Research report · 31 October 2026 · twenty-eighth check
Twenty-eighth check: Fly-NAF tests its wiring, two new entries, and the walking model checked for a re-run
We checked whether the new Nature walking-circuit model can be re-run with shuffled wiring on our small server, re-read three projects with new results (one now compares its real wiring with shuffled copies), graded two new projects with their own scrambled-wiring tests and loaded fly67 again.
Short answer
The walking-circuit model was not re-run with shuffled wiring: it would not fit in our test space. We checked whether the authors' walking model can be re-run with shuffled wiring on an ordinary 2-processor server without a GPU (31 Oct 2026): its software would need an estimated 620 MB once installed, plus 86 MB of wiring data, more than our 600 MB test space. The test was not run. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. Whether the real wiring is needed for the simulated walking rhythm stays not tested; the earlier circuit count is a pattern in the wiring, not a test of walking.
Sometimes: fly wiring beat scrambled wiring in 36 of 94 studies, mostly reflex, sensory and steering circuits; in trained, reservoir, ML and our walking tests it tied or lost (37). Fly-NAF adds a "helps" that we rate unfair, because its game constants were tuned on the real wiring; the two new projects both read "worse".
fly67: runs again on 31 Oct. It downloaded about 33 MB of brain data, identified by their web address, a method we fixed before this check.
The walking-circuit model: why the shuffled-wiring test did not run
Code at commit 10e7661 (MIT licence, stated in the README), package sizes from the Python package index and data sizes from the repository · checked 31 Oct 2026 · nothing installed, nothing simulated
On 30 Oct we read the shuffle test in the code of Pugliese et al. (Nature, 28 Oct 2026): it counts circuit patterns in the wiring, not walking. This time we checked, with limits written down before looking, whether the authors' own simulation could be run on our ordinary 2-processor server without a GPU, once with the real front-leg wiring and then with the authors' shuffle.
- Needed (estimate): about 706 MB = software about 620 MB once installed + wiring data about 86 MB
- Our test space: 600 MB (dashed line in both bars)
What we found. The model is a firing-rate simulation written with the JAX and diffrax libraries (486 files; the main simulation file has 2,038 lines). To load it, Python also needs plotting and storage libraries, because the package imports them on start. The packages to download add up to about 248 MB (the largest: jaxlib 83 MB, llvmlite 60 MB, scipy 35 MB); by our fixed rule of thumb that is about 620 MB once installed. The wiring data adds about 86 MB (an 85.0 MB connection matrix and a 0.7 MB neuron table). Together, about 706 MB, above the 600 MB space we allow for a test, by about 106 MB. So, by the rule set in advance, we installed nothing and ran nothing, and the entry stays not tested. No packages were missing, and no GPU or account would have been needed.
Which shuffle the simulation uses. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. In more detail: the simulation's shuffle reorders which neuron receives which set of inputs, within each of five groups (excitatory and inhibitory descending neurons, excitatory and inhibitory nerve-cord neurons, and motor neurons), so each neuron keeps its sign and the weights it sends, and each set of incoming connections stays whole. The circuit-pattern count of 30 Oct shuffled each nerve-cord neuron's outgoing connections separately, which keeps how many connections each neuron sends but not how many each receives.
Not known. The 620 MB figure is an estimate from a fixed factor, not a measurement; the real install could be smaller or larger. Leaving out the plotting and storage parts might bring the total to about 485 MB, but that changes which parts of the authors' code are loaded, so it would have to be planned before any attempt; it was not tried. How long one simulation would take on our server was not measured (the authors' tutorial mentions about 30 to 60 seconds per simulation on a desktop at relaxed settings).
Catalogue entry: pugliese-2026-walking-cpg, grade A, peer-reviewed. A nerve-cord study, not a whole-brain or body simulation. The claim on Is it real?
3 new ledger rows: one re-read and two new entries
The authors' committed numbers; nothing re-run by us · scored only by our fixed rules, which did not change
| Project | Effect | Real wiring | Scrambled or other wiring | Limits |
|---|---|---|---|---|
| Fly-NAF (re-read) B | Helps | 7 of 8 game threats answered on the threat's side first (one replayed night, 2,618 frames) | 0, 3 and 3 of 8 for three shuffles that keep each neuron's in- and out-degree and sent weights (mean 2.0 ± 1.73, computed by us); eyes swapped 0 of 8 | Rated unfair: the game constants were calibrated on the real wiring, so the shuffles ran with settings made for another network. Weak: one night, open loop, raw logs not committed, no rule-based baseline without a brain. |
| Fly Hearing Lab (new) A | Worse | Male rhythm tuning 0.70 Hz (95% interval 0.68 to 0.72; 5 simulation seeds) | 2.66 Hz (2.60 to 2.71) with degree-preserving shuffled wiring (one shuffle seed) | Partly fair, weak. The author calls every effect not robust; that higher tuning is the better direction is our reading of the author's hypothesis. |
| DrosoCode (new) A | Worse | Next-token accuracy 12.6% after 1,200 lessons (one run) | 18.1% for a rewired copy of the same size (one run); a majority-token baseline scores 9.2% | Partly fair, weak: one run each, no uncertainty; the rewiring keeps each target's number of inputs, not each source's outputs. |
Fly-NAF (re-read 31 Oct): the author now tests the wiring. On one replayed game night the real wiring answered 7 of 8 threats on the threat's side first; three shuffled brains answered 0, 3 and 3. The game constants were tuned on the real wiring, so the shuffled brains ran with settings made for another network. Grade B kept. In a second test, after a sighting the real wiring looked back at that side 24 of 24 times, two shuffles 13 of 24 each and a brain with all synapses removed 4 of 24.
Fly Hearing Lab (new, graded A): in two simulated fly brains the song circuit did not prefer the species' song, and the male rhythm-tuning effect was larger with shuffled wiring (2.66 vs 0.70 Hz), by the author's committed results.
DrosoCode (new, graded A): a fly mushroom body from FlyWire learning to predict code tokens; a rewired copy of the same size learned better than the real wiring (18.1% vs 12.6%, one run each), by the author's committed results.
Ledger now: helps 36, no difference 26, worse 11, mixed 17, not yet scored 4 (each row on Does fly wiring help?).
First look at 21 code projects found by our searches: 2 graded (Fly Hearing Lab and DrosoCode), 14 left out because they have no wiring comparison or their shuffle does not touch the wiring, 4 listed for the next check and 1 not found.
Two more re-reads, grades kept
- Flybrain experiments (re-read 31 Oct): three new experiment pages. In the steering experiment the fly circuit did not beat a direct map and was not reliably better than rewired copies; in the search experiment the fly wiring gave no benefit over rewired wiring. Grade A kept. In numbers: in the steering experiment the fly circuit's tracked fraction minus that of a direct map was −0.021 (95% interval −0.035 to −0.007) with the target visible and +0.007 (−0.005 to +0.020) with it hidden; in the search experiment the real projection-neuron-to-Kenyon-cell wiring scored 2.38% and 0.55% below degree-preserving rewired wiring on the two sides. Its ledger row (a connectome language model) is unchanged and stays mixed. (catalogue entry)
- fly-connectome-reservoir (re-read 31 Oct): a new technical report repeats the same numbers (100% real, 83.9% shuffled, 50% without connections, 3 seeds) and says three shuffled networks are far too few to claim an advantage. Grade B kept. Its ledger row stays helps. (catalogue entry)
Browser: fly67 again
fly67: runs again on 31 Oct. It downloaded about 33 MB of brain data, identified by their web address, a method we fixed before this check. First activity on the page came 11.72 s after loading; 33.28 MB came down in total, with no page errors; we clicked "Feed it sugar", as on 18 Oct. Nothing else ran on the server during the load.
Checked in headless Chromium 154.0.8037.57 on a Linux server without a GPU, 31 Oct 2026. Shows whether the page loads, downloads its brain data and keeps computing; it is not a play test. Safari, Firefox, phones and GPU machines were not tested.
neural-canvas: The copy of this demo at fruitfly.live answered on 30 and 31 Oct with a security certificate made for its hosting service's own addresses, not for fruitfly.live, so browsers refuse to open it; this looks like a setup problem at that address, which we did not check further. The Hugging Face version is the one we link. The block on Ways to try
Tracking and leads
- 31 Oct 2026: we looked up the latest commit of 142 tracked repositories: 127 unchanged, 14 with new commits, 1 gone or private (flydoom). Three of the changed ones were re-read (fly-naf, flybrain-prove-ai and fly-connectome-reservoir); the other changes are followed up in the next check and were not judged today. 57 of 57 videos available; no new view counts.
- 31 Oct: we scanned news, videos and new code from 30 to 31 Oct; nothing met the bar for a new claim check, our news searches found no new coverage of the walking-circuit paper beyond the journal article itself, and two new code projects with their own scrambled-wiring tests were added. Videos: both YouTube searches (by upload date and by views this week) answered in full; 5 new videos, none a claim to check. News: The Transmitter reported a fly taste-wiring study on 30 Oct (research news, no simulation claim). Reddit could not be searched (access refused).
- Dataset pages (FlyWire Codex, neuPrint, FlyEM, BANC) answered as before, with no new version; the Janelia male CNS page still answers 404.
- Odyssey-3 claim: No method, code or write-up was published as of 30 Oct 2026 (checked 30 Oct); no public fly mode found on the preview page (29 Oct). The verdict row stays U.
- Software versions: 12 of the 21 pinned packages we follow have newer releases since 19 Oct (for example pytz and tzdata 2026.6); our pinned versions were not changed.
What we could not resolve
- The walking-model size is estimated, not measured, and the margin over our limit is about 106 MB; whether the real wiring is needed for the simulated walking rhythm stays not tested.
- Fly-NAF's "helps" rests on one replayed night (8 threats) and on constants tuned on the real wiring; we rate it unfair and weak.
- For Fly Hearing Lab, reading higher rhythm tuning as better is our interpretation of the author's hypothesis; the author's own conclusion (not supported) points the same way.
- DrosoCode's rewired copy keeps each target's number of inputs only; one run each.
- 11 further tracked repositories with new commits were not judged in this check.
Sources: Pugliese et al., Nature (28 Oct 2026); code (MIT, commit 10e7661); Fly-NAF (experiments 17 and 19, commit 64cdef2); Fly Hearing Lab (commit f175eeb); DrosoCode (commit 4899921); flybrain experiments (commit e04a8ac); fly-connectome-reservoir (commit 9c385a5); fly67; each project's catalogue row cites the exact commit and file lines.