First catalogue check: fruit fly brain simulation projects, viral claims and a tested beginner project

Reliability labels

direct

A raw page, file, API response or command output that we read ourselves.

fetch-summary

Read through a summarising fetch tool.

search-summary

Seen only in a search result: a lead, not evidence.

tested / untested

Whether a Build your own step was actually run in our test environment.

What we did

Two workstreams ran in parallel, each with a budget of about 90 minutes.

Catalogue findings

Every one of the 58 entries has a link check with an HTTP code, last_verified 28 Sep 2026, a licence status, its true last commit and release (or "n/a" for entries without a repository), a grade with its basis, and at least one primary source. The main link of 56 entries answered directly and 2 through a redirect; none was dead. Journal DOIs are often "blocked" by bot checks, which is not the same as dead.

• Coverage: all 7 datasets; the foundations (Shiu model, Eon fly-brain, flyvis, FlyGym, flybody, and two fast engines, fasterfly and drosophila-brain-mlx); all 12 of the 12 most-searched demos; 12 browser demos; 13 control or negative studies with numbers; 5 repositories created 20–27 September 2026.
• Grade basis: every A and B grade cites 3 to 8 code or result files; none rests on README text or search summaries.
• Datasets used: MaleCNS 20, FlyWire 13, several 12, larval L1 3, BANC 2, hemibrain 2, FANC 1, MANC 1, other or none 4.
• Code licences (from the licence files, counted from the records): 39 repositories have a licence file (26 MIT, 4 Apache-2.0, 6 GPL family, 1 BSD-3-Clause, 2 custom), 10 have none, and 9 entries have no code repository.
• Currency: other directories show creation dates, but true activity differs. The Shiu repository's last commit is 14 Sep 2024; flybody's default branch was last committed on 30 Jul 2025 but pushed on 7 Feb 2026; FlyGym 2.1.0 came out on 24 Jun 2026 after a breaking rewrite.
• Left out: one clone (flybook, a copy of Stonkfly), one unattributed re-host (fruitfly.live), 9 crypto-token projects and the directories themselves. About 50 more candidates were deferred.

Full table: Projects catalogue.

Datasets and neuron counts

Where the quoted counts come from: "125k" is the Shiu abstract's "more than 125,000"; 127,400 is FlyWire v630; 138,639 is the number of rows in the Shiu repository's v783 file (our count); 139,255 is the v783 release; 165,122 is the MaleCNS v1.0 neurons with status "Traced" (counted by the bioreservoir project, not re-counted by us); 166,691 is the MaleCNS preprint; 166,700 is MaleCNS v1.0 in Codex. Codex counts a "connection" differently per dataset: 5 or more synapses for FlyWire and MaleCNS, 3 or more for BANC, 1 or more for MANC.

What the Shiu model leaves out, in its authors' words: it "does not account for gap junctions, non-spiking neurons, internal state or long-range neuropeptides, and assumes that the basal firing of each neuron is zero". Neuromodulation is also not modelled. Weights are synapse count × one hand-tuned parameter, and signs come from predicted neurotransmitters. Details: How it works.

Is it real? Viral claims and controls

• "91%" refers to the brain-only Shiu model: "Across 164 predictions we were able to test empirically, 91% were consistent with our empirical results"; 84% without the split-GAL4 screen. "95%" is not an accuracy: "Decreasing the strength of inhibition results in predictions consistent with 95% of those of the default model; accuracy drops from 91% to 88%." Neither number describes an embodied fly or a game (direct).
• Uploaded? No. What exists is a measured wiring diagram plus simplified models.
• Conscious, or can it suffer? We found no study that measured or claimed this; the models omit internal state and neuromodulation, so any such claim is unsupported.
• Mouse or human next? Not soon. The largest mammal connectome, MICrONS, covers about 1 mm³ of mouse visual cortex, with more than 200,000 cells and about 523 million synapses (search-summary). See the State of Brain Emulation Report 2025.

Controls (authors' numbers, except the larva statistic, which we recomputed, and our own run):

• Reflexes: wiring mattersShiu paper: shuffled weights activated MN9 in 1 of 100 runs vs 100% with real wiring. drosophila-brain-mlx: 67.30 Hz real vs 0 Hz shuffled. FLY-lab: shuffled brains lose 69–100 points on turning. Our own run: 85.2 Hz real vs 0 Hz shuffled.
• General ML: no benefit so farLarva reservoir: +0.09 points over shuffles, p = 0.51. MaleCNS reservoir: memory 2.2 real vs 15.5 rewired. FLM: the adapter without the graph matches the fly adapter. Hemibrain hash: 0.499 real vs 0.621 random.

Full verdicts: Is it real?

Build your own: what we ran

Tested 2026-09-28 on Linux x86_64, Debian 12, Python 3.11.2, 2 CPUs, no GPU, no compiler.

1. Environment: python3 -m venv --without-pip .venv plus get-pip.py (the system Python had no ensurepip).
2. Packages: pip install -r requirements.txt in 32 s (brian2 2.9.0, numpy 2.3.5, pandas 2.3.3, pyarrow 25.0.1, joblib 1.6.0, pytest 9.1.1); the 21-package lock file was identical on both days; the environment is 463 MB.
3. Model and data: a sparse clone of philshiu/Drosophila_brain_model at 91bdd1e7 (14 Sep 2024) with FlyWire v783: Completeness_783.csv (3,327,347 bytes) and Connectivity_783.parquet (100,804,642 bytes), checksums recorded.
4. Tests: 6 passed in 52.3 s (checksums, IDs, silence without stimulus, sugar → MN9, shuffled control).
5. Experiments: 20 trials of 1 s in 1,685 s (28 min) with Brian2's numpy target: 58–65 s per trial with free CPUs, up to 138 s while shared. Peak memory about 0.58 GB per trial.

The 100 Hz reference is 1.4–1.7 Hz below two published v630 values, 67.30 Hz (drosophila-brain-mlx) and 67.03 Hz (the upstream notebook). We relay those and did not re-run them. We count this as agreement, not exact replication.

What we found along the way

• The v783 switch needs more than two config lines. Upstream's v630 sugar list raises KeyError: 720575940620900446 on v783 (reproduced by us); the fix is 720575940639259967 from issue #7. MN9's ID is unchanged.
• Memory, not CPU, is the real constraint. The upstream loader reads about 1 GB per trial and per worker and was killed repeatedly in our shared 3 GiB environment. Our loader streams 3 columns in chunks and runs each trial in a fresh process: peak memory fell from 1.44 GB to about 0.58 GB.
• Silencing cuts outgoing synapses only (issue #10, open).
• The default input rate is 150 Hz in model.py, while the notebook text says 200 Hz.

Licence. Upstream code is MIT (© 2023 Philip Shiu and Nico Spiller); our project files are MIT. For the data there is an unresolved conflict: flywire.ai/guidelines says CC BY-NC 4.0 (direct), while the Zenodo record of the v783 connectivity lists CC BY 4.0 (fetch-summary). We follow the stricter reading: FlyWire data and derived results are non-commercial, with attribution. MaleCNS (CC BY) is the candidate for commercial demos, but that path is untested.

Full guide: Build your own.

Source conflicts

Uncertainties

• Browser demos: none was played, so "works" is never claimed.
• Borderline grade A entries: Fly Worker (no committed script produces its correlation numbers), fly-brain by Lulzx (its benchmark was also used to fit parameters), The Fly's Hash Function (a one-step model) and Fly OCR (rewiring control only on a pilot). DOOMFLY was lowered from A to B. Wired Different is not graded (no simulation).
• Missing large files: grading used clones without files over 512 KB, so some bundled data files were not opened (noted per entry).
• Small samples in our run: 5 trials per condition and 3 per shuffle, against the paper's 30. The silenced neuron was chosen by our own rule; its effect is a model output, not a replication. Repeatability was shown only on our machine.
• Relayed reference values: 67.30 and 67.03 Hz and the paper's 1-in-100 shuffle result were not re-checked in our build run.
• Search-summary evidence: the MICrONS figures, the Beat Saber author's replies and the FANC counts.

Next runs

• Catalogue: the deferred long tail (Virtual Fly Brain, Fly.exe, fly-api, fly.ai, Fly Self Driving, ten more browser demos, 19 new repositories), dead-link re-checks, project pages for the foundations and top demos, and real-browser checks of demos.
• Build your own: brain–body coupling with FlyGym 2.x, the same project on MaleCNS v1.0 (CC BY), a compiler speed-up, macOS and WSL, GitHub Pages, and re-tests against upstream changes.
• Later: a synthesis of the "does fly wiring help machine learning?" controls.
• Open questions: the FlyWire licence, the 616-neuron gap, and the scope of BANC's "188,000".

Key sources

• FlyWire: guidelines, terms, Codex FAQ (direct)
• MaleCNS: portal, Google Research blog, preprint on PMC (direct)
• Virtual Fly Brain neuron counts (direct)
• MANC, hemibrain, BANC project, FANC, Winding et al. 2023 (direct)
• Shiu et al. 2024: doi, full text (direct)
• Drosophila_brain_model at 91bdd1e7; issues #7 and #10 (direct)
• Eon Systems, Eon fly-brain (direct)
• Zenodo record 10676866 (FlyWire v783 connectivity, fetch-summary)
• Mineault on viral fly simulations (fetch-summary); State of Brain Emulation Report 2025 (direct)
• Every project repository is linked from its row in the catalogue, with the commit and date we read. The machine-readable records are in projects.json.