A settling substrate is supposed to be governed by the energy you programmed into it, and lessons 7 and 8 both lean on that: equilibrium is the sample, and the difference of two equilibria is the gradient. Both statements are about equilibrium, and a machine on a clock does not reach one. What you get instead depends on two things the energy never mentions. The first is the relaxation budget. Run equilibrium propagation with one sweep per phase and start the nudged phase where the free phase stopped, which is what the literature specifies and what hardware does, and the free reference is itself unsettled; at zero nudge the nudged phase still travels further than the reference did, so the difference does not vanish when the nudge does, and dividing by the nudge turns that leftover into a divergence. Shrinking beta, which the settled theory says is always safer, makes this machine two orders of magnitude worse. The second is the schedule. Time-multiplexing one physical p-bit across several logical ones is normally argued to be free because it only slows the clock, and measuring the stationary distribution exactly shows it is not: at a reuse factor of three the law the fabric samples differs from the unmultiplexed one on a quarter to two thirds of the probability mass. Neither effect is a bug in the hardware. Both are the schedule and the budget behaving as what they are, which is part of the model.