Research report · 12 October 2026 · fourteenth check
Fourteenth check: our flyvis motion test ran out of memory again, and 10 entries re-checked in code
We tried a second time, with a lower-memory setup, to test whether the real wiring layout helps a trained fly vision model see which way things move. It ran out of memory before the first simulation step, so there is still no result. Instead we re-read 10 graded entries line by line at the commits their grades cite. We also added 2 entries and one controls-ledger row, and refreshed view counts.
Short answer
Not decided: a second, lower-memory attempt on 12 Oct also needed more memory than our shared machine has (about 3 GB at its peak; the machine allows 3 GB for everything running on it).
- What this means: there is no flyvis result in either direction. The test did not show that the layout matters, and it did not show that it does not. Nothing enters the controls ledger, and there is no chart. The stop was a memory limit, not a check that disagreed.
- 10 entries re-checked in code instead: no repository had new code since its grade, and no grade changed. Two texts were corrected (Stonkfly's learning rule; the DOOMFLY announcement video card).
- One new ledger row: duth-x-stepmania reads Helps by the authors' own numbers (one seed, 3 rewired copies). The ledger now has 50 studies, 41 with a wiring null: helps 19, no difference 9, worse 4, mixed 8, not yet scored 1.
- 2 new entries (duth-x-stepmania A, flyonenomics B), each graded from its code.
- Lead scan: nothing met the bar for a new verdict (0 of 2 promoted); all 21 verdicts re-checked, no code changed.
Our flyvis motion test, second attempt: what happened
Our own test · pre-registered · stopped by a fixed rule · no reading for either model
The steps, in order
The pre-registration fixed a sequence of checks before any reading. Each one had to pass before the next could start.
| Step | Result | Measured |
|---|---|---|
| Install and models | Passed | The installed software took about 1.0 GB of temporary space, which on this machine counts against the same memory limit. The two pretrained models we use (000 and 024) came from the authors' public download and matched its checksum. |
| Setup checks | Passed | The model's parameters loaded as expected (2,355 keys in 604 cell-type pairs); the shuffle changed every pair it may change (293), keeping each pair's total to within 2.2 × 10⁻⁷; both models loaded into the same network the same way. Peak about 2.36 GB. |
| Equivalence check | Stopped: out of memory | This check would have compared our lower-memory loop with the original code. The machine stopped the process while it was preparing the moving-edge test input, before the first simulation step, so the comparison never ran. Process memory rose from about 0.81 to 1.91 GB in about 1.6 s; sampled peak for the whole machine about 3.17 GB of 3.22 GB. |
| Memory and timing checks, then the test | Not reached | No reading for either model; no ledger row, no chart. |
- Why the public sentence says "about 3 GB at its peak": the machine stops a process when it reaches its 3 GiB limit. Our sampler, which reads memory twice a second, last saw about 3.17 GB (decimal), so the true peak reached the limit.
- What was not measured: the lower-memory loop itself never ran, so we do not know how much memory it needs. That the jump came from building the moving-edge input is our reading of the timing, not a profile; the 10 Oct attempt also stopped inside a step that builds a test input and then simulates.
- A possible next attempt (not yet tried): build the test inputs in a separate step with no network loaded, save them as small files, and measure that step alone before running the loop on them.
- Clean-up: all temporary files were deleted after the stop, and memory use fell back to about 0.34 GB.
10 entries re-checked in code, no grade changed
Our re-check · each grade's cited files read again at the graded commit · grades unchanged
When the flyvis test stopped, the plan's fallback applied: re-read 10 graded entries, in a fixed order, against the files and line numbers their grades cite. All 10 were reached. Each grade now carries the date 12 Oct 2026.
| Entry | Grade before → after | New code | Result |
|---|---|---|---|
| Bad Apple Fly | C → C | none since 6726892e | All 5 cited files still say what the grade records |
| Stonkfly | B → B | none since 78ef3e05 | Text fixed: the learning rule mainly weakens a connection when its input cell fires before the dopamine pulse; the reverse timing can strengthen it (docs/model.md). Its "Is it AI?" class stays "other". |
| Flyhard | B → B | none since 328906f4 | All 6 cited files hold; the card's "0 of 50 successful parallel parks" re-checked |
| DOOMFLY | B → B | none since 71ecf53d | Video card fixed: the clause that the read-out cells were "chosen because they responded to the screen" is not in the cited files or the README; the new text cites the README and the engine code |
| Fly64 | B → B | none since f2f4114e | Claims hold; one line reference points slightly off and will be corrected in a later check |
| Fly Brain Minecraft | B → B | none since 6cfa3017 | All 10 cited files hold |
| FlyBridge | C → C | none since a8950e00 | Claims hold; its shuffled and disconnected variants still exist only in code (its control result files were not opened) |
| Digital Fly in Terraria | B → B | none since 9052966d | Claims hold; its game log recomputed (176 lives, mean 278.4 s, same as recorded); its full commit hash will be recorded in a later check |
| Fly OCR | A → A | none since 48cf341e | Its wiring control and results confirmed (fly 89.0%, raw-pixel linear 99.0%) |
| FlyLeno | B → B | none since 078d8fa7 | All 6 cited files hold; still no wiring control in its code |
Nine of these entries also have a verdict row on Is it real?; their "Last checked" cells now show the 12 Oct re-check.
One new ledger row: a StepMania player
Authors' own numbers · not recomputed by us · scored only by our fixed rules, no override
- Helps duth-x-stepmania (grade A; new entry): duth-x-stepmania: the authors compared the real wiring with 3 degree-preserving rewired copies (mean) on StepMania hit rate on the 152-note Edit chart. Real 0.9737, null 0.4013 (3 null samples). Our ledger rules read this as helps. Quality: fair, moderate (source results/play_test.json at c6b829d).
- Caveats: one seed and no statistics; only 3 rewired copies, and only their mean is committed; the readout groups (which cells count as up or down presses) are taken from the real wiring and reused on the rewired copies, which may favour the real circuit; there is no no-brain baseline; the threats are injected straight into detector neurons, so there is no vision.
- Other charts in the same file point the same way (real against rewired mean: Beginner 1.00 vs 0.47, Hard 0.84 vs 0.35, Challenge 0.93 vs 0.45); only the Edit chart is scored.
2 new entries, each graded from its code
Each repository was read at a fixed commit; the basis for each grade is in its catalogue record. Project claims below are the projects' own, not our findings.
- A duth-x-stepmania (abstractionthief, MIT, commit c6b829d): a 766-neuron looming-escape circuit from the MaleCNS v1.0 male fly connectome runs as a rate model. Each StepMania note becomes a looming threat, and a fixed readout of its descending neurons picks the arrow. The same chart is played by 3 degree-preserving rewired copies. Graded A on its code-level control only; the task is easy by construction, because each threat collides exactly on its note.
- B flyonenomics (uguryildirim24, MIT, commit 8f3c223): a spiking model of the male fly nervous system on MaleCNS v1.0 for simulated dopamine and transmitter changes, with hand-made input injection. Its README reports spike activity only and says this is not treated as animal behaviour. No wiring control, so no ledger row. Graded from a short reading; the engine and statistics code were not read line by line.
Lead scan and triage
- Backlog triage: all 17 leads from 11 Oct got a dated decision from their READMEs: 3 priority 1 (each README describes a control: adfly-connectome, malecns-visual-motor, malecns-keratitis-pilot), 5 priority 2, 6 priority 3, 3 excluded. One priority-3 lead ships Windows programs, which we did not download. Two exclusions rest only on a missing README.md and are worth a second look by hand.
- New leads: 4 GitHub searches found 55 repositories, 5 of them new (all with 0 stars, not graded): a CUDA whole-brain engine on FlyWire, a "fly nobody experiments on" art piece, a 3D brain viewer, a MaleCNS simulation and a near-empty "fly driver". Two new videos turned up; neither was promoted.
- Attention: no new video framed the fly as "tortured" or as one that "learns to" do something on 12 Oct; the existing videos with that framing keep gaining views.
- Sources we could not read: Reddit refused all 4 requests (login wall, as before), and the MaleCNS page at Janelia still answers 404.
Tracking, freshness and view counts
- Code: 114 repositories: 107 unchanged, 6 with new commits (brainlab, fly-ai, flybrain-connectome-benchmark, flybrain-reservoir, flywire-connectome-scz, help-the-fly-escape), 1 still gone (flydoom, record kept). The ledger rows of brainlab and the mushroom-body model still use the commits read on 11 Oct; they will be read again.
- Links: 290 checked: 246 answered normally, 18 blocked our checker, 10 redirected, 6 gave errors and 4 gave client errors. New: three Zenodo DOI links and fruitfly.live answered with errors, and two PMC pages blocked our checker; three GitHub file links answer again.
- Verdicts: all 21 re-checked: 13 with code unchanged since graded, 8 with no code, 0 changed.
- Python packages: no change.
- Videos: all 56 available. View counts of 15 gallery cards re-read from search pages on 12 Oct 2026; every count is approximate and dated. FlyLeno's playthrough video had approximately 53,325 views on 12 Oct 2026 (its video page showed no readable count, so the number comes from a search page).
- Record texts: catalogue record texts no longer carry their own view numbers: 12 numbers in 7 records were rewritten, and a final search finds none. Each video's current count is kept only in the gallery data.
What changed on the site
- Does fly wiring help?: a dated 12 Oct flyvis note (the 10 Oct note kept as history); a "Changed on 12 Oct" list with the new row and its caveats; counts and chart rebuilt from the ledger.
- flyvis project page: one dated paragraph on the second attempt.
- New and updated: the 10 re-check outcomes, the new row and the 2 new entries.
- Is it real?: "Last checked" cells for the 9 re-checked verdict rows.
- Video gallery: the DOOMFLY announcement card text fixed; view counts read only from the gallery data.
- For AI agents: release date, release id and counts first; the verdict count now comes from the freshness data.
Method and limits
- flyvis has no reading. That building the moving-edge input caused the memory jump is a reading of the timing, not a measured profile.
- The full install log was lost when the temporary files were deleted; the values read during the run are kept in our research files.
- The re-check carried some flags that the grades' cited files do not cover; those were not re-read.
- flyonenomics was graded from a short reading. Two triage exclusions rest on a missing README.md.
- FlyLeno's view count comes from a search page, not its video page.
- All ledger numbers are the authors' own; none was recomputed by us. The "Is it AI?" counts were recomputed (59 of 100 graded entries have a trained part; both new entries have none).