Experiments · E167

Does letting the brain set each walker's stride and patience improve the search?

Withdrawn. The score more than doubled, to 25.9, but a shuffled control matched it: varied strides helped, not the brain's reading.

In the log: octopamine gain and serotonin persistence read from the head (task 0 step 4)

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 10871–10908, lines 10947–10957, lines 10988–11001, lines 11008–11038
exp E167 diagram
What E167 did and how it came out, drawn from this record and the files it names (book/assets/diagrams/exp/E167.svg).

Pre-registration

The pre-registration, as written

E167 interim (seed 0 of 2; seed 1 running) — and a protocol flaw, stated. Readout share after 799 lessons: LAL 3.37%, DNa 0.42%, DN 7.53%, PFL 0.02%, MBON 89.1%. Sites: fb 544 lessons, oa 799, 5ht 0. Gains: step 0.029–0.074 (span 0.045), patience 3.0–8.8.

  • Prediction 3 confirmed: the step modulation is real (span 0.045 ≥ 0.01).
  • Prediction 4 (first half) confirmed: the OA site is taught on every lesson (799).
  • Prediction 1's "never taught" clause fired — but for a reason I wrote into the protocol, not one the head produced: confirm() runs only under --confirm, and the E167 chain does not pass it. The 5-HT site had no teacher. E167 is therefore a test of the state-read gains (step from LAL, patience from DN), and of nothing serotonergic. Withdrawn as a test of serotonin persistence; kept as a test of the gains.
  • Unpredicted: the readout put 7.5% on DN without the 5-HT site ever moving. The delta rule w += lr·err·h favours the nodes with the largest state, and DN carries the largest (E168: 3.8e-3 vs LAL 2.2e-3) — a magnitude effect, not evidence that DN is informative. The E169 DN share must be read with that in mind.

Results

EXPERIMENTS.md · line 10947

E167 amendment, before its results arrive. By the same measurement the OA site (pre = FB) is inert; the 5-HT site (pre = LAL, 71% active) and both gains (read from LAL and DN) are live. E167 therefore tests serotonin persistence plus state-read gains, not octopamine learning. Prediction 1–3 stand as written; a fourth is added: sites[oa].lessons will be ~equal to the readout's lessons and its effect nil — the OA site's |Δw| is below 1e-4 total, so its share of any change is zero.

Housekeeping, stated: the per-seed readout dumps are named by circuit and seed only, so E166 overwrote E163's and the E167 smoke overwrote seed 0 again. E163's seed-1 numbers survive because E166's file is identical to it; E163's seed-0 file is gone (its share was recorded in the E163 entry). Fixed by tagging the filename with FORAGER_TAG from here on.

EXPERIMENTS.md · line 11008

E167 result — falsified upward: the state-read gains more than double the search. Whole brain, broadcast, learned readout, fb+oa+5ht sites, step from LAL state, patience from DN state, 200 × 4 × 2:

arm                                                   AUC_Q            finds          best
E163  same head, constant step 0.05 / patience 6     11.57 +- 8.87    29 +- 17       -111
E167  step and patience read from the head           25.93 +- 11.44   71 +- 15       -136
E136  CEM, 300 rounds                                 13.57 +- 3.79    34.5 +- 8.5    -78

Seeds ≈ 14.5 and 37.4 — both above E163's mean, ×2.2 on AUC_Q, ×2.4 on finds, 25 meV/atom deeper, and for the first time a fly arm is clear of the CEM on every column.

  • Prediction 2 falsified upward. I bounded each seed at 3–20 and called "better" the case of both seeds above 11.57, which I said was not predicted. It happened.
  • Prediction 3 confirmed on both seeds: step 0.029–0.074 and 0.026–0.072 (spans 0.045, 0.046), patience 3.0–8.8 and 3.1–8.9 — the gains are real, walker-specific quantities.
  • Prediction 4 confirmed (first half): OA site taught 799× on both seeds; its effect cannot be isolated here, and by E168 (FB pre-state 1e-4) it is nil.
  • Prediction 1: 5-HT site 0 lessons on both seeds — by protocol (no --confirm), stated in the interim. E167 says nothing about serotonin.
  • Readout: DN 7.5% / 4.1%, LAL 3.4% / 2.1%, DNa 0.42% / 0.25%; fb site 544 / 655 lessons.

What the result does and does not show. The improvement is attributable to the gains, not the sites: the FB and OA sites move synapses by ~1e-5 (E168), the 5-HT site never moved. But two things changed at once relative to E163 — the step and the patience are now walker-specific, and they are read from the head. A swarm with the same spread of steps and patiences assigned at random would separate those. Branch opened: E170, the shuffled control — the same PopulationGain ranges with the z-scores permuted across walkers each call, so each walker's gain is a real head state but the wrong walker's. If E170 matches E167, the gain is heterogeneity, not the head, and "read from the head" is withdrawn as a claim (the mechanism stays, the interpretation goes). If E170 falls back toward E163, the head's state at the walker's own composition carries the information.

The full record

This entry is written in 4 separate places in the log, shown here in log order.

EXPERIMENTS.md · lines 10871–10908

E167 — octopamine gain and serotonin persistence read from the head (task 0 step 4)

What the wiring allowed, which is not what the plan said. The plan's table put serotonin onto ER, FB and SMP. In this asset serotonin cells innervate no ring neuron and no FB cell; they reach 669 of 1,342 descending neurons. Octopamine reaches every ring neuron (282/282), 637 of 665 LAL, 462 of 602 FB, 46 of 50 PFL and all 32 DNa. The head's plan is therefore corrected by its own anatomy: gain = octopamine-gated FB→LAL (1,008 edges, 994 gated), persistence = serotonin-gated LAL→DN (10,658 edges, 4,989 gated). The step-4 row in OVERNIGHT.md is withdrawn as written and replaced by this.

Mechanism (forager/brain/modulation.py, Forager(modulator=), WithSites per-site rewards + teach). Each walker's step is 0.05·(1 + 0.5·tanh z_LAL) and its patience 6·(1 + 0.5·tanh z_DN), z the standardised mean state of the gated population at the walker's composition (running traces over every visit). The OA site is taught by the rung-0 reward the core learns from — arousal follows what pays; stage_b runs no rung-2 kinetics and this is stated, not hidden. The 5-HT site is taught only when confirm() returns, with 2p−1, so persistence is written by the decisive rung and by nothing else. Literature for the two roles: octopamine as locomotor/sensory gain with arousal (Suver, Mamiya & Dickinson 2012; Busch 2009); serotonin as willingness to wait on a course (Miyazaki 2014, 2018; Sitaraman 2008 for central-complex place memory). Nothing here is a knob: both are states of the head, moved by the head's own lessons.

Protocol. E163's arm (whole brain, broadcast input, LAL+DNa+MBON+PFL+DN readout, FORAGER_SITES=fb,oa,5ht FORAGER_MODULATE=1), 200 × 4 × 2. Control: E163 (11.57 ± 8.87 / 29 ± 17 / −111) and E166's fb-only arm when it lands.

Predictions, before the run.

  1. The 5-HT site is taught rarely — confirm() fires only on what the free rung likes — so sites[5ht].lessons will be < 10% of the readout's lessons on both seeds. If it is never taught (0 lessons on a seed), persistence never moved on that seed and the result is a test of the OA gain alone; say so.
  2. AUC_Q within E163's range: between 3 and 20 on each seed. Falsified if either seed is below E159's 2.61 — then the modulated walk is worse than the unmodulated one and step 4 is withdrawn on this wiring.
  3. Step modulation is real, not a constant: the dumped step_last spans ≥ 0.01 across the 24 walkers (base 0.05, span 0.5 permits 0.025–0.075). If it spans < 0.005 the LAL state does not separate compositions and the gain is a no-op — the mechanism is then withdrawn, not the idea.
EXPERIMENTS.md · lines 10947–10957

E167 amendment, before its results arrive. By the same measurement the OA site (pre = FB) is inert; the 5-HT site (pre = LAL, 71% active) and both gains (read from LAL and DN) are live. E167 therefore tests serotonin persistence plus state-read gains, not octopamine learning. Prediction 1–3 stand as written; a fourth is added: sites[oa].lessons will be ~equal to the readout's lessons and its effect nil — the OA site's |Δw| is below 1e-4 total, so its share of any change is zero.

Housekeeping, stated: the per-seed readout dumps are named by circuit and seed only, so E166 overwrote E163's and the E167 smoke overwrote seed 0 again. E163's seed-1 numbers survive because E166's file is identical to it; E163's seed-0 file is gone (its share was recorded in the E163 entry). Fixed by tagging the filename with FORAGER_TAG from here on.

EXPERIMENTS.md · lines 10988–11001

E167 interim (seed 0 of 2; seed 1 running) — and a protocol flaw, stated. Readout share after 799 lessons: LAL 3.37%, DNa 0.42%, DN 7.53%, PFL 0.02%, MBON 89.1%. Sites: fb 544 lessons, oa 799, 5ht 0. Gains: step 0.029–0.074 (span 0.045), patience 3.0–8.8.

  • Prediction 3 confirmed: the step modulation is real (span 0.045 ≥ 0.01).
  • Prediction 4 (first half) confirmed: the OA site is taught on every lesson (799).
  • Prediction 1's "never taught" clause fired — but for a reason I wrote into the protocol, not one the head produced: confirm() runs only under --confirm, and the E167 chain does not pass it. The 5-HT site had no teacher. E167 is therefore a test of the state-read gains (step from LAL, patience from DN), and of nothing serotonergic. Withdrawn as a test of serotonin persistence; kept as a test of the gains.
  • Unpredicted: the readout put 7.5% on DN without the 5-HT site ever moving. The delta rule w += lr·err·h favours the nodes with the largest state, and DN carries the largest (E168: 3.8e-3 vs LAL 2.2e-3) — a magnitude effect, not evidence that DN is informative. The E169 DN share must be read with that in mind.
EXPERIMENTS.md · lines 11008–11038

E167 result — falsified upward: the state-read gains more than double the search. Whole brain, broadcast, learned readout, fb+oa+5ht sites, step from LAL state, patience from DN state, 200 × 4 × 2:

arm                                                   AUC_Q            finds          best
E163  same head, constant step 0.05 / patience 6     11.57 +- 8.87    29 +- 17       -111
E167  step and patience read from the head           25.93 +- 11.44   71 +- 15       -136
E136  CEM, 300 rounds                                 13.57 +- 3.79    34.5 +- 8.5    -78

Seeds ≈ 14.5 and 37.4 — both above E163's mean, ×2.2 on AUC_Q, ×2.4 on finds, 25 meV/atom deeper, and for the first time a fly arm is clear of the CEM on every column.

  • Prediction 2 falsified upward. I bounded each seed at 3–20 and called "better" the case of both seeds above 11.57, which I said was not predicted. It happened.
  • Prediction 3 confirmed on both seeds: step 0.029–0.074 and 0.026–0.072 (spans 0.045, 0.046), patience 3.0–8.8 and 3.1–8.9 — the gains are real, walker-specific quantities.
  • Prediction 4 confirmed (first half): OA site taught 799× on both seeds; its effect cannot be isolated here, and by E168 (FB pre-state 1e-4) it is nil.
  • Prediction 1: 5-HT site 0 lessons on both seeds — by protocol (no --confirm), stated in the interim. E167 says nothing about serotonin.
  • Readout: DN 7.5% / 4.1%, LAL 3.4% / 2.1%, DNa 0.42% / 0.25%; fb site 544 / 655 lessons.

What the result does and does not show. The improvement is attributable to the gains, not the sites: the FB and OA sites move synapses by ~1e-5 (E168), the 5-HT site never moved. But two things changed at once relative to E163 — the step and the patience are now walker-specific, and they are read from the head. A swarm with the same spread of steps and patiences assigned at random would separate those. Branch opened: E170, the shuffled control — the same PopulationGain ranges with the z-scores permuted across walkers each call, so each walker's gain is a real head state but the wrong walker's. If E170 matches E167, the gain is heterogeneity, not the head, and "read from the head" is withdrawn as a claim (the mechanism stays, the interpretation goes). If E170 falls back toward E163, the head's state at the walker's own composition carries the information.

Related entries

Built with PRISMWebsite and visualizations made using Claude