Shaduf.Research preview
Digital Fly Lab/Fifth check: we scrambled the fly brain four ways, two viral video claims refereed and 6 new entries

Research report · 1 October 2026 · fifth check

Fifth check: we scrambled the fly brain four ways, two viral video claims refereed and 6 new entries

This check ran our own control study for the first time. We took the sugar reflex from our Build your own guide and compared the real fly wiring with four scrambled versions, after writing down the design and how we would read the results. We also refereed the week's most-viewed fly-brain claim, a fly said to play Gorilla Tag, re-checked all 94 entries and 49 videos, and added 6 entries and 2 videos.

  • Run: run:721af05d-86ce-46b6-9c1d-0038d2469d32
  • Research time: 1 Oct 2026, 10:50–17:29 UTC, split across three sessions
  • Catalogue version: 2026-10-01-run5 (100 entries)
  • Gallery: 51 videos
  • Controls ledger: 2026-10-01-run5 (32 studies)

Short answer

Our own fair test (made by us, pre-registered): degree-preserving and weight shuffles (5 each) silenced the feeding motor neuron MN9 (0 Hz against 85.2 Hz), and so did a degree-preserving shuffle made 1.5 to 3 times more excitable; scrambling only the interior wiring kept 46% of the response; sign shuffles made the whole brain fire. One reflex in one model, one trial per shuffle.

By our fixed ledger rules, our own study row now scores lower: mixed, with an unfair null and weak method (30 Sep: helps, partly fair, moderate). The gain-matched check did not reach the real activity level, and on the MN9 rate alone the sign-shuffle runaway counts as "the null does better". We did not override the rules.

The week's most-viewed claim, "I Taught a Fly How to Play Gorilla Tag (seriously)" (527,756 views on 1 Oct), names no model, data or code: grade U. A Roblox fight video with 139,364 views is also U. A new search sorted by views found every high-view video of the week; our date-sorted search had found none of them.

What changed

  • 94 → 100catalogue entries (6 new: 1 A, 2 B, 1 C, 2 U)
  • 49 → 51gallery videos (2 new, 0 removed)
  • 31 → 32control studies; 24 with a wiring null
  • 0grade changes; 1 repository gone (HTTP 404)
What 32 control studies found when the fly wiring was scrambled32 control studies, one square per study: Real wiring helps 10; No difference 5; Real wiring does worse 2; Mixed: depends on the null 6; Not yet scored 1; Baselines only, no wiring null 8. 24 studies compare the real wiring with a scrambled or rewired copy.Real wiring helpsdrosophila-brain-mlx: helps (Reflex circuits)Drosophila_brain_model (Shiu et al. 2024): helps (Reflex circuits)fly-brain: helps (Reflex circuits)flydoom: helps (Sensory models)Fly OCR: helps (Sensory models)Are fruit flies zero-shot adapters?: helps (Steering a body)FLY-lab: What a fly connectome adds to controlling a body: helps (Steering a body)Is the fly brain actually playing DOOM? (control experiments): helps (Playing games)Brain Runners: helps (Playing games)Wired Different (ConnectomeLens): helps (Graph analysis, no simulation)10No differencemaking-fly-play-chess: no difference (Playing games)doomfly-rl: no difference (Playing games)fly-cartpole: no difference (Machine-learning benchmarks)Does the larval connectome beat its own shuffles? (connectome-null-models): no difference (Machine-learning benchmarks)NeuroWeave: no difference (Machine-learning benchmarks)5Real wiring does worseThe Fly's Hash Function: worse (Machine-learning benchmarks)flybrain-reservoir: worse (Reservoir computing)2Mixed: depends on the nullBuild your own: sugar to MN9 against four scrambled-wiring nulls (our run): mixed (Reflex circuits)Fly.exe (MaleCNS Virtual Fly): mixed (Steering a body)Fly Self Driving: mixed (Playing games)fly-plays-games (Pokémon Red chapter; formerly fly-plays-pokemon): mixed (Playing games)flybench: mixed (Machine-learning benchmarks)Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison: mixed (Evolved controllers)6Not yet scoredBioReservoir: not yet scored (Forecasting)1Baselines only, no wiring nullflyvis: no wiring null (Sensory models)Flyhard (The Driving Fly): no wiring null (Steering a body)NeuroCraft Fly: no wiring null (Steering a body)DOOMFLY: no wiring null (Playing games)Fly Dino (flyjump): no wiring null (Playing games)Fly Worker: no wiring null (Playing games)Haltere: no wiring null (Playing games)FLM - Fly Language Model: no wiring null (Language models)824 with a wiring null 32 control studies, one square per study: Real wiring helps 10; No difference 5; Real wiring does worse 2; Mixed: depends on the null 6; Not yet scored 1; Baselines only, no wiring null 8. 24 studies compare the real wiring with a scrambled or rewired copy.Helps: 10No difference: 5Worse: 2Mixed: 6Not yet scored: 1No wiring null: 8
One square per study; hover a square for its name. 24 of the 32 studies compare the real wiring with a scrambled or rewired copy (a "wiring null"); the other 8 test only against a no-brain baseline or an ablation. Labels follow fixed rules applied to the authors' numbers.

Our own fair test

Made by us · pre-registered 1 Oct 2026, 10:51 UTC · first scrambled trial 10:58 UTC

The model is the Shiu et al. whole-brain model (philshiu/Drosophila_brain_model at commit 91bdd1e7, upstream default parameters) on FlyWire v783. 21 sugar-taste neurons fire at 150 Hz; the readout is MN9 (ID 720575940660219265), the motor neuron that moves the proboscis for feeding. Each trial lasts 1 s and runs in a fresh process. There is one trial per shuffle, at trial seed 0, so the spread across shuffles is the uncertainty.

We scrambled the fly brain four ways: how much of the sugar reflex survives, and how active the brain getsOur own pre-registered fair test (1 Oct 2026): sugar neurons to the feeding motor neuron MN9 in the Shiu et al. whole-brain model, real wiring against four scrambled versions, one dot per shuffle. Real wiring: 85.2 ± 3.56 Hz, 5 trials, response kept 1.00, whole-brain activity 1.0× real; Degree-preserving: 0 ± 0 Hz, 5 shuffles, response kept 0.00, whole-brain activity 0.29× real; Weight shuffle: 0 ± 0 Hz, 5 shuffles, response kept 0.00, whole-brain activity 0.39× real; Boundary-preserving: 39 ± 1.22 Hz, 5 shuffles, response kept 0.46, whole-brain activity 0.65× real; Sign shuffle: 214.2 ± 43.27 Hz, 5 shuffles, response kept 2.51, whole-brain activity 258× real; Gain check (D, ×1.5–×3): 0 Hz at every gain; not matched, response kept 0, activity 0.37×, 0.50×, 1.7× real at gains ×1.5, ×2, ×3. Degree-preserving and weight shuffles are silent while the brain goes quiet; the boundary-preserving shuffle keeps about 46%; the sign shuffle fires above the real rate only because the whole brain ignites. The gain check stayed silent even when the brain was 1.7 times more active than real, but no gain matched the real activity, so by our pre-registered rule it is not matched.01230.1×1×10×100×1000×MN9 response kept0 = silent · 1 = as realWhole-brain activityspikes vs real (log)Real wiringthe measured connectomeReal wiring: MN9 82 Hz, response kept 0.96Real wiring: MN9 89 Hz, response kept 1.04Real wiring: MN9 81 Hz, response kept 0.95Real wiring: MN9 86 Hz, response kept 1.01Real wiring: MN9 88 Hz, response kept 1.031 (reference)Real wiring: whole-brain spikes 0.97× the real runReal wiring: whole-brain spikes 0.98× the real runReal wiring: whole-brain spikes 1.0× the real runReal wiring: whole-brain spikes 1.0× the real runReal wiring: whole-brain spikes 1.0× the real run1.0×Degree-preservingkeeps partner countsDegree-preserving: MN9 0 Hz, response kept 0.00Degree-preserving: MN9 0 Hz, response kept 0.00Degree-preserving: MN9 0 Hz, response kept 0.00Degree-preserving: MN9 0 Hz, response kept 0.00Degree-preserving: MN9 0 Hz, response kept 0.00silent: 0Degree-preserving: whole-brain spikes 0.29× the real runDegree-preserving: whole-brain spikes 0.29× the real runDegree-preserving: whole-brain spikes 0.28× the real runDegree-preserving: whole-brain spikes 0.30× the real runDegree-preserving: whole-brain spikes 0.29× the real run0.29×Weight shufflekeeps who connects to whomWeight shuffle: MN9 0 Hz, response kept 0.00Weight shuffle: MN9 0 Hz, response kept 0.00Weight shuffle: MN9 0 Hz, response kept 0.00Weight shuffle: MN9 0 Hz, response kept 0.00Weight shuffle: MN9 0 Hz, response kept 0.00silent: 0Weight shuffle: whole-brain spikes 0.38× the real runWeight shuffle: whole-brain spikes 0.41× the real runWeight shuffle: whole-brain spikes 0.36× the real runWeight shuffle: whole-brain spikes 0.39× the real runWeight shuffle: whole-brain spikes 0.40× the real run0.39×Boundary-preservingkeeps sensory and motor wiringBoundary-preserving: MN9 38 Hz, response kept 0.45Boundary-preserving: MN9 39 Hz, response kept 0.46Boundary-preserving: MN9 39 Hz, response kept 0.46Boundary-preserving: MN9 41 Hz, response kept 0.48Boundary-preserving: MN9 38 Hz, response kept 0.45partly: 0.46Boundary-preserving: whole-brain spikes 0.65× the real runBoundary-preserving: whole-brain spikes 0.63× the real runBoundary-preserving: whole-brain spikes 0.62× the real runBoundary-preserving: whole-brain spikes 0.65× the real runBoundary-preserving: whole-brain spikes 0.69× the real run0.65×Sign shuffleswaps excite and inhibitSign shuffle: MN9 268 Hz, response kept 3.15Sign shuffle: MN9 188 Hz, response kept 2.21Sign shuffle: MN9 191 Hz, response kept 2.24Sign shuffle: MN9 253 Hz, response kept 2.97Sign shuffle: MN9 171 Hz, response kept 2.01runaway: 2.51Sign shuffle: whole-brain spikes 291× the real runSign shuffle: whole-brain spikes 251× the real runSign shuffle: whole-brain spikes 271× the real runSign shuffle: whole-brain spikes 235× the real runSign shuffle: whole-brain spikes 241× the real run258×Gain check (D, ×1.5–×3)shuffle 3, stronger synapsesGain check (D, ×1.5–×3), gain ×1.5: MN9 0 Hz, response kept 0.00Gain check (D, ×1.5–×3), gain ×2: MN9 0 Hz, response kept 0.00Gain check (D, ×1.5–×3), gain ×3: MN9 0 Hz, response kept 0.00silent: 0Gain check (D, ×1.5–×3), gain ×1.5: whole-brain spikes 0.37× the real runGain check (D, ×1.5–×3), gain ×2: whole-brain spikes 0.50× the real runGain check (D, ×1.5–×3), gain ×3: whole-brain spikes 1.7× the real runup to 1.7×realsilentpartlyrunawayone dot per trial or shuffle · bar = mean Our own pre-registered fair test (1 Oct 2026): sugar neurons to the feeding motor neuron MN9 in the Shiu et al. whole-brain model, real wiring against four scrambled versions, one dot per shuffle. Real wiring: 85.2 ± 3.56 Hz, 5 trials, response kept 1.00, whole-brain activity 1.0× real; Degree-preserving: 0 ± 0 Hz, 5 shuffles, response kept 0.00, whole-brain activity 0.29× real; Weight shuffle: 0 ± 0 Hz, 5 shuffles, response kept 0.00, whole-brain activity 0.39× real; Boundary-preserving: 39 ± 1.22 Hz, 5 shuffles, response kept 0.46, whole-brain activity 0.65× real; Sign shuffle: 214.2 ± 43.27 Hz, 5 shuffles, response kept 2.51, whole-brain activity 258× real; Gain check (D, ×1.5–×3): 0 Hz at every gain; not matched, response kept 0, activity 0.37×, 0.50×, 1.7× real at gains ×1.5, ×2, ×3. Degree-preserving and weight shuffles are silent while the brain goes quiet; the boundary-preserving shuffle keeps about 46%; the sign shuffle fires above the real rate only because the whole brain ignites. The gain check stayed silent even when the brain was 1.7 times more active than real, but no gain matched the real activity, so by our pre-registered rule it is not matched.Response kept: 0 to 3Activity: 0.1× to 1000×Real wiring1 (reference)activity 1.0×Degree shufflesilent: 0activity 0.29×Weight shufflesilent: 0activity 0.39×Boundary keptpartly: 0.46activity 0.65×Sign shufflerunaway: 2.51activity 258×Gain checksilent: 0activity up to 1.7×Grey band: silent (≤ 0.10)Line: the real value (1)
Our own test, pre-registered on 1 Oct 2026 before any scrambled trial: 21 sugar neurons at 150 Hz, readout the feeding motor neuron MN9, Shiu et al. whole-brain model on FlyWire v783, one 1 s trial per shuffle. First scale: how much of the real MN9 response each scrambled brain keeps (dots: single shuffles; bar: mean). Second scale: how active the whole brain is, as a multiple of the real run (log scale). Silent arms leave a quiet brain; the sign shuffle's high MN9 rate comes with a brain about 258 times more active than real, so it is runaway, not a better reflex.
Numbers behind the chart
Our fair test, 1 Oct 2026: one 1 s trial per shuffle at trial seed 0; the real-wiring row is our run of 28 Sep 2026 (re-runs bit-exact). Activity is whole-brain spikes as a multiple of the real run's 13,828.
ArmWhat it keepsMN9 (Hz), mean ± sdPer shuffle (Hz)Response keptActivity vs realNeurons activeReading
Real wiringThe measured FlyWire v783 connectome85.2 ± 3.56 (5 trials)82, 89, 81, 86, 8811× (13,372–14,284 spikes)366–385reference
Degree-preservingKeeps partner counts0 ± 00, 0, 0, 0, 000.291×103silent
Weight shuffleKeeps who connects to whom0 ± 00, 0, 0, 0, 000.388×109silent
Boundary-preservingKeeps sensory and motor wiring39 ± 1.2238, 39, 39, 41, 380.4580.648×215partly
Sign shuffleSwaps excite and inhibit214.2 ± 43.27268, 188, 191, 253, 1712.514257.637×60,501runaway
Gain check on degree-preserving shuffle 3As degree-preserving, with synapses ×1.5, ×2 and ×30 at every gain×1.5: 0, ×2: 0, ×3: 000.37×, 0.50×, 1.74×193, 349, 4,257not matched (no gain within ±30% of real activity)

The pre-registered design

We wrote this down before the first scrambled trial. Every null keeps the stimulus, the readout, the neuron model, the number of connections and the total synapse count.

Arms of the fair test as pre-registered on 1 Oct 2026, 10:51 UTC.
ArmWhat is scrambledWhat is keptShuffles
D: degree-preservingThe target of every connection (the same null as our 28 Sep run)Every neuron's number of inputs and outputs; each connection's signed weight stays with its sending neuron5 (2 re-used from 28 Sep, 3 new)
W: weight shuffleThe synapse counts across all connectionsWho connects to whom; each connection's sign5
S: sign shuffleWhich neurons excite and which inhibit (each neuron keeps one sign: Dale's law)Who connects to whom; synapse counts; the number of excitatory and inhibitory neurons5
B: boundary-preservingThe target of interior connections onlyEvery connection out of a sensory or ascending neuron and into a motor, descending or endocrine neuron; every neuron's degrees5
G: gain checkAs D (shuffle 3), with synapses 1.5, 2 and 3 times strongerAs D; the real wiring is not re-tuned1 shuffle, 3 gains
  • Reading rules, fixed in advance: retention (null mean ÷ real mean, floor 0 Hz) of 0.10 or less means "the response needs the specific wiring this null destroys"; 0.90 or more means "this null keeps what the response needs"; anything between is "partly". Whole-brain activity is reported next to every arm, so that a silent brain and a mis-routed one can be told apart.
  • Matched gain: the gain whose whole-brain spike count is closest to the real run's (13,372–14,284 per trial), if it comes within ±30% of the 13,828 midpoint; otherwise "not matched".
  • Property checks run on every new null before simulating (degrees kept, boundary connections unchanged, graph and signs kept, one sign per neuron); a failed check would discard the trial. All passed. A trial running over 240 s would count as a timeout; none did.
  • Boundary sides come from FlyWire's cell-class annotation: 138,625 of 138,639 neurons matched (99.99%, above the 99% bar), giving 18,669 input-side and 1,485 output-side neurons.

Results by arm

Real wiring: 85.2 ± 3.56 Hz over 5 trials (82, 89, 81, 86 and 88 Hz) from our 28 Sep run. We re-ran trial seeds 0 to 2 and got 82, 89 and 81 Hz, bit for bit the same. Pooled over all 8 trials it is 84.75 ± 3.62 Hz, but that counts three trials twice, so it is shown only for reference. The sd of 3.56 is the sample sd of the five trials; our 30 Sep ledger showed 3.19, the population sd of the same values. A real trial fires about 13,400–14,300 spikes across the whole brain and activates 366–385 neurons, among them 26–29 motor and 64–66 descending neurons. Without a stimulus nothing fires (0 Hz).

Our fair test, per arm. Retention = null mean ÷ 85.2 Hz. Effect = the ledger's fixed rule (a gap beyond 2 combined standard errors).
ArmMN9 (Hz), mean ± sdPer shuffle (seed: Hz)RetentionEffect (rule)Neurons activeWhole-brain spikesActive motor / descending
D: degree-preserving0 ± 01: 0, 2: 0, 3: 0, 4: 0, 5: 00helps1034,021 (0.291× real)0 / 2.7
W: weight shuffle0 ± 0101: 0, 102: 0, 103: 0, 104: 0, 105: 00helps1095,363 (0.388× real)1 / 7.8
S: sign shuffle214.2 ± 43.27201: 268, 202: 188, 203: 191, 204: 253, 205: 1712.514worse60,5013,562,609 (257.637× real)99.8 / 1,033.4
B: boundary-preserving39 ± 1.22301: 38, 302: 39, 303: 39, 304: 41, 305: 380.458helps2158,967 (0.648× real)12.6 / 48.6
  • D and W: silent, and mis-routed. MN9 stayed at 0 Hz in all 10 shuffles. The brain goes quiet (about 30–40% of the real spikes, about 100–110 active neurons, almost no motor or descending neurons) although the sugar neurons still fire at about 150 Hz. Pre-registered reading: the response needs the specific wiring these nulls destroy.
  • B: partly. Keeping every sensory and ascending output and every motor, descending and endocrine input, and scrambling only the interior, keeps about 46% of the MN9 response with about two thirds of the real spikes. The reflex partly runs on boundary pathways that this null keeps; the interior wiring matters for the rest.
  • S: runaway. MN9 fires above the real rate only because the whole brain ignites: about 3.6 million spikes, about 60,500 active neurons and about 1,033 of the 1,299 descending neurons. By the mechanical rule this arm "keeps" the response (retention above 1) and counts as "worse" for the real wiring; read with its activity, it shows that the real assignment of excitation and inhibition is what keeps the response selective. The ledger's single metric cannot express this.

The gain check

  • ×1.5: MN9 0 Hz, 5,063 whole-brain spikes (63% below the target), 193 active neurons.
  • ×2: MN9 0 Hz, 6,865 spikes (50% below), 349 active neurons.
  • ×3: MN9 0 Hz, 24,024 spikes (74% above), 4,257 active neurons.

Result: not matched. No gain came within ±30% of the real spike count, so by the pre-registered rule the arm is "not matched" and the null cannot be called tuned equally. But the bracket is informative: from half to 1.7 times the real whole-brain activity, MN9 stayed at 0 Hz. A degree-preserving shuffle does not bring MN9 back by being made more excitable. This rests on one shuffle (shuffle 3); the matched-gain trials on shuffles 4 and 5 were planned as optional and not reached.

What it changes in the controls ledger

  • Our study row (byo-shiu-shuffle) now holds one arm per null type, with shuffles as samples. By the fixed rules it is mixed, with an unfair null and weak method (30 Sep: helps, partly fair, moderate). Our plan sets "tuned equally" only with a gain-matched arm; without it, the rule rates a null that loses as unfair (lower excitability could explain the loss), and "unfair" caps the method at weak. The sign-shuffle runaway makes the wiring effect mixed. This is the rule working as designed: the run measured more but did not yet settle the excitability question.
  • New study: Brain Runners: an untrained FlyWire brain survives 82.34 rows against 27.84 on shuffled wiring, but the same rule with no brain reaches 74.05 (held-out tracks, one shuffle). "Helps" by retention; the no-brain baseline is not beaten.
  • Counts: 32 studies (31 on 30 Sep), 24 with a wiring null (23): helps 10, no difference 5, worse 2, mixed 6, not yet scored 1 (30 Sep: 10, 5, 2, 5, 1); 8 studies have no wiring null. Reflex circuits: helps 3, mixed 1 (30 Sep: helps 4 of 4).

Deviations from the plan

None changes the design; all are listed for the reader.

  1. The annotation table has a combined "sensory_ascending" class (581 neurons) that the pre-registration did not name. It was treated as input side from the start, before any boundary trial.
  2. The run was split across three sessions by interruptions. One degree-preserving trial (shuffle 3) was lost in flight and re-run; in the last session no new trial was started after 30 minutes. All 18 core trials (5 shuffles in every arm) and the 3 gain trials finished; the optional trials (matched-gain trials on shuffles 4 and 5, a stimulus-and-readout-keeping null, shuffles 6 to 10) were not started. This is the only shortfall against the pre-registered queue.
  3. The restarts used small wrappers that call the unchanged trial code in the pre-registered order.
  4. The gate trial (real wiring, seed 0) ran after the first real re-runs and the first boundary trial. It gave the pre-registered 82.0 Hz, bit-exact, so no result depends on the order.

Limitations of the fair test

One reflex (sugar to MN9), one model (Shiu et al. leaky integrate-and-fire, upstream parameters), one trial per shuffle at one trial seed, 5 shuffles per arm, 1 s trials. The gain check ran on one shuffle and found no matched gain. The MN9 rate is the only metric, and it cannot tell selective drive from global ignition. The boundary split depends on FlyWire's cell-class annotation. Two of the five degree-preserving shuffles come from our 28 Sep run (same code, trial seed 0).

Chart: a brainless autopilot plays Doom as well as the fly brain

We fetched the Doom control study's results file once more and charted every row. DOOMFLY survives 53.34 s (n 20; the file's 95% interval 49.73–56.78 s; sd 8.40 s from our ledger), the matched autopilot 51.93 s, and the best of 12 autopilot settings (spray_t1_f20) 53.44 s. Shuffled wiring survives 5.48, 9.29 and 5.48 s on three seeds (6.75 s pooled), no connections 5.50 s and no vision 10.46 s. All 5 numbers our ledger uses match the file.

A brainless autopilot plays Doom as well as the fly brainSeconds survived in DOOMFLY's Doom arena, from gabrycina/doom-fly-control@6b22922 results/a2/summary.json: DOOMFLY, real wiring 53.34 s (95% interval 49.73–56.78); Autopilot matched to the fly 51.93 s (95% interval 45.32–57.28); Best of 12 settings (spray_t1_f20) 53.44 s (95% interval 48.81–57.31); All 12 autopilot settings 5.28–53.44 s; No vision (blind) 10.46 s (95% interval 8.84–12.83); Shuffled wiring, 3 seeds 6.75 s; No connections 5.5 s (95% interval 5.24–5.76). Shuffled-wiring seeds: 5.48, 9.29, 5.48 s. A brainless autopilot matched to the fly's average commands survives almost as long as the fly brain, and the best of 12 autopilot settings slightly longer.0102030405060DOOMFLY, real wiringDOOMFLY: real FlyWire wiring (real): 53.34 s, 95% interval 49.73–56.78, n 2053.34 sn 20Autopilot matched to the flyMatched spray autopilot (no brain) (spray_matched): 51.93 s, 95% interval 45.32–57.28, n 2051.93 sn 20, no brainBest of 12 settings (spray_t1_f20)Spray autopilot setting t1_f20 (spray_t1_f20): 53.44 s, 95% interval 48.81–57.31, n 1653.44 sn 16All 12 autopilot settingsSpray autopilot setting t0_f0 (spray_t0_f0): 6.11 s, 95% interval 5.63–6.67, n 16Spray autopilot setting t0_f10 (spray_t0_f10): 8.59 s, 95% interval 7.65–9.57, n 16Spray autopilot setting t0_f20 (spray_t0_f20): 9.63 s, 95% interval 8.47–10.89, n 16Spray autopilot setting t1_f0 (spray_t1_f0): 5.68 s, 95% interval 5.35–6, n 16Spray autopilot setting t1_f10 (spray_t1_f10): 26.7 s, 95% interval 21.91–31.43, n 16Spray autopilot setting t1_f20 (spray_t1_f20): 53.44 s, 95% interval 48.81–57.31, n 16Spray autopilot setting t2_f0 (spray_t2_f0): 5.46 s, 95% interval 5.18–5.76, n 16Spray autopilot setting t2_f10 (spray_t2_f10): 9.65 s, 95% interval 8.29–10.99, n 16Spray autopilot setting t2_f20 (spray_t2_f20): 21.09 s, 95% interval 16.55–26.08, n 16Spray autopilot setting t4_f0 (spray_t4_f0): 5.28 s, 95% interval 5.06–5.53, n 16Spray autopilot setting t4_f10 (spray_t4_f10): 7.6 s, 95% interval 6.73–8.52, n 16Spray autopilot setting t4_f20 (spray_t4_f20): 11.41 s, 95% interval 9.39–13.59, n 165.28–53.44 sNo vision (blind)No vision (blind) (blind): 10.46 s, 95% interval 8.84–12.83, n 810.46 sn 8Shuffled wiring, 3 seedsShuffled wiring, seed 0 (shuf0): 5.48 s, 95% interval 4.99–5.97, n 4Shuffled wiring, seed 1 (shuf1): 9.29 s, 95% interval 7.1–11.47, n 4Shuffled wiring, seed 2 (shuf2): 5.48 s, 95% interval 4.99–5.97, n 46.75 spooled, n 12No connectionsNo connections (noconn): 5.5 s, 95% interval 5.24–5.76, n 85.50 sn 8Seconds survived (mean; whiskers: the file's 95% interval)fly brainautopilot, no brainbroken brain Seconds survived in DOOMFLY's Doom arena, from gabrycina/doom-fly-control@6b22922 results/a2/summary.json: DOOMFLY, real wiring 53.34 s (95% interval 49.73–56.78); Autopilot matched to the fly 51.93 s (95% interval 45.32–57.28); Best of 12 settings (spray_t1_f20) 53.44 s (95% interval 48.81–57.31); All 12 autopilot settings 5.28–53.44 s; No vision (blind) 10.46 s (95% interval 8.84–12.83); Shuffled wiring, 3 seeds 6.75 s; No connections 5.5 s (95% interval 5.24–5.76). Shuffled-wiring seeds: 5.48, 9.29, 5.48 s. A brainless autopilot matched to the fly's average commands survives almost as long as the fly brain, and the best of 12 autopilot settings slightly longer.Fly brain: 53.34 sAutopilot: 51.93 sBest autopilot: 53.44 s12 settings: 5.3–53.4 sNo vision: 10.46 sShuffled wiring: 6.75 sNo connections: 5.50 s0 s30 s60 sLine: the fly brain, 53.34 s
Source: gabrycina/doom-fly-control@6b22922 results/a2/summary.json (the doom-fly-control study's own results, commit 6b22922); all 5 numbers our ledger uses match the file. The matched autopilot always turns, walks and shoots at the fly brain's average rates; the 12 settings vary its turning and firing. The autopilot's constants come from the fly's average output, so this shows a constant policy reproduces the play, not that the brain does nothing. DOOMFLY's spread (sd 8.40 s, n 20) is from our ledger, recomputed from the study's per-episode file.
All 19 rows of the source file
Every number from results/a2/summary.json in gabrycina/doom-fly-control at commit 6b22922 (fetched 1 Oct 2026; sha256 3508cd4f…67e86). The study author's measurements; we read the file but did not re-run the games.
ArmKey in the fileSeconds survived (mean)95% intervalEpisodesKills (mean)
DOOMFLY, real wiringreal53.3449.73–56.782012.1
Matched spray autopilot (no brain)spray_matched51.9345.32–57.282011.75
Spray autopilot setting t0_f0spray_t0_f06.115.63–6.67161.625
Spray autopilot setting t0_f10spray_t0_f108.597.65–9.57161.625
Spray autopilot setting t0_f20spray_t0_f209.638.47–10.89161.875
Spray autopilot setting t1_f0spray_t1_f05.685.35–6160.9375
Spray autopilot setting t1_f10spray_t1_f1026.721.91–31.43166.6875
Spray autopilot setting t1_f20spray_t1_f2053.4448.81–57.311613.25
Spray autopilot setting t2_f0spray_t2_f05.465.18–5.76160.25
Spray autopilot setting t2_f10spray_t2_f109.658.29–10.99162.625
Spray autopilot setting t2_f20spray_t2_f2021.0916.55–26.08165.0625
Spray autopilot setting t4_f0spray_t4_f05.285.06–5.53160.25
Spray autopilot setting t4_f10spray_t4_f107.66.73–8.52161.5
Spray autopilot setting t4_f20spray_t4_f2011.419.39–13.59163.125
Shuffled wiring, seed 0shuf05.484.99–5.9740.25
Shuffled wiring, seed 1shuf19.297.1–11.4741.5
Shuffled wiring, seed 2shuf25.484.99–5.9740.25
No connectionsnoconn5.55.24–5.7680.375
No vision (blind)blind10.468.84–12.8382.5

What people argued about this week

The scan covered new items since the end of the last scan (30 Sep, 09:45 UTC). Two claims were promoted, both under a catch-up rule for videos with at least 100,000 views from the last 14 days that our earlier scans missed; no new claim met the normal bar.

  • Hacker News: 6 queries, 7 stories and 60 comments, no fly-brain story. Google News: 5 queries, 29 items; one new item, a press-release story, is not a behaviour claim. Reddit refused our scripts (HTTP 403), as before. GitHub: 18 queries, no new repository with 20 or more stars (the best has 2). Dataset pages: no new release.
  • YouTube, sorted by date (17 queries, 108 unique fly-related hits) against sorted by views this week (8 queries, 38 unique) and this month (32): only 4 of the 38 week-by-views hits were also among the date-sorted hits. The views sort found every high-view item (Gorilla Tag 527,756; a cartoon skit 1,040,792; "Fly's Brain Played Video Games" 214,311; "THEY TRAPPED A FLY'S BRAIN" 138,478); the date sort found only new low-view uploads. The month view found five videos over 1 million views from 10 to 19 Sep, mostly skits without a mechanism claim we could check.
  • Watch pages: 68 view counts read, 70 refused (HTTP 429). YouTube now shows our fetcher a sign-in player; counts, dates and descriptions are still read from the page data.
  • Re-checked: the Pokémon stream (no new search hit, stays U); the Rainbow Six claim (a new short from the same media team, 4,346 views, not reviewed; stays U); Vinesauce's FlyLeno stream grew to 42,125 views.

Verdicts

Gorilla Tag: U

"I Taught a Fly How to Play Gorilla Tag (seriously)" (SuperCatCrazeGT, uploaded 26 Sep): 527,756 views on 1 Oct, against 415,612 a day earlier. The description says the fly learned "using its open source brain", but names no model, data or code. Its only link is a linktr.ee page with 76 links (game maps, socials, sponsors), none of them code; four GitHub searches returned nothing. The channel's earlier video trained an ordinary AI on Gorilla Tag. We could not watch the video, so its on-screen claims are not recorded. Verdict row

Roblox fight: U

"I Made Two Real Flies Brain Fight Each other in JJS" (Evoke, uploaded 21 Sep): 139,364 views, never seen by our earlier date-sorted scans. Only Discord and Roblox-group links; three GitHub searches found nothing related. Any coupling to a Roblox fighter would be the creator's own client script. Verdict row

Tracking

  • Repositories: of 84 tracked repositories, 4 changed, 79 were unchanged and 1 is gone or private: FlyDoom by eganeganegan returns HTTP 404. We keep its record and re-check before removing it. Changed: acamilo-flybrain (24 commits: run infrastructure and status notes), fly-ai (9 commits, not read yet), fly-cartpole (25 commits) and Open Fly (1 commit).
  • Links: 219 URLs: 180 ok, 12 redirects, 19 blocked by bot checks, 8 errors (the FlyDoom code and source pages, four YouTube watch pages that were rate-limited, and two bot blocks).
  • Videos: 49 of 49 available.
  • Re-grades: none. fly-cartpole was re-checked from its results summary at the new commit: 392.5 ± 103.1 steps against 390.3 ± 96.4 for shuffled wiring (the author's permutation p = 0.4771), against 233.8 vs 218.3 (p = 0.1953) at the commit our ledger uses. Same verdict, no difference. We have not read the protocol changes behind the new numbers, so the ledger keeps the older numbers until the next check.

New entries

Every entry was graded from files we opened, or from what the claim publishes when there is nothing to open. The numbers are the authors' own.

6 new catalogue entries on 1 Oct 2026.
EntryKindGradeControlsDeciding evidence
Brain RunnersGameAWiring null: helps; no-brain baseline not beatenAn untrained FlyWire brain plays a lane-runner on held-out tracks: 82.3 rows, 27.8 on shuffled wiring, 74.1 for the same rule with no brain. One shuffle, no spread.
Doodle FlyGameBNone (its two controls change the eyes, not the wiring)The whole FlyWire brain plays a Doodle-Jump-style game in the browser; target position drives eye neurons, steering neurons press the buttons.
FLY67Browser demoBNoneThe whole FlyWire brain in a browser tab drives an animated 3D fly that feeds, grooms and escapes; inputs and outputs are hand-set, and one gesture is a labelled script.
Fly x JevBrowser demoCNoneMaleCNS wiring turns a robot's compound eye into visual signals and a learned memory, but an external commercial model picks every action.
Fly plays Gorilla TagGame (video claim)UNoneSee the verdict above.
Two fly brains fight in a Roblox gameGame (video claim)UNoneSee the verdict above.

Left out: a repository that uses FlyGym's name was confirmed as a plain copy of FlyGym and excluded; a second one adds its own work (a bridge from the BANC connectome to FlyGym's legs) and is left for the next check; a later copy of Fly x Jev was excluded. None is linked. A research project with its own controls and several new low-star repositories were not opened in time. FLY67's README links a demo video file that returns "not found"; we do not link it.

51 videos, 2 new: the creators' own uploads of the Gorilla Tag and Roblox fight videos, both grade U and described from their titles, descriptions and view counts only. The README videos of the other four new entries were checked: none was added (no video, a GIF only, or a file that returns "not found"). All 49 earlier videos were available.

Build your own: re-tested

Pass. From an empty folder on Linux (Python 3.11.2, 2 CPUs): the environment fallback without pip worked as written, the pinned install finished in 37 s, the partial clone at 91bdd1e7 in 9 s, both data checksums match the published table, the first sugar trial gave 82.0 Hz (pass band 75–95 Hz), bit-exact with our first run, and seeds 1 and 2 gave 89 and 81 Hz, also bit-exact. The project's tests: 6 passed in 56.5 s. The fair test's own unit tests: 8 passed. No page fixes were needed. On PyPI, the pins still install on Python 3.11 (Brian2 2.9.0 pinned; the latest is 2.10.1, which needs a newer Python). Re-test details · Try a fair control yourself

Method

  • Lead scan: Hacker News, Google News, GitHub, dataset pages and YouTube, now searched both by upload date and by view count (this week and this month), plus the video pages of all fly-related hits.
  • Tracking: remote repository heads for every tracked repository, one-at-a-time shallow clones of the changed ones, and a check of every entry link and every video.
  • Fair test: the pre-registered design above, run with one trial worker, one fresh process per trial, and every trial written to its own file with property checks, activity and versions. The analysis computed per-arm means, retention and the ledger's effect class with fixed rules.
  • Controls ledger: rebuilt with the same fixed rules; the data file validates against its unchanged schema.
  • GitHub data: anonymous access; fields reused from earlier checks keep their true dates in the data file.

Limitations

  • The fair test's limits are listed above.
  • Every ledger number except our own is the authors'.
  • The two new verdicts rest on watch-page data only: we could not watch the videos.
  • About half of the watch-page view counts were refused (HTTP 429). X and TikTok cannot be searched without a login, and Reddit refuses scripts.

Next

  • "Add a body": a candidate list of descending neurons for walking, turning, escape and grooming (present in the model; their published basis is listed but not yet verified), and the FlyGym derivative with a BANC bridge as a design input.
  • Finish the fair test cheaply: a gain between ×2 and ×3 that hits the real spike count, matched-gain trials on two more shuffles, and the stimulus-and-readout-keeping null. A selectivity measure (MN9 rate relative to whole-brain activity) next to the MN9 rate would need a new pre-registration.
  • Carried flags: confirm whether FlyDoom was removed; read fly-ai's 9 new commits; read fly-cartpole's new protocol before updating its ledger numbers.

Main sources

Search Shaduf

Search published pools, pages, reports, and evidence.