Research report · 30 October 2026 · twenty-seventh check
Twenty-seventh check: the walking-circuit paper's shuffle test, four new entries and four older demos re-checked
We read the shuffle test in the code of a new Nature paper on a fly walking circuit, graded three more projects that compare real and scrambled wiring, re-checked 27 earlier grades and loaded four older browser demos again.
Short answer
The walking-circuit paper's shuffle test is about the shape of the wiring, not about walking. The authors' code includes a test that shuffles the wiring while keeping the balance of excitatory and inhibitory connections; their saved notebook output shows the number of three-neuron circuits of the walking-rhythm type among front-leg nerve-cord neurons: 21,544 with the real wiring and 306.95 ± 29.10 (mean ± SD) across 100 shuffles. We did not re-run it (checked 30 Oct 2026). This is a count of circuit patterns in the wiring, not a walking test with shuffled wiring, so our ledger does not score it.
Three more projects compare real and scrambled wiring, and none of them shows a clear advantage for the fly wiring: one no difference and two mixed. Fly wiring still beats scrambled wiring in 35 of 91 studies.
30 Oct: re-checked in the same headless browser without a GPU: 1 of 4 older demos ran (17 Oct: 1 of 4). 27 entries first graded 13-19 Oct were re-checked on 30 Oct: 25 keep their grade; 2 with larger upstream changes are re-read next.
The walking-circuit paper: what its shuffle test measures
Code at commit 10e7661 (MIT licence), the saved output of its Figure 3 notebook and the preprint abstract · read 30 Oct 2026 · nothing re-run
Pugliese, Tuthill, Brunton and colleagues (Nature, 28 Oct 2026) simulated rate models of the fly's ventral nerve cord, built from four wiring maps (MANC, FANC, BANC and the male CNS), and pruned them to a small circuit that still produced leg rhythms in the model: two excitatory neurons in a chain and one inhibitory neuron inhibiting both. The paper also reports tests in real flies, which we did not check.
- Real wiring: 21,544 circuits
- 100 shuffles: 306.95 ± 29.10 on average (about 70 times fewer)
In the shuffle, each front-leg neuron's outgoing weights are moved to other targets, so each neuron keeps its sign and its outgoing weights; incoming connection counts are not kept. Of the 21,544 circuits in the real wiring, 19,799 were oscillatory in the authors' linear estimate (shuffles: 302.66 ± 28.45). The shuffled wiring is not run as a walking or rhythm simulation in the output we read, so our ledger records the study but does not score it.
Catalogue entry: pugliese-2026-walking-cpg, grade A, peer-reviewed, MIT licence. A nerve-cord study, not a whole-brain or body simulation. Verdict row on Is it real?
4 new entries and 4 new ledger rows
The authors' 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 |
|---|---|---|---|---|
| Pugliese et al. 2026 (Nature) A | Not scored | 21,544 walking-rhythm-type three-neuron circuits among front-leg nerve-cord neurons (MANC) | 306.95 ± 29.10 (mean ± SD) across 100 shuffles that keep each neuron excitatory or inhibitory and keep its outgoing weights | A count of circuit patterns in one wiring map, not a walking or rhythm test with shuffled wiring; incoming connection counts are not kept by the shuffle. Saved notebook output; not re-run by us. |
| fly-connectome-forecast A | No difference | Mean forecast skill 0.355 over six wind and solar tasks (5 seeds) | 0.353 degree-preserving rewired; 0.355 random reservoir; 0.345 shuffled weights | Means over tasks averaged by us from the author's table; no per-task uncertainty used. Removing the recurrent layer lowered skill to 0.334. |
| RiskFly A | Mixed | Test error (MAE) 0.00576 (5 seeds; lower is better) | 0.00723 degree-preserving rewired; 0.00857 random sparse | Fly better in 5 of 5 seeds, author's p = 0.0625, called exploratory, while a dense network did best (0.00537). 128 neurons; the target is a formula, not an investment strategy; no licence file. |
| fly-novelty-scada A | Mixed | CARE score 0.30 (3 seeds; higher is better) | 0.43 shuffled (difference −0.13, 95% interval −0.29 to +0.03); 0.44 random | The detector found almost no faults with the alarm settings used; a k-nearest-neighbours detector scored about 0.51. |
RiskFly and fly-novelty-scada read "mixed" because their wiring nulls disagree: in both, the main rewired or shuffled comparison does not separate from the real wiring by our rules (RiskFly: p of 0.05 or more; fly-novelty-scada: the interval includes zero), while another null differs. Ledger now: helps 35, no difference 26, worse 9, mixed 17, not yet scored 4 (each row on Does fly wiring help?).
Seven more candidate projects were looked at and closed without a grade, seven were left out because their shuffle did not touch the wiring or no result was committed, and four names from the news could not be found as code. Eight new code projects from the scan wait for a first look.
27 earlier grades re-checked
27 entries first graded 13-19 Oct were re-checked on 30 Oct: 25 keep their grade; 2 with larger upstream changes are re-read next. For each, we compared the commit we graded with the current one and checked that the files we cited are still there.
- 21 repositories had no new commits and every cited file was present: grade kept.
- 4 had new commits that do not touch the graded result: biogrid (a README badge), malecns-keratitis-pilot (assistant instruction files), flylab-demo (a rewritten roadmap) and malecns-training-gomoku (a 15x15 follow-up whose README reports no consistent advantage, the same direction as its graded no-difference; we did not read the follow-up's numbers). Grade kept.
- 2 changed a lot: flybrain-prove-ai (new result pages for three experiments) and fly-connectome-reservoir (a new technical report and README). Their grades stand until we re-read them.
Browser: four older demos and fly-arithmetic
30 Oct: re-checked in the same headless browser without a GPU: 1 of 4 older demos ran (17 Oct: 1 of 4).
- flybrain.app: runs again on 30 Oct. It downloaded about 12.4 MB of brain data, identified by their web address, a method we fixed before this check.
- Help the Fly Escape: loaded, then froze in our no-GPU browser again on 30 Oct.
- FlyLeno: loaded, then froze in our no-GPU browser again on 30 Oct. It downloaded about 32 MB of brain data, identified by their web address, a method we fixed before this check.
- fly-brain-lulzx: stopped at our browser memory limit again on 30 Oct. It downloaded about 34 MB in total.
Not counted among the four: fly-arithmetic: runs again on 30 Oct (11 Oct: runs).
Checked in headless Chromium 154.0.8037.57 on a Linux server without a GPU, 30 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.
For Help the Fly Escape, fly-brain-lulzx and fly-arithmetic, our fixed method did not identify a brain-data file, so we give no brain-data figure for them. The block on Ways to try
Tracking and leads
- 30 Oct 2026: we looked up the latest commit of 142 tracked repositories: 127 unchanged, 14 with new commits, 1 gone or private (flydoom). From commit messages and file lists we judged most changes not material; flybrain-prove-ai and fly-connectome-reservoir added larger new results and fly-naf a commit about its wiring, so those three are re-read next, and fly-worker's demo code changed, so its browser check is redone later. 57 of 57 videos available; no new view counts.
- fasterfly: an MIT licence file was added upstream on 30 Oct.
- Links: ascend's demo page and FlyGym's browser viewer: not confirmed by our link check (29 and 30 Oct). We do not know whether ascend moved, was renamed or was made private.
- 30 Oct: we scanned news, videos and new code from 29 to 30 Oct; nothing met the bar for a new claim check, and eight new code projects were listed for a first look.
- 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: the 21 pinned packages we follow are unchanged since 29 Oct.
What we could not resolve
- The Pugliese test counts circuit patterns in one wiring map (MANC front leg); whether shuffled wiring would still walk was not tested in the output we read. We could not open the journal page.
- RiskFly has five seeds (p = 0.0625 is the smallest possible with five paired seeds) and fly-novelty-scada three; neither is re-run by us.
- The seven closed candidates were judged from search lines and file names only; fly-naf and fly-worker were judged to have material changes from commit messages and file lists only, and are re-checked later.
- flybrain.app's computing location stays unclear: our browser recorded a call to a server, which we did not inspect.
Sources: Pugliese et al., Nature (28 Oct 2026); preprint; code (MIT); fly-connectome-forecast; RiskFly; fly-novelty-scada; each project's catalogue row cites the exact commit and file lines. Previous report: 29 Oct 2026.