Does a second learning site in the brain's navigation centre change the search?

E166 — Task 0, step 3: a second plastic site, in the fan-shaped body

The mushroom body is the fly's fast associative learner; the fan-shaped body (FB) is its navigation centre, where dopaminergic tangential cells shape hΔ/vΔ and the PFL steering output (E156: FB 602, hΔ 189, vΔ 430, PFL 50; PFL is fed 35% by FB; 20 FB dopamine cells). Every lesson so far has landed on one site, KC→MBON. Step 3 adds a second: rung-1 verdicts — the ordering temperature's margin to the window — delivered as FB-tangential dopamine, gating plasticity at the synapses those cells innervate (FB → hΔ/vΔ → PFL), with PFL added to the readout. The rule is the same three-factor rule as KC→MBON, selected by transmitter and family from Atlas, not by name.

Predicted, when it runs (E159's protocol, broadcast input, both plastic sites):

1. Rung-1 verdicts move PFL: the learned readout's share on PFL is above 2% after two hundred rounds — a site that receives its own reward takes weight, as LAL did in E163.
2. The two sites do not fight: the KC→MBON site's contribution (share on MBON, and the swarm's ladder-p with only that site on) is within 20% of its E163 value.
3. AUC_Q does not fall below E163's, and rises if rung-1 margins carry information the rung-0 reward does not — which E142 says they do (the gate under-reports ordering).

Falsified if PFL takes under 0.5% of |w| after two hundred rounds — then either the FB dopamine cells do not reach PFL in this asset's graph at the depth the recurrence runs (three steps), or rung-1 verdicts are too sparse (one per verified proposal) to teach a site that only sees them; the E156 edge counts say the path exists, so sparsity would be the suspect and the repair a rung-1 verdict on every proposal, not only the verified ones.

E166 gate (build only, run queued): 255 passed, 2 skipped, 263 warnings in 820.58s (0:13:40) — tests/test_plastic_site.py adds 3.

E166: smoke passed — site: FB->hDelta/vDelta (602->619) gated by 20 FB dopamine cells; PFL (50) added to the readout — and the full run started from it.

E166 interim (seed 0 of 2, seed 1 running). Readout share after 799 lessons: LAL 4.28%, DNa 0.72%, MBON 94.97%, PFL 0.03%. Prediction "PFL > 2%" not met; the stated falsification bar "< 0.5%" is met on seed 0. The FB site is learning (it is taught on every proposal with an ordering verdict) but the readout does not come to rely on PFL. Final verdict waits for seed 1 and the AUC_Q/finds line.

E166 result — a null, and not the kind the prediction was about. Whole brain, broadcast input, learned readout, plus the FB→hΔ/vΔ dopamine-gated site: AUC_Q 11.57 ± 8.87, 29 ± 17 finds, −111 meV — identical to E163 to every digit, and the seed-1 readout dump is E163's (LAL 2.361%, DNa 0.383%, 799 lessons). PFL share 0.03% / 0.008%: the "< 0.5%" falsification bar is met on both seeds, but the reason is not that the readout ignored PFL — it is that the site changed nothing at all. Two seeded runs cannot coincide otherwise. The site was taught (E167's smoke shows fb lessons arriving: 1 of 4 verifications carried an ordering verdict), so its updates were exact zeros. The only zero in the rule is the eligibility, cache["pre"] at the FB source cells.

Prediction for the diagnosis, written before it runs: under the afferent broadcast with three recurrent steps, the FB cells' pre-state is exactly 0 (they sit deeper than three synapses from any afferent), while LAL and DN carry non-zero state (E167's smoke already shows their gains varying: step 0.026–0.075, patience 3.2–8.9). If FB is non-zero, this explanation is wrong and the null needs another cause (the readout's PFL slice, or an operator that is never rebuilt) — that branch is opened and stated as such.

E166 diagnosis (E168 measurement, runs/e168_fb_state.json). Mean |pre| per population under the broadcast head, four random compositions:

fan_shaped_body  1.0e-4  (13% of cells non-zero)     hdelta  0 exactly     vdelta  1.0e-5 (1%)
LAL              2.2e-3  (71%)      descending  3.8e-3 (54%)      PFL  2.4e-4 (66%)
ring             1.6e-6  (1%)       kenyon      7.4e-7 (0.4%)     MBON 3.1e-4 (14%)

The prediction "exactly 0" is falsified in its letter: the FB is not silent, it is nearly silent — 13% of its cells at 1e-4. The substance stands: the eligibility at the FB is 20–40× smaller than at LAL or DN, so a lesson with lr 0.1 and a standardised error of order one moves an FB→hΔ/vΔ synapse by ~1e-5, and E166's 799 lessons could not move the walk out of E163's trajectory. The FB site is inert on this head — not because the rule is wrong but because the broadcast does not reach the fan-shaped body in three steps. This is E157's question again: the whole brain is driven only as deep as the recurrence carries the input, and the central complex sits below that depth. Branch opened: either more recurrent steps (the FB is ≥4 synapses from the afferents) or an input that reaches it (compass_input is already in the afferent set and still leaves ring cells at 1.6e-6) — that is a measured statement about how much input the head needs, and it is the next experiment, not tonight's.