Is the brain's own learning rule the one that makes the walkers do worse?

core rule only                mushroom off → on        whole brain off → on
swarm mean p                 0.153 → 0.107 (−0.046)   0.133 → 0.190 (+0.057)
top-four p                   0.158 → 0.119 (−0.040)   0.076 → 0.053 (−0.023)
E162, both rules:            0.153 → 0.101            0.133 → 0.119

EXPERIMENTS.md · lines 10711–10739

E165 — Which learning rule lowers the ladder-p of the walk: the core's or the readout's?

E162 measured lessons lowering the swarm's rung-0 p over forty rounds on both substrates with the learned readout on. Two rules were active: MushroomBody.learn_head on KC→MBON, and PopulationReadout's delta rule on the readout weights. Same four passes as E162, with --readout mbon — the compartment readout, so only the core learns. Same seeds, rounds, walkers, step, lr.

Predicted: (1) the mushroom body's learning contribution in p is still negative (swarm mean over the last ten rounds lower with learning than without by at least 0.02) — the core rule is the one that hurts, because it is the one E155 already showed relocating walkers more with learning on; (2) the whole brain's is negative too, by less; (3) if instead both contributions turn positive, the readout's delta rule was the culprit and the core is cleared — that is the branch I consider less likely, since the readout starts from the compartment weights and moves slowly.

Falsified if the mushroom contribution is positive by ≥ +0.02 — then the readout rule is what damaged E162's walk, PopulationReadout gets its learning rate cut or its rule changed to a proper least-squares update, and E159's learned-MBON gain has to be re-explained.

Wake-up collection — nothing new decided, everything accounted for. The E161 → E162 → pytest chain finished: 252 passed, 2 skipped (1,393 s under contention; 249 plus the three broadcast-encoder tests). E163's smoke pass ran end to end — encoder: broadcast onto 17,479 afferent nodes, two rounds, no finds, as a two-round smoke should — and the full whole-brain broadcast run (200 x 4 x 2) started from it; prediction on record under E163. E165 waits behind it. The Δ-check stands at 11 of 14 anchors; the six binaries and Hf, Mo, Nb, Ta, Ti are in, and the residual on Ta/V/W needs the V and W elemental cells, which are the last to run — so no partial residual is quoted. The rung-4 pw.x is still converging. No new job was launched this hour: E163 is the one heavy job beside pw.x.

EXPERIMENTS.md · lines 10840–10860

E165 result — mixed, and it separates the two rules. Compartment readout (core KC→MBON learning only), rung-0 p at every walker position, last-ten-round means:

core rule only                mushroom off → on        whole brain off → on
swarm mean p                 0.153 → 0.107 (−0.046)   0.133 → 0.190 (+0.057)
top-four p                   0.158 → 0.119 (−0.040)   0.076 → 0.053 (−0.023)
E162, both rules:            0.153 → 0.101            0.133 → 0.119

Prediction 1 confirmed: on the mushroom body the core rule alone lowers the swarm's p — it is the core rule that hurts there. Prediction 2 falsified: on the whole brain the core rule alone raises the swarm's p by +0.057, and it was the readout's delta rule that pulled E162 down to 0.119. So the two substrates disagree about which rule helps, and the readout rule — which E159 credited with the whole-brain gain — costs the whole-brain walk 0.07 in swarm p while (E161) it also carried that gain. Prediction 3's clean "both positive" branch did not occur; the falsification condition (mushroom contribution ≥ +0.02) is not met. Top-four p falls with learning on both substrates under both readouts — the proposals chosen for verification get worse in ladder units over forty rounds — while E163's two-hundred-round finds are the best any fly arm has had. That tension is recorded as open: either the top-four metric over ten rounds is not what finds are made of, or the first forty rounds are the bad ones. A two-hundred-round ladder-p pass is the experiment that resolves it, and it is not run tonight.