Measured by the ladder's own verdict, do lessons move the walkers toward better alloys?

E162 — The walk measured in the ladder's own units, not the circuit's

E160's climb metric divides by a landscape sd fixed before learning; a learned readout rescales the score as its weights grow, so "climb in sd" is not scale-free and E160's −0.58 for the whole brain may be an artefact of the ruler. The scale-free measure is the ladder's own verdict at each walker's position: Verifier.screen mapped to p exactly as the search's reward is, evaluated on all 24 walkers every round. Same four passes as E160 (both assets, learned readout, learning off then on), same seeds; recorded per round: mean p over the swarm, and mean p over the round's top four by circuit score — the proposals the search would actually send for verification.

Predicted:

1. The mushroom body's learning contribution is positive in p: mean swarm p over the last ten rounds is at least +0.05 higher with learning on than off.
2. The whole brain's sign flips back: its learning contribution in p is positive (≥ +0.03), which would make E160's −0.58 the rescaling artefact I suspected.
3. Top-four p rises with learning on both assets, and the whole brain's top-four gain is at least the mushroom body's — consistent with E159's 7 finds against 5.5: the readout improves what gets verified more than how far the swarm climbs.

Falsified if the whole brain's top-four p with learning on is no higher than with it off. Then the walk explains none of E159's 17x, the metric was not the problem, and what is left is verification-time selection or two lucky seeds — in which case E161's second pair of seeds, with the readout saved, is the arbiter and step 2b should not be built on E159 alone.

E162 result — falsified on all three counts, and the pre-registered branch fires. Rung-0 p at every walker position, mean over the last ten of forty rounds:

learned readout           mushroom off → on      whole brain off → on
swarm mean p             0.153 → 0.101 (−0.052)   0.133 → 0.119 (−0.014)
top-four p               0.158 → 0.134 (−0.024)   0.076 → 0.072 (−0.004)
first-ten swarm p        0.077   0.077            0.107   0.106

Prediction 1 falsified (mushroom gain ≥ +0.05: it is −0.05). Prediction 2 falsified (whole-brain sign positive: −0.014). Prediction 3 falsified (top-four p rises with learning: it falls on both). The falsification condition is met, so: the walk explains none of E159's 17x, and E160's sign was not the ruler. The greedy walk alone doubles the swarm's p over forty rounds on the mushroom body (0.077 → 0.153) — the circuit's initial landscape correlates with the ladder — and lessons make it climb less, on both substrates, measured in the units the search is paid in. Finds in E159 therefore do not come from the swarm rising; they come from the round's top four being verified out of a swarm that is not improving. Two learning rules were on at once — the core's KC→MBON plasticity and the readout's delta rule — and this cannot say which one hurts.

Second reading, stated so it is tested rather than absorbed: forty rounds is ten times fewer lessons than E159's two hundred, and the first two lessons teach nothing; a rule that overshoots early and settles late would look exactly like this. E165 runs forty rounds with the compartment readout to remove one rule; a two-hundred-round ladder-p pass is the follow-up if E165 clears the core.