From Prompt to Field Engineering: Govern the Context, Not the Output

Context-conditioned AI systems are commonly “governed” at the output layer through filters, prompt rules, and policy text, while the conditioning substrate remains underspecified. This mismatch produces a persistent audit gap: reviewers can inspect the answer but cannot reconstruct what actually executed, which tools were invoked, which retrieval snapshots were used, or which constraints were active. Our central claim is architectural rather than linguistic: stable behavior under drift requires governing context as a first-class object, not styling outputs after the fact. We introduce Field Engineering, a receiver-first approach that treats context as a Field F = ⟨D, G, C, R⟩, where D denotes domain scope, G denotes operational regime, C denotes an explicit contract with falsifiers and limits, and R denotes replayable receipts. In this view, prompts are dials that operate within a governed admissible space; they are not the governance boundary. We propose a minimal three-plane blueprint — Meaning Plane, Control Plane, and Evidence Plane — to separate semantic handling from enforcement and from auditability. We define admissible transforms Τ(F) as declared re-descriptions or environment changes under which stability claims are evaluated, preventing silent semantic drift disguised as “prompt improvement.” We introduce station tests as replay runs in independent environments that validate receipt completeness, contract satisfaction, and stability under Τ(F). We adopt a fail-closed gate discipline for promotion decisions: absent receipts and station-test stability, default promotion should not occur. We further argue that reliable governance is dominated by worst-slice behavior rather than average-case scores, motivating explicit worst-slice metrics and drift corridors. The proposed framework does not depend on model internals and applies across LLM, RAG, and tool-augmented stacks. We provide operational templates for contract spines, receipt spines, transform registries, and station-test harnesses to support implementation. This paper positions Field Engineering as a practical foundation layer for subsequent work on promotion compilers, transport tests, and enterprise tool governance.