This paper presents a harness-engineering pattern for turning enterprise LLM prototypes into auditable applications. Deterministic requirements are moved into code, manifests, schemas, and validation artifacts around a replaceable model-composition boundary, while source-backed claims remain authoritative at runtime. On a public-data slice covering five Korean corporate groups and 25 listed companies, the authors report 270 composition-boundary runs across three hosted models, with harness-enforced contracts passing throughout. Prompt-only enforcement allowed recommendation-language and internal-trace-leakage violations to reach readers. An external guardrail blocked those violations but reduced utility to 88/120, whereas the harness retained 120/120 utility in the ablation.
No heat snapshots are available in the last 24 hours.