The authors present a harness-engineering approach that transforms enterprise LLM prototypes into traceable, auditable architectures by moving deterministic behavior into code and validation artifacts around a replaceable composition boundary.
- The system preserves source-grounding, entity-routing, trace, output-hygiene, and recommendation-language contracts across fixed validation scenarios.
- Validators successfully flag deliberately broken contracts in fault-injection controls.
- Enforcement checks held under model substitution, passing all 270 composition-boundary runs across three hosted models.
- Code-owned guarantees prevent violations that prompt instructions alone cannot block, such as internal-trace leakage.
- Unlike external guardrails that drop utility to 88/120 due to over-refusal, the harness preserves full utility (120/120) while maintaining safety.
This reusable engineering pattern allows developers to turn exploratory prototypes into auditable applications with versioned source, control, and validation artifacts.