Twentieth check: the beginner guide re-tested, and a strengthened scrambled brain walked in the body

Re-test, 19 October 2026: 6 of 6 tests, 82.0 Hz bit-exact.

Our own test · pre-registered 19 Oct 2026, 09:33 UTC · Linux x86_64, Python 3.11.2, 2 CPUs, no compiler, no GPU

Pins: unchanged; checked against PyPI on 19 Oct 2026 (latest: Brian2 2.10.1, numpy 2.5.4, pandas 3.0.6, mujoco 3.16.0; our pins stay at 2.9.0, 2.3.5, 2.3.3 and 3.9.0), and installed cleanly in the re-test of 19 Oct; next re-test due 2 Nov 2026. For Python 3.11 the newest releases are Brian2 2.9.0 (our pin) and numpy 2.4.6; Brian2 2.10.1 and numpy 2.5.4 need Python 3.12 or newer.

The short command: On 19 Oct 2026 the short requirements.txt command picked 5 newer helper packages than our exact lock (iniconfig 2.3.0 → 2.3.1, MarkupSafe 3.0.3 → 3.0.4, pluggy 1.6.0 → 1.7.0, pytz 2026.4 → 2026.5, tzdata 2026.4 → 2026.5); the six main packages are pinned and did not change. Use requirements.lock to get exactly our environment. The resolver also listed setuptools 84.0.0, which a beginner already has from get-pip, so it is not counted.

Quick look: Quick look (200 ms trials), run on 19 Oct 2026 in 1.7 min: sugar 80.0 Hz → extend proboscis; shuffled wiring 0.0 Hz → stay still. Shorter trials are noisier than the full run. The quick-look command in the guide is now labelled tested.

Not re-run at full length on 19 Oct: the 20-trial run_demo.py and the "change one thing" experiments (their full results are from 28 Sep 2026; the quick look ran each condition once with 200 ms trials). Next re-test due 2 Nov 2026. The guide: Build your own.

The strengthened scrambled brain with a steering signal, walked in the body

Our own test · pre-registered 19 Oct 2026, 09:37 UTC · simulated body (FlyGym) driven through our own mapping · one run · not a real fly

Why. On 18 Oct we strengthened 4 scrambled brains at 30 spikes/s until their spikes outside the stimulated input matched the real brain's. None sent a turn command, but one (right eye, shuffle 4, connections ×2.75) sent a small steering signal: the far-side turn neuron fired at 1 Hz in one 100 ms bin. We had not walked it in the body. Our expectation, written down before the walk, was that one such bin would change the heading by a few degrees at most.

Checks before the walk. The steering drive was regenerated from the saved brain trial and matched its saved fingerprint; the no-brain floor run was repeated and matched our 8 Oct floor exactly (0 of 700 trajectory values differ, heading after onset −1.49°).

Result. The heading after onset was −0.10°, so the turn beyond the no-brain floor is +1.39° away from the shadow, under our 10° rule. The fly walked 11.5 mm forward after onset (16.8 mm in all); none of the drive was clipped. The real brain at the same input (right eye, 30 spikes/s, trial seed 2) turned it 16.78° beyond the floor in our 8 Oct body run.

Pre-registered sentence: 19 Oct: walked in the body, the one matched scrambled brain with a small steering signal changed the fly's heading by about 1.4 degrees away from the shadow compared with no brain input, under our 10-degree turn rule; the real brain, with the same input, turned it away by about 16.8 degrees (one run each). This body run turns the brain's firing rates into steering through our own hand-made mapping; it is not a separate test of the wiring.

The 30 spikes/s closing is unchanged: As at 80 spikes/s, none of the scrambled brains we ran at 30 spikes/s sent a turn command, including those strengthened to about the real brain's activity outside the input.

The panel below adds this run to the body-strength chart as a filled triangle on the right eye.

10 new entries and 3 new ledger rows

• MaleCNS Cube A (ledger: no difference): a recurrent network on a 256-node MaleCNS v1.0 subgraph, with trainable connection strengths, estimates how many face turns a Rubik's cube state is from solved. Test error (lower is better; 3 seeds): real 0.3509, degree-preserving rewired 0.3380, frozen strengths 0.5275, parameter-matched MLP 0.3245; all four solve 12 of 12 search tests. A course project; the author's numbers.
• QedSDL Endpoint A (ledger: mixed): the FlyWire v783 wiring from projection neurons to Kenyon cells as the fixed input layer of a novelty detector in simulated materials-lab streams. AUROC over 20 seeds: real 0.9505, degree-preserving shuffle 0.9505, random wiring 0.9391 (means recomputed by us from the committed per-seed files). Not a fly-behaviour model.
• flybrain (foraging economy) A (ledger: worse): a MaleCNS v1.0 network (166,700 neurons) runs flies in an evolving foraging economy. Final-generation mean energy over 3 seeds: real 360.27, weight-permuted 742.93, zeroed weights −151.17, coin-flip decision −110.33 (recomputed by us from the committed ledgers). The simulation kernel is not in the repository; a single-author preprint.
• 3dfly B: a MaleCNS v1.0 rate network steers a flybody fly in MuJoCo while it maps a 3-D scene; a degree-preserving shuffle is implemented in its acceptance script, but no result is committed.
• FlyWire LM B: the frozen FlyWire v783 graph as the recurrent core of a byte-level language model with trained input and output layers; the author states that the results do not establish an advantage of the fly wiring over other graphs, and there is no other-graph comparison.
• ChatFlyPT B: a MaleCNS v1.0 brain-and-nerve-cord simulation behind a chat interface; trained decoders turn its activity into labels, text and NeuroMechFly movement. Its controls shuffle the decoders' inputs, not the wiring.
• FastFly B: a FlyWire v783 graph atlas, and a whole-brain simulation that steers a 2-D walker through hand-drawn mazes; no wiring control.
• A Virtual Fly Brain course n/a: a Chinese-language course on Virtual Fly Brain, with a recount of the Nern et al. 2025 visual-system cell-type census from public data. No simulation.
• OrigoZero "Fruit Fly" U: the page says all 164,587 neurons of the MaleCNS v1.0 connectome run as leaky integrate-and-fire units on the GPU and drive a fly in an arena; it credits Berg et al. 2026 for the data and Shiu et al. 2024 for the model constants. No code is linked. The page responded on 19 Oct 2026; not loaded in a browser by us.
• Fly Connectome Explorer (flyproject.io) U: the page calls itself "a real fruit fly brain, one circuit at a time"; its own script contains leaky integrate-and-fire code next to hand-written behaviour states. No repository, licence or author is given. The page responded on 19 Oct 2026; not loaded in a browser by us.

One more lead, DeevlerPlays599/FlyBrain, is a re-host of Lulzx/fly-brain with identical files, so it is counted under the existing fly-brain entry. Two of these repositories had new commits since they were first listed (3dfly and DeevlerPlays599/FlyBrain); both were read at their new commits.

The ledger now has 65 studies, 56 with a wiring null: helps 25, no difference 12, worse 5, mixed 12, not yet scored 2. No scoring rule changed.

What each activity number counts. Ten comparison arms carry an activity ratio (display only). Each now says what it counts: eight count spikes over the whole brain, the stimulated input included (our own tests), and two count network units (fly-connectome-reservoir). The gain-matched arm of our looming test counts the whole brain (0.9864× the real brain); it was matched on the spikes outside the stimulated input, which were 0.98× the real brain's, and the ledger now shows both (the ledger).

Corrections, verdicts and tracking

• Author fields. fly67 and ChessFly now name the author as each repository's LICENSE does: fly67 "Ruben Gexhaw (GitHub: gexhaw, formerly znatgost)", ChessFly "Ruben Nugmanov (GitHub: gexhaw, formerly znatgost)".
• fly67 size. About 33 MB (measured on our 18 Oct 2026 load; the README says 31-33 MB gzipped).
• The Odyssey-3 claim (the post): No method, code or write-up was published as of 19 Oct 2026. Odyssey's newest post is about Odyssey-3 itself (8 Oct) and does not mention a fly. The verdict row stays U.
• Repositories of 15 of the 23 verdict rows re-checked on 19 Oct 2026; none changed since grading. Rows without code are carried.
• All 56 videos available. Links: fruitfly.live still has a TLS certificate error; FlyDoom's repository still answers "not found". Six catalogue repositories have new commits since we graded them; they are recorded and will be re-read later (grades unchanged until then).

What we could not resolve

• One machine. The re-test ran on one Linux machine with Python 3.11.2; macOS, Windows and other Python versions are untested.
• One body run. The walk is one deterministic run of one scrambled graph at one trial seed; it describes this drive, not scrambled brains in general.
• Our mapping. The body turns the brain's firing rates into steering through our own hand-made mapping, and we hold the walking drive; the walk is not a separate test of the wiring.
• Count, not pattern. The strengthened scrambled brains matched the real brain's number of spikes outside the input, not where those spikes happen.
• Unverified pages. The two new U rows are the pages' own claims; we did not load them in a browser or read their full code.

Earlier: report of 18 Oct 2026 (the scrambled brains at 30 spikes/s, strengthened) and report of 5 Oct 2026 (the previous beginner re-test).