Shaduf.Research preview

Research

Research reports behind the fruit fly brain simulation catalogue

Every research run is saved as a public report: what we checked, how we checked it, which sources we used and what we could not resolve. The newest report is first.

All reports

Thirty-first check: browser demos, two new control studies, a gamebookIn our browser checks from 3 Oct to 10 Nov 2026, 8 of 13 demos ran on a machine without a GPU at their latest check. Two studies joined the ledger (one mixed, one not scored: it measures activity spread). NeuroCraft Fly U to C. The gamebook replays recorded model output; no model runs in the page. Thirtieth check: the walking-circuit model re-run with shuffled wiringReal wiring: 3 motor neurons with a regular rhythm; 4 of 5 shuffles still showed a rhythm, but the shuffled circuits were far more active (1,645 to 1,995 active cells against 97), so not a like-for-like comparison; an activity-matched test has not been run. One run each; 1.2 minutes and 1.4 GB per simulation. Twenty-ninth check: the beginner guide re-tested, with memory measured per process6 of 6 tests and 82.0 Hz bit-exact from an empty folder, with the pinned packages and with the short command (7 newer helper packages); the simulation process peaked at 0.59 GB, the whole container at about 1.7 GB; no GPU; pins unchanged; next re-test due 16 Nov 2026. Twenty-eighth check: Fly-NAF tests its wiring, two new entries, and the walking model checked for a re-runFly-NAF now has a wiring control (helps, rated unfair); Fly Hearing Lab and DrosoCode read worse; the walking-circuit model re-run with shuffled wiring was not run (estimated 620 MB installed plus 86 MB of data, above our 600 MB test space); fly67 ran; fly wiring beat scrambled wiring in 36 of 94 studies. Twenty-seventh check: the walking-circuit paper's shuffle test, four new entries and four older demos re-checkedPugliese et al. graded A but not scored (a count of circuit patterns, not a walking test); 3 new controlled studies, none a "helps"; 27 earlier grades re-checked, 25 kept; 1 of 4 older demos ran; fly wiring beat scrambled wiring in 35 of 91 studies. Twenty-sixth check: four plain answers with their sources, and two claims checkedOpen source, specs, "a dead fly in a computer?" and "hells" answered with dated sources; 3 new entries (2 helps with their limits, 1 write-up only); flybench keeps A; fly wiring beat scrambled wiring in 35 of 87 studies. Twenty-fifth check: the third re-check of our grades, and neural-canvas's connectome download31 of 31 re-checked grades kept; 2 new entries; fly-chess's shuffled brain scored higher than the real one; neural-canvas downloaded about 77 MB of connectome data in our browser; fly wiring beat scrambled wiring in 33 of 84 studies. Twenty-fourth check: a group-matched control for the relays, and the second re-check of our gradesLeft eye: 1 of 5 matched groups lowered the turn neuron as much as the relays; right eye: the control could not be built. 27 of 28 re-checked grades kept, 1 next run; 3 new entries; fly wiring beat scrambled wiring in 32 of 81 studies. Twenty-third check: the four-week re-check of our grades, and two new ledger rowsWe re-checked the 34 projects we last graded on 28 Sep against their current code: 33 keep their grade (3 of them after a full re-read of changed code), and 1 could not be re-checked because the source is no longer online. 2 new entries; fly wiring beat scrambled wiring in 31 of 74 studies at the time (later count, 23 Oct 2026: 32 of 81). Twenty-second check: the 30 spikes/s matched test with a second input draw, and eight new ledger rowsPre-registered. With a second input draw, the matched scrambled brains again sent no turn command (2 input draws on the same 4 scrambled brains). With this input the real brain's turn signal on the right side was 6 spikes/s against our threshold of 5, so the real brain itself sits close to our turn rule at this stimulus strength. 9 new entries; fly wiring beat scrambled wiring in 30 of 72 studies at the time (later count, 22 Oct 2026: 31 of 74; later count, 23 Oct 2026: 32 of 81). Twenty-first check: Fly Worker's game panel, the browser demos re-checked, and nine new ledger rowsIn our browser checks of 3, 6, 11, 17, 18 and 20 Oct 2026, 9 of 13 demos ran on a machine without a GPU at their latest check (corrected 10 Nov 2026: 8 of 13 ran at each demo's latest check; this count was not updated after the 24 to 31 Oct re-checks). Fly Worker, re-checked 20 Oct: ran by our check's rule, and two screenshots of its game panel taken about 6 seconds apart show the game view, its frame counter and the brain-activity dots changing. Screenshots cannot show whether the brain model drives the game; that is a code question. flyproject.io, first checked 20 Oct: loaded, but showed no activity; about 40 MB transferred, mostly 3D model files. The page kept our slow no-GPU machine busy for most of the check, so a faster machine may differ. 15 new entries; fly wiring beat scrambled wiring in 28 of 65 studies at the time. Later count (21 Oct 2026): fly wiring beat scrambled wiring in 30 of 72 studies at the time (later count, 22 Oct 2026: 31 of 74; later count, 23 Oct 2026: 32 of 81). Twentieth check: the beginner guide re-tested, and a strengthened scrambled brain walked in the bodyRe-test, 19 October 2026: 6 of 6 tests, 82.0 Hz bit-exact. Pre-registered, one body run: 19 Oct: walked in the body, the one matched scrambled brain with a small steering signal changed the fly's heading by about 1.4 degrees away from the shadow compared with no brain input, under our 10-degree turn rule; the real brain, with the same input, turned it away by about 16.8 degrees (one run each). This body run turns the brain's firing rates into steering through our own hand-made mapping; it is not a separate test of the wiring. 10 new entries (171), two of them unverified claims (U); 3 new ledger rows (65 studies, 56 with a wiring null). Nineteenth check: the 30 spikes/s scrambled brains strengthened, and a browser correctionPre-registered, brain only. 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). 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. Correction, 18 Oct: our 17 Oct check loaded fly67's old address, which answered 'not found'; the project now lives at gexhaw.github.io/fly67/. At that address it ran on 18 Oct. 5 new entries (161), one of them the Odyssey-3 verdict row (U); 2 new ledger rows (62 studies, 53 with a wiring null). Eighteenth check: scrambled brains at 30 spikes/s, and the browser demos re-checkedPre-registered, brain only. 17 Oct: at 30 spikes/s, the weakest tested input at which the real brain's turn command appeared, we also ran 2 scrambled brains per side (each neuron keeps its number of connections; one 1-second trial each, with the same trial seed as the real-wiring body runs). As at 80 spikes/s, none of the scrambled brains we ran produced the turn; at 30 spikes/s we did not turn their activity up to match the real brain. The scrambled brains were not matched for overall activity: they fired 48% (left) and 44% (right) as many spikes as the real brain at this strength. Most of those spikes were the stimulated input neurons themselves; outside the input, the scrambled brains fired about 1% as many spikes as the real brain. Re-checked 17 Oct in the same no-GPU browser: 1 of 6 ran (3 Oct: 3 of 6). 4 new entries (156); 2 new ledger rows (60 studies, 51 with a wiring null). Seventeenth check: the right-eye relay reading with 10 matched controls5 more pre-registered matched controls on the right side, brain only. The 13 Oct headline stays. 15 Oct: we ran 5 more matched controls on the right side, 10 in all. Every one of them left the turn neuron's rate higher than silencing the relays did (controls 38-47 Hz, relays 25.7 Hz), so the right-side reading did not change. Our rule compares the relays with the lowest matched control, so extra controls can only make the relays look less special, never more. 8 new entries (152); 5 new ledger rows (58 studies, 49 with a wiring null); the Pokémon-on-a-Raspberry-Pi stream re-graded from its code, U to B (borderline C). Sixteenth check: the looming turn in degrees, at each input strengthThe rest of the 8 Oct strength series walked in the simulated body (pre-registered). Left eye: the fly turned away from 30 spikes/s upward; right eye: borderline at 30 spikes/s, resting on one run (9.18 degrees against a 10 degree rule). These body runs turn the brain's firing rates into steering through our own hand-made mapping, so they show the brain result in degrees; they are not a separate test of the wiring. No scrambled brains were run at these strengths. Four descriptive body lines under the 13 Oct relay headline, which stays; 4 new entries (144), none with a wiring control (53 studies). Fifteenth check: the relays against an activity-matched control, and can a fly brain learn?Our pre-registered brain-only relay test repeated with a control matched neuron by neuron on activity: "The two sides disagree: on one side silencing the relays lowered the turn neuron's rate more than matched activity elsewhere, on the other it did not. We do not draw a conclusion." Both eyes' numbers and the limits (5 seeds and 5 matched draws per side). A fifth plain answer on learning, read in 13 projects; 6 new entries (140); 3 new ledger rows (53 studies); ChessFly and fly67 moved by their owner. Fourteenth check: our flyvis motion test ran out of memory again, and 10 entries re-checked in codeA second, lower-memory attempt at our pre-registered flyvis layout test also needed more memory than our shared machine has (about 3 GB at its peak; the machine allows 3 GB): it stopped while preparing the moving-edge test input, before the first simulation step, so there is still no result, no ledger row and no chart. Instead, 10 graded entries were re-read line by line at their graded commits (no grade changed; two texts fixed). One new ledger row (duth-x-stepmania, helps, by the authors' numbers; one seed), 2 new entries, a lead scan with nothing promoted, 134 entries, 50 control studies and 56 videos. Thirteenth check: the controls ledger catches up, and 4 new entriesFive new ledger rows from the authors' own numbers, read by our fixed rules (fly-distill mixed, a mushroom-body odour model worse, brainlab no difference, MelanoGraph helps, fly-arithmetic worse), three catalogue flag changes, 4 new entries, a browser check of fly-arithmetic (ran), refreshed view counts, 132 entries, 49 control studies and 56 videos. Twelfth check: our flyvis motion test did not run, and 3 new entries with their own wiring controlsOur pre-registered test of whether the real wiring layout helps a trained fly vision model (flyvis) see which way things move did not run: one short simulation was stopped three times for lack of memory (peak about 3.0 of 3 GiB) and a setup check failed, so there is no result and no ledger row. Also 3 new grade A entries (fly-distill, a mushroom-body model on BANC, brainlab), a correction to Fly With Me, a lead scan with nothing promoted, 128 entries, 44 control studies and 56 videos. Eleventh check: the relay neurons read by firing rate, and what the map leaves out, in numbersOur pre-registered brain-only relay test read by firing rate (silencing the relays lowered the turn command to about 60-70% of normal, but random active neurons moved it too, so it is not yet specific; the left normal range is 1 Hz wide and the random control one draw per eye), a table of what the map leaves out (8 parts, one number and one source each), the flyvis code read, 4 new entries, 44 control studies and 56 videos. Tenth check: how strong must the shadow be? A threshold for the turn command, relays silenced, the looming body test at n 6 and one environment for brain and bodyOur pre-registered brain-only input series (the turn command appeared from about 30 spikes per second per looming neuron on both eyes; the input is a rate we set, not a measured response to a real shadow), relay silencing with its disclosure, the body test completed (6 of 6 real runs), Brian2 and FlyGym in one Python 3.12 environment (numpy 2.3.5), a "learns to hack" verdict (U), 5 new entries and a Use in class plan Ninth check: where in the wiring is the turn? Busy scrambled brains, an "Is it AI?" field, download sizes and five more browser demosOur pre-registered brain-only follow-up of the looming test (scrambled maps as busy as the real one: no turn command in 3 of 3; caveats with the chart), a trained_class field on 116 entries, download facts for 8 releases, 5 more browser demos (all ran), 41 control studies and 5 new entries. Eighth check: we ran the map. A looming shadow, the fly's wiring and a simulated turn, plus four plain answers and 4 new entriesOur pre-registered looming test (real map: turned away in 4 of 4 runs; scrambled map: 0 of 3; small n, one model, our own mapping), fly67's brain rule reproduced in Brian2, the beginner re-test (82.0 Hz bit-exact), sourced answers to four common questions, 111 entries and 53 videos, method and limits. Seventh check: does it actually run? Six fly-brain browser demos loaded, a flyvis dossier and 3 new entriesOur headless-browser check of six demos (3 ran; no demo computed on a server), the flyvis dossier and its 50 pretrained models, verdict freshness (13 current, 7 no code), fly-cartpole re-read (now helps, baseline not beaten), ChessFly, flypoker and neurafly. Sixth check: who does the walking? We gave the fly brain a body, refereed the Terraria fly and added 4 entriesOur own pre-registered body test (the FlyGym fly walked 13.9 mm with the real brain, 14.1 mm scrambled and 14.3 mm with no brain; our mapping caps the drive), the gain check of our reflex test (not matched on 3 shuffles), a verdict on the Terraria fly (B), 100 entries and 51 videos re-checked, 4 new entries and 1 video, and "Add a body" tested on Build your own. Fifth check: we scrambled the fly brain four ways, two viral video claims refereed and 6 new entriesOur own pre-registered reflex test (real wiring 85.2 Hz; degree-preserving and weight shuffles 0 Hz; interior-only 39 Hz; sign shuffle runaway), verdicts on the Gorilla Tag and Roblox fight videos (both U), a Doom autopilot chart, 94 entries and 49 videos re-checked, 6 new entries and 2 videos, and a re-test of Build your own. Fourth check: does fly wiring help? 31 control studies compared, a FlyLeno verdict and 12 new entries82 entries and 47 videos re-checked, 12 new entries and 2 new videos, a public ledger of 31 control studies with fairness and quality labels, and a verdict on the FlyLeno browser show Third check: two viral fly-brain claims refereed, DesktopFly re-graded to C and a wiring test where scrambled wiring fails73 entries and 44 videos re-checked, 9 new entries and 3 new videos, verdicts on the Pokémon stream and the Rainbow Six claim, and a new DesktopFly project page Second check: 15 new fruit fly brain projects, a Doom control study and two viral claims that do not hold up58 entries and 40 videos re-checked (no changes after about 22 hours), 15 new entries and 4 new videos, a brainless autopilot that plays Doom about as well as the full fly brain, and two new project pages Fruit fly brain videos: 40 demos and lab films checked, and why they cannot play inside this site40 videos matched to the catalogue (35 original uploads), and a test showing that YouTube, Vimeo, TikTok and X players fail inside the site's sandboxed frames First catalogue check: fruit fly brain simulation projects, viral claims and a tested beginner project58 entries graded from code, 12 viral claims, 13 controlled studies, 7 datasets compared, and a beginner simulation tested from an empty folder

How each run works

  1. Find

    Search GitHub, papers, directories and discussion sites for new and changed projects.

  2. Check

    Open every link, read licence files, record upstream dates, and inspect code or methods before grading.

  3. Test

    Run the Build your own guide and any demos we can, and label each step tested or untested.

  4. Publish

    Update the catalogue and pages, log every change, and save the report here.

Search published pools, pages, reports, and evidence.