With lessons switched on, do the whole-brain walkers climb higher?

E160 — E155's walk statistics, re-run under the learned LAL + DNa + MBON readout

E155 measured the walk under the MBON compartment readout: the whole-brain walkers accepted as often (x0.95) but relocated half again as much (x1.49), climbed a quarter less (x0.76), and forty rounds of lessons moved their climb by +0.03 sd against +0.27 on the mushroom body. E159 then changed one thing — the readout — and the whole brain went from 0.15 to 2.61 AUC_Q. This asks which of E155's numbers the readout changed. Same protocol: 12 elements, 24 walkers x 10 moves x 40 rounds, step 0.05, lr 0.1, seed 0, learning off then on with the rung-0 reward; FORAGER_READOUT's construction applied inside the script.

Predicted:

1. The learning contribution on the whole brain rises from +0.03 sd to at least +0.15, i.e. more than half of the mushroom body's +0.27 — lessons now reach the readout that steers the walk, which is what E159's 17x says must have happened.
2. Relocations on the whole brain fall from x1.49 toward parity (below x1.2): the learned readout smooths the landscape's gradient direction where the fixed compartment readout, fed by a shortcut it could not weigh, made headings fail.
3. The mushroom body's learning contribution is within 30% of E155's +0.27 under its learned MBON readout — the control moves little, matching E159's 1.48 → 2.06 inside error.
4. Acceptance stays within 20% on both, as before; greedy acceptance is scale-free.

Falsified if the whole-brain learning contribution stays below +0.08 sd — then E159's gain came from the starting landscape the learned readout produces (the initial MBON valences read alongside silent LAL/DNa nodes), not from lessons, and the mechanism is a better prior rather than better learning; that would send the next hour to saving and reading the readout weights rather than to step 2b.

E160 result — falsified on three of four, and the pre-registered branch opens.

learned LAL+DNa+MBON readout      mushroom (off → on)      whole brain (off → on)
accept                            0.923 → 0.930            0.878 → 0.867     (x0.93)
relocations / round               2.78 → 2.50              4.13 → 4.45       (x1.78 on)
travel / walker / round           0.317 → 0.294            0.406 → 0.418
climb over 40 rounds, sd units    2.44 → 2.90  (+0.47)     1.85 → 1.28  (−0.58)
E155 for comparison, compartment readout:  +0.27                +0.03

Prediction 1 falsified: the whole brain's learning contribution is −0.58 sd, not ≥ +0.15 — with lessons on, the walkers climb less than with them off. Prediction 2 falsified: relocations moved away from parity (x1.49 → x1.78). Prediction 3 falsified in the other direction: the mushroom body's contribution rose to +0.47, outside the ±30% I allowed around E155's +0.27. Prediction 4 confirmed: acceptance x0.93, scale-free as ever. The learning-off passes reproduce E155 to every digit, as the identity test in tests/test_readout.py says they must.

The branch that opens is the one I pre-registered: E159's 17x did not come from lessons steering the whole-brain walk better, at least not inside forty rounds. Before reading more into the sign, one caveat against my own instrument: score_rise_sd divides the walkers' mean score by a landscape sd measured before any lesson, and a learned readout rescales the score as |w| grows, so climb "in sd units" is not scale-free under this readout — the mushroom body's jump from +0.27 to +0.47 with no other change is the same warning. Two readings survive it: (i) the readout's gain lives in which proposals are chosen for verification (the round's top four), not in how far walkers climb; (ii) forty rounds is too few — E159 ran two hundred, and the first two lessons teach nothing (E159's unit-test observation). Neither is separable with this metric.

Next measurement, scale-free: the same four passes, recording per round the ladder's own rung-0 verdict (Verifier.screen p) at every walker's position — where the swarm actually stands, in the units the search is paid in — instead of the circuit's score. Prediction goes in before it runs. E161's saved readout (running) decides the other half: whether LAL/DNa took weight at all.