Experiments · E171b

Does switching on the navigation centre's learning site make the whole-brain search better?

Withdrawn. The score reached 39.96, but a later run without the site matched it exactly; the extra propagation step did the work.

In the log: seed 0 interim

withdrawnDate not stated in the log; it was written between the commit of 2026-09-16 19:06 and the first commit that contains it, 2026-09-19 08:35generator · fly brain0 predictions · 2 result paragraphsEXPERIMENTS.md lines 11363–11366, lines 11393–11414

Checked · 2 discrepancies

Files checked (5)

runs/e171b_wholebrain_fbscaled.log
runs/readout_fly_malecns_full_e171b_seed0.json
runs/readout_fly_malecns_full_e171b_seed1.json
runs/e171b_seed2.log
runs/e178_seed2.log

Numbers confirmed (4)

claim: AUC_Q 39.96 +- 7.06, 82 +- 1 finds, best -137 (seeds 32.90 and 47.02 per E184b); source: runs/e171b_wholebrain_fbscaled.log; ok: true
claim: fb 563 / 646 lessons, |dw| 269.8 / 224.9; PFL share 0.19 % / 0.13 %; source: runs/readout_fly_malecns_full_e171b_seed{0,1}.json; ok: true
claim: step span 0.049, patience 3.0-8.8; source: modulation field, runs/readout_fly_malecns_full_e171b_seed0.json; ok: true
claim: later third seed with and without the FB site: both 47.04 AUC_Q, 102 finds, best -137; source: runs/e171b_seed2.log, runs/e178_seed2.log; ok: true

Discrepancies

Withdrawn later (N6, EXPERIMENTS.md l. 12378-12389): with and without the FB site the third seed is bit-identical, E171b's standing as best arm on account of the FB site is withdrawn, and the fourth propagation step is the whole of the gain.
Later context (EXPERIMENTS.md l. 12753, 2026-09-19): the reward was found coupled to process history (fixed in 247ed9d); the log says only within-run comparisons survive from before the fix, so the cross-run AUC_Q comparisons here carry that caveat.

Status corrected from the catalog's recorded to withdrawn, the status shown above

The arm and site numbers reproduce from the log and dumps; the later bit-identical seed-2 pair withdrew the FB-site reading.

From runs/site_data/exp_checks/E171b.json; the log itself is not edited.

Results

EXPERIMENTS.md · line 11363

E171b seed 0 interim (steps = 4, FB site connected): fb 563 lessons, |Δw| 269.8 (×15 E170's 18.3 — the lr scaling and the larger step-4 eligibility together); oa 799, 3.7; 5ht 0 (no --confirm, as E167). PFL readout share 0.19% (E166: 0.03%). Gains: step span 0.049, patience 3.0–8.8 — prediction 4 confirmed. AUC_Q waits on seed 1.

EXPERIMENTS.md · line 11393

E171b result — the best arm yet, and a confound that needs one control. Steps = 4 (the central complex reached), FB site lr 2.0, E167's protocol otherwise, 200 × 4 × 2:

arm                                       AUC_Q            finds         best
E167  steps 3, gains from own state       25.93 +- 11.44   71 +- 15      -136
E170  steps 3, gains shuffled             31.54 +- 18.81   68 +- 24      -137
E171b steps 4, FB site connected          39.96 +-  7.06   82 +-  1      -137

Seeds ≈ 33 and 47; finds 81 and 83 — the tightest pair any fly arm has produced.

Prediction 1 (fb |Δw| > 1e-2): confirmed and uninformative, as stated in advance (270 and the seed-1 value in the dump).
Prediction 2 (not identical to E167): confirmed — the FB site's synapses now reach the readout, and the run differs.
Prediction 3 (mean AUC_Q in 14–38): falsified upward at 39.96, by the same 2-seed margin that has made every "within E167's range" prediction too narrow. The prediction was calibrated on E167's spread; the result's spread is half that.
Prediction 4 (gain spans within 20% of E167's): confirmed (0.049; patience 3.0–8.8). The confound. Two things changed against E167: the FB site became causally connected, and every node's state changed, because a fourth recurrent step alters the whole propagation — the gains' inputs (LAL, DN), the readout's inputs, the KC→MBON site's eligibility. The improvement cannot be assigned to the FB site without a run at steps = 4 with the FB site absent.

The full record

This entry is written in 2 separate places in the log, shown here in log order.

EXPERIMENTS.md · lines 11363–11366

EXPERIMENTS.md · lines 11393–11414

E171b result — the best arm yet, and a confound that needs one control. Steps = 4 (the central complex reached), FB site lr 2.0, E167's protocol otherwise, 200 × 4 × 2:

arm                                       AUC_Q            finds         best
E167  steps 3, gains from own state       25.93 +- 11.44   71 +- 15      -136
E170  steps 3, gains shuffled             31.54 +- 18.81   68 +- 24      -137
E171b steps 4, FB site connected          39.96 +-  7.06   82 +-  1      -137

Seeds ≈ 33 and 47; finds 81 and 83 — the tightest pair any fly arm has produced.

Prediction 1 (fb |Δw| > 1e-2): confirmed and uninformative, as stated in advance (270 and the seed-1 value in the dump).
Prediction 2 (not identical to E167): confirmed — the FB site's synapses now reach the readout, and the run differs.
Prediction 3 (mean AUC_Q in 14–38): falsified upward at 39.96, by the same 2-seed margin that has made every "within E167's range" prediction too narrow. The prediction was calibrated on E167's spread; the result's spread is half that.
Prediction 4 (gain spans within 20% of E167's): confirmed (0.049; patience 3.0–8.8). The confound. Two things changed against E167: the FB site became causally connected, and every node's state changed, because a fourth recurrent step alters the whole propagation — the gains' inputs (LAL, DN), the readout's inputs, the KC→MBON site's eligibility. The improvement cannot be assigned to the FB site without a run at steps = 4 with the FB site absent.

Related entries

E170 — the shuffled control for E167
E167 — octopamine gain and serotonin persistence read from the head (task 0 step 4)
E166 — Task 0, step 3: a second plastic site, in the fan-shaped body