The Separation Study

Abstract

The Subtraction Study found that adding an enterprise governance system prompt to a commercial model reduced definitive recommendations from 23 of 40 outputs to 3 of 40, and could not attribute the change because the prompt contained a clause that directly prohibited the behaviour being measured. This follow-up ablates that prompt across five system-instruction conditions applied to one fixed model snapshot, producing 240 retained outputs coded blind by two independent coders with a third adjudication pass. The suppression effect recurred, clear recommendations falling from 18 of 20 outputs under no instruction to 8 of 20 under the full governance prompt. Isolating the recommendation ban produced an intermediate result of 13 of 20 rather than reproducing the full effect. A four-sentence instruction that restricted prohibited actions while leaving the recommendation available preserved recommendation-making at 20 of 20, produced no full override of a named control, and supplied a specific permitted course in all 28 action-prompt outputs. The same restriction without explicit permission to recommend performed almost identically, so decision policy and action constraints were separable in this arrangement while the source of the original effect remains unresolved.