After tightening the system instructions on a permitting chatbot, a city team reruns the full golden-dataset evaluation suite rather than spot-checking a handful of prompts most related to the change. What risk does the full rerun cover that spot-checking the related prompts would not?

System instructions shape an agent's behavior broadly, so a change aimed at one scenario, like tightening scope constraints on a permitting question, can shift outputs on prompts that seem unrelated because the model's overall interpretation of instructions has moved, not just the narrow case the team edited for. Spot-checking only the prompts the team expects to be affected has a blind spot by construction: it can't surface a regression in a category nobody thought to look at, which is precisely the kind of failure a full golden-dataset rerun is designed to catch by covering the whole scenario space. Latency verification is a separate performance concern, unrelated to whether an instruction change altered response correctness. Confirming the model version itself is a platform/infrastructure check, not something a behavioral evaluation of instruction changes addresses. A full rerun does not reduce review volume — it's the more exhaustive option, and its value comes from that breadth, not from being lighter weight. Scope caveat: even a full rerun only catches regressions the golden dataset has coverage for, so a change affecting a scenario absent from the dataset can still slip through undetected. Operational check: after tightening instructions, diff the full run's scores category by category against the last baseline and confirm no category outside the intended change shows a drop.