Enterprise AI teams often defer reproducibility infrastructure for production LLM deployments, assuming it is expensive, but a defensible cost model shows the infrastructure costs are an order of magnitude lower than most teams imagine. The deferral typically rests on unvalidated guesses, such as "maybe six figures a year" or "a lot of engineering time," rather than a cost model. The article identifies four cost categories: audit-trace storage, which includes durable persistence of the full execution trajectory for each decision, stored for an audit retention window of typically five to seven years for regulated industries; compute overhead for measurement, covering additional computation for attribution stability under variance; engineering build cost, a one-time effort for deterministic seed handling, environment fingerprinting, persistence pipelines, and measurement protocols, plus ongoing maintenance; and operational overhead. The engineering cost of building this infrastructure once is similarly bounded, while the cost of not having it—recreating audit defense in three years under regulatory pressure—is often underestimated. The math has shifted in the last eighteen months.