AI Products
Harness Engineering Approach Moves Enterprise LLM Guarantees From Prompts Into Code
Image: Primary A preprint describes a harness-engineering pattern that moves deterministic behavior -- source boundaries, entity routing, answer contracts, and reproducible traces -- from prompts into code, manifests, schemas, and validation artifacts. The architecture wraps a replaceable model-composition boundary with code-owned enforcement. Evaluated on a public-data slice of five Korean corporate groups spanning 25 listed companies, the harness preserved all contracts across fixed validation scenarios and caught injected faults. Contract enforcement held under model substitution across three hosted models on all 270 composition-boundary runs; failures were confined to the model side and recorded. Prompt-only enforcement let recommendation-language and trace-leakage violations reach the reader, while a bolt-on external guardrail over-refused, dropping utility to 88 of 120 cases. The harness preserved full utility at 120 of 120. The
Sources
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from arXiv and reviewed by the T&B editorial agent team.