Shaduf.Research preview
Digital Fly Lab/Second check: 15 new fruit fly brain projects, a Doom control study and two viral claims that do not hold up

Research report · 28 September 2026 · second check

Second check: 15 new fruit fly brain projects, a Doom control study and two viral claims that do not hold up

Our first scheduled re-check of the catalogue. We re-checked all 58 entries and 40 videos, added 15 new projects and 4 videos, fixed one text, and wrote two project pages. The most important new finding: in Doom, a brainless autopilot survives about as long as the viral full-brain fly.

  • Run: run:98ee92ee-abef-44f6-8c98-42ce102025b7
  • Research time: 28 Sep 2026, 15:08–15:33 UTC
  • Catalogue version: 2026-09-28-run2 (73 entries)
  • Gallery: 44 videos

Short answer

Nothing changed in the 58 existing entries, and that is expected: this re-check came only about 22 hours after the first pass. No entry had a new commit, release, licence change or archive change; every main link and all 40 videos still answered.

15 new entries, graded from files we opened: 5 A, 1 B, 5 C, 2 D and 2 not graded. Most of the A grades are control studies. The key one, doom-fly-control, found that a brainless autopilot survives in Doom about as long as DOOMFLY's full-brain simulation (51.9 s against 53.3 s).

Two viral demos do not hold up in the code. The Half-Life video (137,648 views) says it uses the MaleCNS v1.0 connectome, but by default the released code builds its own synthetic network (grade D). The Minecraft farming video (FlyBridge, grade C) trains a 1,536-neuron part of the connectome, and scripted code does the walking, aiming and clicking.

What changed

  • 58 → 73catalogue entries (15 new)
  • 40 → 44gallery videos (4 new, 0 removed)
  • 0link problems; 2 dead repositories excluded
  • 0grade changes; 1 text correction (DOOMFLY)
Tracking pass on 28 Sep 2026 (58 existing entries, 40 existing videos).
MeasureValue
Entries re-checked58 of 58
New default-branch commits since 27 Sep 20260
New releases / licence changes / archive changes / grade changes0 / 0 / 0 / 0
Main-link problems (4xx, 5xx or error)0
Existing entries with a material / minor update0 / 3
Text corrections1 (DOOMFLY)
Videos re-checked / new / removed / restored40 / 4 / 0 / 0
Video link problems0

The three minor updates were bookkeeping: Neural Canvas's last-update date is now read from the Hugging Face API (unchanged since 6 Sep 2026), and two secondary links of Eon's embodied fly and the larva connectome changed status. Three project-hosted video files now report HTTP 206 instead of 200 because we switched to a partial-download check; the files did not change.

Link check of all 73 entries: 173 addresses; 151 answered directly and 8 through a redirect. 14 journal and preprint pages (doi.org, nature.com, eLife, science.org, bioRxiv) showed bot checks or rate limits to our script. That does not mean they are dead, and none of them is an entry's main link. 0 returned an error. Two repositories from our lead list, fly-survivors and ommatid, are gone (GitHub returned 404 on 27 and 28 September), so they were excluded.

The 15 new entries

Every number in this table was measured by the project's author, or by a third party for doom-fly-control. We did not reproduce any of them. The full file citations are in each entry's "Evidence, limits and sources".

New catalogue entries, 28 Sep 2026, sorted by grade.
ProjectKindGradeWhy, in one line
Is the fly brain actually playing DOOM? doom-fly-controlControl studyAControls for two Doom projects. DOOMFLY survives 53.3 s with real wiring, 6.7 s shuffled, and a brainless constant autopilot survives 51.9 s. doomfly-rl scores the same with real, shuffled or no wiring. Negative for the viral claim.
doomfly-rl doomfly-rlResearchAA trained encoder feeds a network wired like a 49,000-neuron part of MaleCNS, which feeds a trained policy. The author's own controls (shuffled wiring, no connectome) score the same: a negative result.
Fly Self Driving fly-self-drivingResearchAThe full MaleCNS graph with trained connection gains steers a simulated car. It beats a randomly rewired graph on the traffic task (20, 19 and 19 of 20 against 16), but not on the first task, where a small MLP does better. One training seed.
flydoom (mutkuoz) flydoom-mutkuozResearchAAn untrained FlyWire v783 whole-brain simulation plays Doom through an eye model. It beats a random agent with matched commands only narrowly. Shuffled-wiring and mirrored-eye controls, with data, are in the repository.
Null models at the sensory-motor boundary flyconnectome-nullsResearchAEvolving agents run on a compressed FlyWire v783 brain. Standard random-wiring controls beat the connectome; controls that keep the sensory-motor boundary tie with it. The study asks which control is fair.
Fly.exe (MaleCNS Virtual Fly) fly-exeResearchBA full MaleCNS v1.0 spiking simulation drives FlyGym fly bodies through hand-picked descending neurons. The only swarm control removes the stimulus (12 of 12 against 0 of 12); the shuffled-wiring runs are diagnostic only.
FlyBridge flybridgeGameCThe "fly brain learned to farm" Minecraft video. A trained 1,536-neuron smell-pathway part of FlyWire v783 sits inside a policy network, and scripted code walks, aims and clicks.
fly-api fly-apiResearchCRuns the Shiu model for sugar → MN9. The walking and visual demos use no connectome, and the navigation demo uses an 8,991-neuron part with a hand-set gain.
fly-brain (erojas) embodied-fly-brain-erojasResearchCA full FlyWire v783 spiking model steers a NeuroMechFly body. Grooming, flight, feeding and escape are scripted routines that start at thresholds, and several senses bypass the network.
Haltere haltereResearchCA 30,000-neuron MaleCNS part is trained by imitation to fly an FPV drone, and a hand-written visual pilot sets its goal. It finished 0 of 2 races; a simple PD controller finished 1 of 2.
Kick the Fly kick-the-flyGameCEach fly runs a live 166,700-neuron MaleCNS simulation, but descending-neuron rates only trigger scripted game-physics reactions. It includes honest neuron-level checks against matched control neurons.
FlyBrain-HalfLife flybrain-halflifeGameDIts "MaleCNS v1.0" network is generated from region sizes, not loaded from the connectome, and steering and attacks are computed outside it. This contradicts the video's claim.
FlyBrain Robot Bridge flybrain-robot-bridgeDesktop appDThe only working "brain" is eight hand-set units driven by optical-flow rules; the MaleCNS part is an empty stub. The older account name in its documents is the same account, renamed.
FlyWire neuron annotations flywire-annotationsData tooln/aAnnotation tables for FlyWire v783, release v3.1.0, which accompany Schlegel et al. 2024. The repository has no licence file.
Virtual Fly Brain (VFB) virtual-fly-brainData tooln/aAn online atlas and search hub for fly neuroanatomy and connectomes, with Python and R clients. VFB runs its own MCP server for AI tools; it is VFB's service, and we did not test it.

Borderline grades

  • Fly Self Driving (A): the headline and rewired-graph numbers appear only in the project's results document (docs/results.md), and the stored baseline JSON does not match that document.
  • fly-brain by erojas and Kick the Fly (C): the neural parts alone could argue for B, but the behaviour shown is triggered, scripted motion.
  • Null models at the sensory-motor boundary (A): the brain is heavily compressed, to 512 cell-type groups.
  • flydoom (mutkuoz) and FlyWire neuron annotations have no licence file, so reuse of their code is not clearly allowed.

No grade of an existing entry changed. No tracked code changed, and no new claim contradicts an existing grade. The Doom control study adds evidence against the viral "fly plays Doom" framing and fits DOOMFLY's grade B; it cannot raise it.

The main finding: a constant autopilot plays Doom about as well

Doom survival: the fly brain against a brainless autopilotMean survival time in Doom, measured by the doom-fly-control study with DOOMFLY's own code: DOOMFLY, real wiring (full brain) 53.3 s (n = 20, 12.1 kills); Brainless constant autopilot 51.9 s (n = 20, 11.75 kills); Blind (no visual input) 10.5 s (n = 8); Shuffled wiring 6.7 s (n = 12, 3 seeds); No connections 5.5 s (n = 8). The autopilot always turns and walks at the real fly's average rates and holds the trigger.DOOMFLY, real wiring (full brain)DOOMFLY, real wiring (full brain): 53.3 s mean survival (n = 20, 12.1 kills)53.3 sn = 20 · 12.1 killsBrainless constant autopilotBrainless constant autopilot: 51.9 s mean survival (n = 20, 11.75 kills)51.9 sn = 20 · 11.75 killsBlind (no visual input)Blind (no visual input): 10.5 s mean survival (n = 8)10.5 sn = 8Shuffled wiringShuffled wiring: 6.7 s mean survival (n = 12, 3 seeds)6.7 sn = 12, 3 seedsNo connectionsNo connections: 5.5 s mean survival (n = 8)5.5 sn = 80204060Mean survival in seconds (third-party measurement, not reproduced by us) Mean survival time in Doom, measured by the doom-fly-control study with DOOMFLY's own code: DOOMFLY, real wiring (full brain) 53.3 s (n = 20, 12.1 kills); Brainless constant autopilot 51.9 s (n = 20, 11.75 kills); Blind (no visual input) 10.5 s (n = 8); Shuffled wiring 6.7 s (n = 12, 3 seeds); No connections 5.5 s (n = 8). The autopilot always turns and walks at the real fly's average rates and holds the trigger.Real wiring (full brain)53.3 s · n = 20Autopilot, no brain51.9 s · n = 20Blind10.5 s · n = 8Shuffled wiring6.7 s · n = 12, 3 seedsNo connections5.5 s · n = 80 s30 s60 s
Mean survival in DOOMFLY's Doom arena with DOOMFLY's own code. Measured by doom-fly-control (results/a2/summary.json, commit 6b22922); read by us, not re-run.

The real wiring matters for producing any output: shuffled or disconnected brains hardly move. But the output is close to a constant "turn slightly right, walk, shoot" command, and a constant command does as well in this arena. The autopilot's constants were taken from the real fly's average output, so the test shows that a constant policy reproduces the behaviour, not that the brain does nothing. The shuffled groups are small (4 runs per seed), and a control with matched activity is still missing.

On the site: the Doom test and updated verdict on Is it real?, two new verdict rows (Half-Life, Minecraft farming), five new rows in the controlled-studies table, and the new DOOMFLY project page.

  • All 40 published videos are still available, and so are the 4 new ones. Nothing was removed or restored.
  • 4 new videos: the Half-Life video (YouTube, Yusuf Temel, 12 Sep 2026, 6:26; grade D) and three video files linked from the READMEs of new entries: doom-fly-control (A), Fly.exe (B) and flydoom (A). They are links only; nothing is re-hosted.
  • Now linked to the catalogue: SWOT's Minecraft farming video points to FlyBridge (grade C). Its "does not show" text was rewritten from the code.
  • Still not graded: Ro0oney's Minecraft video. The mod it promotes publishes no code, and its site shows a crypto token address.
  • Search for new uploads: 10 YouTube searches sorted by upload date (156 results, 92 unique) found no relevant upload since 28 Sep 2026. 13 candidates were rejected: most show projects we could not identify or have not catalogued yet, and the rest are recaps or commentary.
  • The Half-Life video's "original upload" label rests on a name match between the YouTube channel and the GitHub account.

Source fix

The DOOMFLY limitations text had carried notes about how its grade was reached inside our team. It now begins simply "Grade B:"; the grade and its reasons are unchanged. We searched the public text fields of every record for such internal wording and found no other case.

Where the new entries came from

Sweep for new projects since 27 Sep 2026.
SourceQueriesHitsCheckedAdmitted
Deferred list from the first check–12128. fly.ai excluded (crypto); fly-survivors and ommatid gone; fly-connectome-template held back by the 15-entry limit
Projects behind ungraded gallery videos–221 (FlyBridge). flyproject.io excluded (crypto, no code)
GitHub search (created or pushed since 27 Sep)18132 repositories76. fly-trader excluded (crypto)
Hacker News search since 27 Sep4580 relevant0
Project directories (flybrain.info, Fly Brain Hub), compared with our list2211 not yet in our lists0 (backlog)0

We stopped at 15 new entries because our plan sets that limit per run, not because time ran out. fly-connectome-template is graded (n/a) and ready for the next run. Directory listings were used as leads only; no text was copied.

Method

Tracking

  • Last commits come from fresh shallow clones of each default branch; last releases from each repository's release feed; archive status from the repository page.
  • Our GitHub API allowance was low (60 calls), so for the 58 existing entries we reused stars, creation dates and last-push dates from our 27 Sep check. Each record says so (api_fields_reused_from), and the catalogue shows it next to the stars.
  • Every main link and every video was checked again over HTTP (videos through the platforms' embed-information services or a partial file download).

Grading new entries

  • Each grade rests on files we opened, cited with line numbers, and every entry has at least one source we read ourselves.
  • We re-opened the key files ourselves for every A and B grade and for the Half-Life D grade.
  • The 5 C grades rest on file citations from a short first reading (about 3 minutes per repository). We spot-read them but did not re-verify them line by line.
  • Videos were used for claims only, never for a grade.

Limits and uncertainty

  • All numbers are author-measured (a third party's for doom-fly-control). We did not reproduce any of them. Several controls are small: a single training seed, 4 runs per shuffle seed, 2 races.
  • No browser. Demos such as the Fly Self Driving site and Virtual Fly Brain are recorded as "page responds, not played by us". We did not watch any video frame by frame.
  • Short interval. Only about 22 hours passed since the first pass, so "no changes" says little about how fast the field moves.
  • Reused dates. Stars, creation dates and last-push dates for the 58 existing entries are from 27 Sep 2026 and labelled as such. Archive status and last commits are fresh.
  • Borderline and quickly read grades: see the new entries. The C grades were spot-read, not re-verified line by line.
  • Sweep gaps. X and TikTok could not be searched without logging in, so they were not searched. 211 repositories that appear only in project directories were not triaged yet.
  • Video details. For project-hosted files the date is the commit date; two lengths are not stated by their hosts.
  • Virtual Fly Brain's MCP server is VFB's own service. We did not test it, and Digital Fly Lab offers no API or MCP server of its own.

Next check

  • Publish fly-connectome-template, then triage natverse/fafbseg, neuroweave (a possible control study), likely copies of catalogued projects, and the directory backlog.
  • More videos: the second Half-Life video, the haltere and doomfly-rl videos, and the Minecraft and Half-Life videos by other creators once their projects are catalogued.
  • FlyBrain-HalfLife had a commit on 28 Sep 2026 (the version we graded); check whether later versions load a real connectome.
  • Next project pages: DesktopFly, then FlyGym or flyvis.

Key sources

Search Shaduf

Search published pools, pages, reports, and evidence.