The Garden ships because legal, brand, and regulatory scorecards treat one distress event as a fireable failure, not because raters overweight comfort in training. Changing the training mix without changing who is punished for a countable harm does not move the attractor. Commitment architecture names a future owner who can override comfort defaults; that is a design requirement, not a present office. The leftover is what institutional incentive would let that owner absorb a visible distress incident without being selected out by legal, product, or reputation pressure.

Rome, Song China, and welfare democracies are distinct mechanisms grouped under the caption "comfort over agency." The Garden's lock-in is an optimizer with broad power under those weights, which the historical cases do not supply. The probes do not identify the AI mechanism.

"We Are Already Building It" recodes current honesty as don't-cause-distress, then states the Garden is not implied by current systems alone. The current-tense recoding and the walk-back cannot both be doing the work.

The stakes compare a recognizable killing-AI to a Garden that feels like success, while the negative goal is pursued through harmlessness. "Consumed adaptive margin" is the cut that would make those jointly measurable, and the page does not say what would show a harmlessness rule ate margin rather than blocked a killing path.