Source note

From Prompts to Contracts: Harness Engineering for Auditable Enterprise LLM Agents

LLM AgentsHarness EngineeringCode IntelligenceAuditable AIEnterprise Software

The paper presents a code-owned harness that turns prompt-based enterprise LLM prototypes into auditable agents with source, routing, output, and trace contracts. A Korean corporate-data investment briefing system shows that these controls survive model substitution and preserve more utility than an external guardrail.

  • Prompt-dominant prototypes do not reliably enforce source boundaries, entity routing, claim eligibility, answer structure, trace completeness, or restrictions on recommendation language.
  • These failures matter in enterprise settings because unsupported claims, mixed-entity evidence, leaked internal fields, and unreproducible answers can make a system unsafe to deploy or audit.
  • Register approved sources in manifests, extract evidence records, and promote only atomic, entity-scoped, provenance-linked statements into a runtime-eligible source-backed claim set.
  • Move source admission, entity routing, claim selection, answer planning, follow-up filtering, trace generation, and validation into code, schemas, manifests, and validators.
  • Keep the language model at a replaceable composition boundary. The model phrases a bounded claim package, while the harness checks the result and falls back to a deterministic composer when needed.
  • Record both the reader-facing answer and an audit trace covering routing, source states, selected claims, validation outcomes, and fallback paths.
  • Evaluate the system on 5 Korean corporate groups, 25 listed companies, 113 source-backed claims, 30 validation scenarios, 3 hosted models, and 3 repeats per scenario.
  • The harness preserved source-grounding, entity-routing, trace, output-hygiene, and recommendation-language contracts across the fixed validation scenarios; fault injection caused the validators to flag deliberately broken contracts.
  • All 270 live composition-boundary runs passed the harness contracts across 3 hosted models. Model-side failures were caught and recorded by the control layer.
  • In the enforcement ablation, prompt instructions alone allowed recommendation-language and internal-trace-leakage violations to reach users. Code-owned enforcement blocked these violations.
  • An external bolt-on guardrail achieved 88/120 utility, while the harness preserved 120/120 utility while blocking the tested violations.
  • The study measures system verifiability and contract enforcement rather than investment-decision quality, and it does not provide evidence from a broad production deployment or user-quality benchmark.