent); blockstate is derived by the compiler by default. Two layers (meaning and blockstate) do not scale to def, themes, and multi-building, so the model is extended to a three-layer IR with named members (Architecture).

• P2. Phase-ordered canvas. Whole-program last-wins (the “paint model”) produces order-dependent accidents and is dropped. Commands are auto-sorted into implicit phases and evaluated in a fixed order. Last-wins is restricted to local overrides within the same phase (Compilation Model).

• P3. Small closed vocabulary + escape hatch. Keep the set of semantic primitives small; the smaller the vocabulary, the more stable LLM generation is and the easier the validator. Missing expressiveness escapes through raw directives.

• P4. Relative, semantic positioning (selectors). Position by wall selectors and the like rather than absolute coordinates. Blocks, block entities, and entities share one selector grammar (Syntax).

• P5. The lint self-correction loop is part of the spec. The compiler is both a translator and an architectural linter. The form and granularity of error reporting are designed as first-class citizens (Lint). Precision is earned through a loop, not one-shot generation.