Regulated financial systems demand deterministic behavior -- auditability, repeatability, and strict policy enforcement. AI systems are probabilistic. Introducing AI into these environments without principled constraints creates compliance risk and institutional liability. This paper formalizes the Epistemic Harness, an architectural pattern for governing AI in regulated financial systems. The pattern positions LLMs as grammar-derivation engines: given a formal specification, the model produces a scoped intermediate representation that downstream deterministic systems can consume and enforce. The harness wraps and constrains AI behavior. It does not replace existing processes. The core argument is that existing engineering infrastructure -- version-controlled specifications, CI/CD pipelines, and human review gates -- is sufficient to govern AI in high-compliance environments. No new platforms or governance tooling are required. The pattern enables non-disruptive AI retrofit into legacy systems and positions organizations for migration readiness as requirements evolve. A reference implementation applying this pattern to NACHA/ACH payment file validation is available at github. com/torjoshi/nachaₚarser.
Rajesh Joshi (Tue,) studied this question.