Nineteenth check: the 30 spikes/s scrambled brains, strengthened, and a browser correction

Scrambled brains at 30 spikes/s, strengthened

Our own test · pre-registered 18 Oct 2026, 09:34 UTC · brain only · one model · not a real fly

Design. We took the same 4 scrambled graphs as on 17 Oct (each neuron keeps its number of connections; shuffle seeds 4 and 5 for each eye) and multiplied every connection strength inside the brain by one factor, leaving the input alone. The target was the real brain's count of spikes outside the stimulated looming neurons in the same 1-second trial at 30 spikes/s: 5,266 for a shadow on the left eye and 5,901 on the right, with a band of 30% either side. The search followed a fixed step rule with a cap of 12 trials; it used 6. We read the turn neuron (DNa02) on the far side, the giant fibres and MDN by the same rules as on 17 Oct.

Checks before reading. The trial program was byte-identical to the 17 Oct one; the saved drives of the no-brain floor, the real brain and one 17 Oct scrambled brain were reproduced exactly; and one 17 Oct scrambled trial was repeated with identical spike counts, active neurons and descending-neuron rates. Every strengthened trial used the same scrambled graph as its 17 Oct twin.

Pre-registered sentences. 18 Oct: we then strengthened the connections inside scrambled brains (not the input) until their neurons outside the stimulated input fired about as many spikes as in the real brain (within 30%; same input, one 1-second trial each).

• Left eye: On the left side, 2 scrambled brains matched this way, and neither sent a turn command.
• Right eye: On the right side, 2 scrambled brains matched this way, and neither sent a turn command.

Matching the number of spikes does not match the pattern: 1,152 to 2,228 neurons fired in the matched scrambled brains, against 379 (left) and 432 (right) in the real brain. 1 of the matched scrambled brains sent small steering signals, which we did not walk in the body.

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 repeats the 14 Oct body results with the 17 Oct and 18 Oct scrambled brains, and the chart of the strengthening search.

fly67 at its current address, and Fly Worker again

On 17 Oct our browser loaded fly67 at the address it had on 3 Oct, which answered "not found". The project moved to gexhaw.github.io on 13 Oct, so this time we loaded it there. We also loaded Fly Worker again, because on 17 Oct the text we track on its page had not changed. The browser ran before any brain trial, never at the same time.

In our catalogue fly67 and ChessFly now point to their gexhaw addresses; their author field still names the earlier account, and we did not change it.

5 new entries and 2 new ledger rows

• FlySnake A (ledger: helps): a MaleCNS v1.0 network plays Snake through a readout trained on descending-neuron spike counts. On 100 held-out games the real wiring scored 20.03 food per game, a rewired graph with its own trained readout 3.63 (paired 95% interval of the difference 15.32-17.46), random actions 0.17, and a hand-written rule 21.52. One rewired graph; the author's numbers, not recomputed by us.
• flywire-snn-stdp A (ledger: mixed): a FlyWire v783 hearing circuit (1,239 neurons) trained with a local learning rule tells apart five courtship-song rhythms with accuracy 0.88, against 0.724 for a configuration-model graph and 0.428 for a random graph (5 seeds). Without learning the real wiring is not ahead (0.566 vs 0.604), and a random input mapping with learning reaches 0.920.
• Fruit Fly Plays Terraria (Moon Lord) B: the Shiu et al. FlyWire model fights Terraria's final boss through a readout trained to imitate a teacher. A shuffled-wiring control exists in its code, but its numbers appear only in the README, with no result file, so it has no ledger row. A different project from the earlier fly-terraria entry.
• BioGrid B: a FlyWire v783 whole-brain screen of 140 sensory types for a steering neuron, with knockouts compared with random knockouts; its controls are lesions, not a wiring control.
• The Odyssey-3 claim U (a verdict row; see below).

Of 31 new leads, 4 were graded, 9 wait for a later check and 18 were excluded on 18 Oct, each for a stated reason (mostly no committed result file or wiring comparison we could read; one is a build-only copy of another lead): someone405-ship-it/fruitfly-roblox-agent, Hodaska/fly-chess, JacoMoolman/fly-brain, hotocoo/malecns, 777LVB777/flymarket, amirhaghani-brain/fly-connectome, DeevlerPlays599/FlyBrain-Build, dimspmj/Flyware-brain, Jeanne-Dagmar/flybrain-game-server, joseeto433/mosca-pista, koishiSaikou/embodied-fly-lab, Kyoto-Designs/Fly-brain, marginli/bsc-topics, Mostly1sAnd0s/flygym-stream, pascaldisse/nari-brain, Sneha-Karki/Female-Adult-Fly-Brain-Network, TBuksik/fly-brain-lab and vadantosaha/Fly-brain. The fruitfly-roblox-agent repository links no Roblox video or account, so the Roblox avatar claim stays U.

A caveat marker in the ledger. Eight rows now carry a marker for reading only: for Flycraft, a control that keeps the wiring but scrambles or replays the activity came close to the real result; for fly-cartpole, fly-connectome-reservoir and Wired Different, a control without the fly wiring came close; for flight-test-the-fly (p = 0.0909), flyarm (p = 0.125), flyaim (p = 0.66) and the larva reservoir (p = 0.51), the authors' own test did not separate the real wiring from a control. The marker changes neither the rows' wiring effect nor any count (the ledger). Flycraft's own p-value (0.29) is only in its free text, so its marker comes from the replay control.

The ledger now has 62 studies, 53 with a wiring null: helps 25, no difference 11, worse 4, mixed 11, not yet scored 2. Our looming row is not re-scored; the 18 Oct trials are in its note, brain only.

A new verdict row: the Odyssey-3 claim

Odyssey (@odysseyml) said on X on 17 Oct 2026: We placed the fruit fly brain inside Odyssey-3, a frontier world model. The lil' guy is now flying around totally imagined, simulated worlds all by itself (the post, which has a short video). We read the post's text from the page without logging in; we could not see the video or the replies. Odyssey's Odyssey-3 announcement page does not mention a fruit fly, a fly brain or a connectome, and we found no paper, repository, documentation or demo page about the fly.

No method, code or write-up was published as of 18 Oct 2026, so the claim cannot be checked; we do not know which fly brain model was used (a connectome simulation, a learned policy or something else), what drives the flight or how the world model feeds it input. What would settle it: a write-up or code that names the fly brain model, how its outputs drive the flight and how the world model feeds it sensory input. Grade U (the verdict row).

Verdicts, tracking and freshness

• Repositories of 15 of the 23 verdict rows re-checked on 18 Oct 2026; none changed since grading. Rows without code are carried.
• All 56 videos available. Two new videos about the Roblox claim name no builder, game code or account behind it, so they were not promoted.
• Links: fly67 and ChessFly answer at their gexhaw addresses; fruitfly.live still has a TLS certificate error; FlyDoom's repository still answers "not found". No PyPI package in our list changed version.

What we could not resolve

• One trial each. Each strengthened scrambled brain ran one 1-second trial at one trial seed. Two of the four matches sit near the lower edge of the band (0.71 and 0.76 of the real count, both on the left).
• Count, not pattern. We matched the number of spikes outside the input, not where they happen: 3 to 6 times as many neurons fired in the matched scrambled brains as in the real brain.
• One steering signal not walked. One matched brain (right eye, shuffle 4) sent a small steering signal (far-side turn neuron 1 Hz in one 100 ms bin). We did not walk it in the simulated body.
• Fly Worker screenshots. Our rule reads "ran", but our screenshots show only the static top of the landing page, so we did not see the fly. A later check could scroll the page or use a larger window.
• Brain only, one model. Shiu et al. LIF on FlyWire v783; the input is a rate we set for every looming neuron of one eye, not a measured response to a real shadow; not a real fly.

Earlier: report of 17 Oct 2026 (the scrambled brains at 30 spikes/s before strengthening; the six browser demos re-checked).