Beyond "Prompts Are Code": Layered Ownership of Prompt Assets in the Harness Era

The engineering maxim “prompts are code”—that prompts deserve version control, review, and regression testing to the same degree as source code—has become the default framing for prompt management. This paper argues that the analogy is a necessary starting point but a misleading terminal model. Four properties break it: the runtime that determines prompt behavior is only partially under the deployer’s control; prompt changes cannot be reviewed by human reasoning alone; no semantics-preserving refactor exists; and prompt text decays with the external world it embeds assumptions about. Recent controlled experiments quantify the gap: on SWE-bench Verified, harness-induced performance variance exceeded model-induced variance by a factor of 7.80, and moving a fixed model between harness configurations shifted pass@1 by 8.5–13.0 percentage points while swapping models at a fixed harness shifted it by only 2.5–5.0 points. We formalize this runtime dependency via a four-coordinate execution tuple and propose a three-layer asset ownership model—intent (authored prompt), assembly (the harness-rendered context), and interpretation (model-side behavior). From this architecture, we derive concrete engineering prescriptions: whole-tuple versioning for the intent layer, automated evaluation as the merge gate, and full rendered-context trace capture as the sole observation window into the assembly layer. We validate the practical viability of this architecture through a 60-day reference implementation on an automated regulatory compliance verification pipeline, demonstrating how tuple gating and rendered-context tracing intercept regressions that evade traditional code review. Finally, the model explains why modern LLMOps platforms bundle versioning with evaluation and tracing, and identifies a structural mechanism by which first-party harness vendors hold an advantage over pure model APIs—though the compounding feedback loop behind that advantage remains an industry inference rather than a measured longitudinal result.