Prose Is the Database: Compiling a Signed, Adoption-Probed Controlled Language into Running Applications with Small Local Models

PREPRINT. Large language models are usually pointed at code or machine formats and bent toward validity with prompts, schemas, or constrained decoding. This work inverts the direction of fit: fix a small, closed, human-readable fact language and MEASURE which sentence forms small local models already speak, closing the grammar around their measured speech (98.2% strict adoption across 489 unconstrained lines on a 4B model). The language serializes a graph; the graph projects deterministically onto a spec-driven code generator; the result is a pipeline from a client's verbatim need-sentence to a running application whose permission rules are enforced exactly as signed in prose. A signed, hash-pinned 'wisdom pack' lifted four small local models across three families (composites 4-29 baseline to 89-102.5 guided) with a built-in negative control (-160, same model, self-authored manual). Everything runs on consumer hardware; nothing is fine-tuned. Withheld artifacts are committed to by SHA-256 in the appendix. License on publication: CC BY 4.0 (text only).