Did the learned readout put real weight on the newly read brain regions?

MBON   99.62%      LAL   0.37%      DNa   0.011%      DNa02   1.2e-5

EXPERIMENTS.md · lines 10454–10478

E161 — Where the learned readout puts its weight (E159's unmeasured prediction 2)

E159 changed the readout to a learned linear map over LAL (665) + DNa (32) + MBON (97) and the whole brain went from 0.15 to 2.61 AUC_Q. The run did not save the readout, so whether the gain came from reading LAL, DNa, or simply from learning the MBON weights instead of assuming compartment valence is unknown. stage_b will now write the final readout and its share of |w| by population at the end of every learned-readout run; one whole-brain run, identical to E159's arm, supplies the number.

Predicted:

1. MBON keeps the majority of |w| — above 60% — because under olfactory input E157 measured LAL/DNa at a small fraction of MBON activity, and a delta rule grows weight in proportion to the node's activity at each lesson.
2. LAL carries more than DNa (LAL 665 nodes, fed 9% by MBON directly; DNa 32 nodes, fed through LAL) and DNa02's weight is exactly zero — it never fired under this input (E157).
3. The share on LAL + DNa is nonetheless above 10%: if it were near zero, E159's 17x would be the learned MBON weights alone and reading the extra populations bought nothing — that is the falsification, and it would mean the mushroom-learned control's 2.06 and the whole-brain 2.61 are the same effect measured twice.

Falsified if LAL + DNa carry under 10% of |w| (E159 = learning, not location) or over 50% (the extra populations dominate under an input that barely drives them, which would contradict E157 and need explaining before anything is built on it).

EXPERIMENTS.md · line 10480

E161 gate: 249 passed, 2 skipped, 227 warnings in 865.88s (0:14:25) (no forager/ change this hour; stage_b gained the readout dump).

EXPERIMENTS.md · lines 10544–10551

E161 / E162 — killed under memory pressure, relaunched; no result was lost that existed. The E161 whole-brain run (2.4 GB) and the E162 chain waiting behind it died while the rung-4 pw.x (up to 6.5 GB) was resident and I had started this hour's pytest (2 GB) beside them — the third heavy process was mine, and it was the likely tip. stage_b writes its readout dump and result only at the end, so E161 had nothing to resume; its partial log is kept as runs/e161_wholebrain_laldna.killed.log and the run is repeated, not resumed. Relaunched as one strict chain — E161, then E162, then pytest — so that at most one of my jobs runs beside pw.x. Predictions for both stand as written before the first launch.

EXPERIMENTS.md · lines 10657–10678

E161 result, seed 0 (seed 1 running) — prediction 3 falsified; the location claim in E159 is withdrawn. After 799 lessons the learned readout over 794 nodes carries:

MBON   99.62%      LAL   0.37%      DNa   0.011%      DNa02   1.2e-5

Prediction 1 confirmed (MBON above 60% — it is nearly everything). Prediction 2 confirmed (LAL above DNa; DNa02 at zero to the precision that matters). Prediction 3 falsified: LAL + DNa carry 0.4% of |w|, not more than 10%. The pre-registered branch is the finding: E159's 17x came from learning the MBON readout instead of assuming compartment valence — not from reading LAL or DNa. The mushroom-learned control (2.06 ± 1.68 against compartment 1.48 ± 0.48) said the same in its own noisy way. "Reading the brain where its signal goes" is withdrawn as the mechanism of E159; what E159 showed is that the readout the anatomy hands us is a poor prior and a learned one is a much better one, on either substrate. Why the delta rule cannot put weight on LAL/DNa is E157: under olfactory input they are barely driven, so their eligibility is ~0 at every lesson. The test of location is therefore step 2b — the afferent broadcast that drives them — and not E159. Seed 1 will be appended when its dump lands; the conclusion does not hinge on it.

Δ-check (E164): 3 of 14 anchors finished at this wake-up (TaV_a2.813, TaW_a3.230, VW_a3.067); ~20 min each on four cores beside two other jobs. Collected when all are in. Suite: deferred to the chain's own pytest (E161 → E162 → pytest) — one heavy job beside pw.x at a time, after the kill two hours ago.

EXPERIMENTS.md · lines 10680–10682

E161 result, both seeds — DONE. Seed 1: MBON 99.56%, LAL 0.43%, DNa 0.014%; the arm reproduced E159 exactly (2.61 ± 0.48, 7.0 finds, −90 — same seeds, deterministic). Prediction 3 falsified on both seeds; the withdrawal of E159's location claim stands.