Research report · 24 October 2026 · twenty-fifth check
Twenty-fifth check: the third re-check of our grades, and neural-canvas's connectome download
We re-checked the 31 projects we last graded between 6 and 12 Oct against their current code, finished the two re-reads left from 23 Oct, graded two new projects with wiring controls, and loaded neural-canvas once more in our browser.
Short answer
24 Oct: we re-checked the 31 projects we last graded between 6 and 12 Oct against their current code: 31 keep their grade (1 after a full re-read). 24 Oct: the project left from the 23 Oct re-check, acamilo-flybrain, keeps its grade C after a full re-read.
fly-chess keeps B, and its wiring result is now worse: the author committed a shuffled-wiring control, and under the same readout probe the shuffled brain scored higher than the real one (20.9% against 18.0% top-1; board-blind 11.7%; on the author's Elo ladder 682 against 653 and 631). Only a readout is trained on the frozen connectome.
Fly wiring beat scrambled wiring in 33 of 84 studies. somahouse A helps only against its rewired copy: dense wiring and a no-brain controller both cost less than the fly wiring.
The re-check: 31 projects graded 6 to 12 Oct
9 A, 15 B and 7 C in the 23 Oct catalogue, each checked against the commit it was graded at and today's code
24 Oct: we re-checked the 31 projects we last graded between 6 and 12 Oct against their current code: 31 keep their grade (1 after a full re-read).
| Outcome | Projects |
|---|---|
| Kept without a full re-read (30) | bad-apple-fly, connectocopter, connectome-control-exploration, doomfly, duth-x-stepmania, flappy-fly, fly-arithmetic, fly-brain-minecraft, fly-brain-tetris-ag, fly-clash-royale, fly-distill, fly-jumpshot, fly-ocr, fly-racer, fly-tennis, fly-terraria, fly-vnc-walking-sim, fly64, flybrain-flappy, flybridge, flyhard, flyleno, flyonenomics, flyvl, flywire-playground, flywire-pong, let-it-flyppen, melanograph, stonkfly, synaptera |
| Kept after a full re-read (1) | brainlab |
| Left from 23 Oct, decided after full re-reads (2) | acamilo-flybrain (C kept), fly-chess (B kept; fields updated) |
27 of the 31 still point at the commit they were graded at, and every cited file is still there. Four have new commits: fly-brain-tetris-ag, fly-racer, synaptera (a new folder, a demo interface or one added score, none touching what the grade rests on), and brainlab, which was re-read in full.
brainlab. On 23 Oct 2026 the author added a report that records a failed check of a different result, whether the real wiring helps the frozen olfactory circuit integrate angular velocity (the real wiring did no better than a no-information guess); the first-campaign report its grade cites did not change. The grade stays A, and its ledger row is unchanged.
fly-chess: the shuffled brain scored higher
The author's committed metrics · single runs, one shuffle seed · not re-run by us
fly-chess keeps B, and its wiring result is now worse: the author committed a shuffled-wiring control, and under the same readout probe the shuffled brain scored higher than the real one (20.9% against 18.0% top-1; board-blind 11.7%; on the author's Elo ladder 682 against 653 and 631). Only a readout is trained on the frozen connectome. The connection strengths stay as measured; training them too made held-out accuracy worse, by the author's account.
In the catalogue, fly-chess's controls change from "none" to a wiring null, its wiring result to "worse" and its trained part to "a trained readout"; the grade stays B, because the control shows that this wiring does not help, not that its specific wiring matters.
2 new entries and 3 new ledger rows
- Helps somahouse: from the author's committed held-out results (60 test scenarios, 3 training seeds; not re-run by us), the fly-derived circuit cost 0.1589 against 0.1622 for a degree-preserving rewired copy (lower is better), but dense wiring (0.1538) and a conventional controller with no brain (0.1172) both cost less than the fly wiring. Fair, strong; the no-brain baseline is not beaten.
- No difference fly-trader: from the author's committed matched comparison (3 seeds, 10 unseen tickers; not re-run by us), the excess return over buy-and-hold (the no-brain baseline, 0 by definition) was +280.3 points with the FlyWire circuit against +211.0 with a random-weight circuit of the same size (paired difference +69.3, 95% interval -186.6 to +325.3). Back-test only; partly fair (random weights, not a degree-preserving rewiring), strong.
- Worse fly-chess: from the author's committed metrics (single runs, one shuffle seed; not re-run by us), the shuffled wiring matched Stockfish's best move more often than the real MaleCNS wiring (20.9% against 18.0% top-1; board-blind baseline 11.7%); cutting the central brain gave 17.9%. Partly fair, weak.
Of 25 repositories checked for a committed wiring comparison, 3 passed; 2 were graded, and the third will be graded first in the next run. The other 22 were not admitted on 24 Oct.
neural-canvas in our browser
Our browser check · procedure written down before the load · headless Chromium 154, Linux, no GPU · one load · not a play test
24 Oct: in a headless browser without a GPU, after we pressed 'Download & start', the page downloaded about 77 MB of connectome data within 82 seconds and started a background worker, and the picture kept changing in our screenshots after that. Screenshots cannot show whether the brain drives the movement.
Our screenshot requests timed out from about 143 seconds after the click, so we have no picture of the page for the rest of our watch.
- How we identified the files:
We identified the connectome files by their web address, a method we fixed before this check.
All 27 connectome files the page lists, about 76.7 MB in total, were seen. - Status line:
The page's own status line read "JavaScript · 0.08× realtime · 0 spikes".
- Catalogue status:
runs in a headless browser without a GPU (checked 2026-10-24; not a play test)
- Count: In our browser checks of 3, 6, 11, 17, 18, 20, 22 and 24 Oct 2026, 8 of 13 demos ran by our check's rule on a machine without a GPU at their latest check. neural-canvas is not counted among them: our check's rule marks a page as unresponsive when its screenshot requests fail, and on 24 Oct our screenshot requests timed out from about 143 seconds after the click, so we have no picture of the page for the rest of our watch.
The 22 Oct check stays on the record: 22 Oct: in a headless browser without a GPU, after we pressed 'Download & start', the page prepared its scene for about 73 seconds, then started a background worker, which downloaded about 79 MB within about 94 seconds of the click; the picture changed between our later screenshots, and at the end of our about 187-second watch the page's status line read 'JavaScript · 0.02× realtime · 0 spikes'. Screenshots cannot show whether the brain drives the movement.
Tracking and leads
- 143 tracked repositories checked on 24 Oct 2026: 138 unchanged, 2 already gone, 3 changed (among them flybench, whose new commit will be checked against its grade in the next run); no new broken links; 56 of 56 videos available.
- View counts: 12 of the 56 videos were refreshed from YouTube search pages on 24 Oct 2026 (approximate snapshots); the other 44 keep their earlier dated values.
- Odyssey-3 claim: No method, code or write-up was published as of 24 Oct 2026; no public fly mode found on the preview page (24 Oct). The verdict stays U.
- Lead scan: YouTube search pages answered in English, French, Spanish and Russian; news, Hacker News and the repository sweep gave nothing new to admit. One debunk video found on 24 Oct goes through our usual gallery check first.
What we could not resolve
- neural-canvas: after about 143 seconds from the click our screenshot requests timed out, so we cannot say what the page showed for the rest of our watch; our check's rule therefore marks the page unresponsive, while the download and the worker were seen.
- somahouse's "helps" rests on a small difference over 3 training seeds, and dense wiring and the no-brain controller both did better.
- fly-chess's control is a single run with one shuffle seed, at the short probe stage only.
- fly-trader's intervals are very wide (3 seeds), and it is compared with random weights, not a degree-preserving rewiring.
- View counts are approximate snapshots from search pages on one day. All ledger numbers in this report are the authors' own; we did not re-run them.