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 body

What changed

• 30 per secondper looming neuron: the turn command from here up, on both eyes (brain only; 38% of our strongest input)
• 6 of 6real-map body runs turned away (0 of 3 scrambled body runs; inside-only scramble 0 of 6)
• 1.1 GBbrain and body in one Python 3.12 environment, with numpy 2.3.5
• 43control studies (2 new), 34 with a wiring null: helps 16, no difference 8, worse 2, mixed 7, not yet scored 1

How strong must the shadow be?

The objection to an 80-per-second input on every looming neuron is that it is so strong that any wiring connecting the eye to the turning neurons would turn. We kept the model, the seed (2, as in the 6 Oct real trials) and the readouts, and changed only the input: 40, 20, 60, 10 and 30 per second, alternating eyes, so that a time cut would leave balanced sides. The 80-per-second points are the 6 Oct trials. At the threshold and one step below it, a second seed (0) checked the result.

Open the share card (SVG): the sentence, the threshold, the caveat and the date on one image.

Descriptive, pre-registered as optional, n 1 per eye We also walked the threshold drives in the body: At 30 spikes per second per looming neuron, the fly turned away by 11° (left eye) and 17° (right eye). Both above the 10° line of our body test (11.1° and 16.8° beyond the no-brain floor, which repeated our 5 Oct floor run exactly in this environment). One run per eye; not in the ledger.

The design, written down first

Deviations, in plain words

• The combined environment could not run a brain trial on numpy 2.5.3 (the import error above), so the pre-registered fallback E1 was taken: numpy 2.3.5 in the same environment. The pre-registered check of that environment (gate E) then passed on every criterion, and the trial was identical to the 6 Oct trial (Python 3.11.2 then, 3.12.11 now).
• The environment was grown while the body runs were still going, to save time; no body package changed, and every body run recorded numpy 2.5.3.
• The installer added setuptools 84.0.0, a Brian2 dependency not in the planned list.
• Wording made explicit before the first brain trial, no rule changed: in this series "the turn rule passes" means fly67's turn rule only; the escape rule is the giant fibre above 50 Hz.
• The scratch folder moved to the path the run's scratch rule allows. No effect on results.

Limits of this test

• The input is ours. Every LC4 and LPLC2 neuron of one eye at one rate, with no timing or geometry of an approaching object. We found no measured real-fly LC4/LPLC2 firing rate during looming to compare with (one abstract search of 8 Oct 2026; full papers were not read), so "38%" describes this model only.
• Brain only. Descending-neuron rates over 1 s. Two threshold drives walked in the body as a descriptive extra (one run per eye, below the chart); nothing from the strength series enters the ledger's effect.
• The relay measure saturates. The pre-registered LI ratio stays near 1 whenever the near side is silent, so it cannot show a weaker far-side command; the far-side rates are disclosed next to the readings. A rate-based measure would have to be pre-registered for a later test.
• "Silenced" removes output synapses only. The silenced neurons still receive input and spike.
• One model on one fly's map; not a real fly.

Silencing the relays

The census of 6 Oct counted 21 two-step relays (left eye) and 23 (right eye) between the looming neurons and the far-side turning neuron DNa02, with PVLP141 on the eye's side as the strongest (one cell per side: 361 / 451 synapses in, 98 / 81 out). The lists were rebuilt and written down before the first relay trial and match 6 Oct. At 80 per second, seed 2, we silenced PVLP141 alone, then all two-step relays of that side.

The looming body test, completed to n 6

The seven drives of 6 Oct that differed from the no-brain floor walked in FlyGym 2.1.0, with the 5 Oct body code and lock file unchanged and every drive file checked by its sha256. First the floor run F: "repeated exactly", all 140 × 5 stored trajectory values equal to the 5 Oct run (and again in the combined environment with numpy 2.3.5). The chart with every path is on Is it real?

• Pre-registered Real arm, n 6: 6 of 6 turned away (5 required), mean 49.9° ± 13.1 sd; scrambled retention 0.0; no noise flag. Reading: "the turn away needs the wiring".
• Pre-registered Giant fibre silenced: With the escape neuron silenced, the fly still turned away. 2 of 2.
• Pre-registered Inside-only scramble: With only the inside of the map scrambled, the fly turned away in 0 of 6 runs. One body run and five drives identical to the floor (same sha256), which take its walk by determinism.

Limits: one model, open loop, our hand-made mapping, a constant external walking drive of 0.8, 1 s; the body walks with FlyGym's own controller; "away" rests on our mapping's DNa02 sign and fly67's claim; not a real fly. The scrambled body runs are few, and the floor-identical drives rest on determinism, now confirmed across sessions and across numpy 2.5.3 and 2.3.5 for the floor run.

One environment for brain and body

Until now the guide needed two environments: Python 3.11 for the brain and Python 3.12 for FlyGym. We grew the body environment in place with the brain packages, using the body's lock file as constraints so that no body package changed, and checked it with the pre-registered gate E.

• Gate E on numpy 2.5.3: failed at import. import brian2 (2.9.0) raised AttributeError: type object 'numpy.ndarray' has no attribute 'ptp' (Brian2 wraps np.ndarray.ptp in brian2/units/fundamentalunits.py). No trial could run.
• Fallback E1: numpy 2.3.5 in the same environment (the only change to the freeze). Gate E then passed on every criterion: fly67's rule, giant fibre 116.5 Hz, LI 0.962, stimulated rate 80.13 Hz, 28,153 whole-brain spikes, 1,087 MiB on disk. As a fact, not a criterion: the trial was identical to the 6 Oct trial in every readout.
• The tested line on Build your own: "Brain (Brian2 2.9.0) and body (FlyGym 2.1.0) installed together in one Python 3.12 environment: 1.1 GB on disk, tested 8 Oct 2026." with the numpy 2.3.5 condition and the error as a common-errors row (the commands).
• Whether Brian2 2.10.x imports with numpy 2.5 was not tested. The environment was deleted at the end of the run.

Controls ledger: 43 studies

• Connectocopter (new, grade A): looming-ball escapes 15 of 15 with the real wiring against 0 of 15 with one degree- and sign-preserving rewiring (raw files recounted by us). "Helps"; but a hand-written controller also escapes 15 of 15, so the no-brain baseline is not beaten. Unfair (one rewiring seed, thresholds tuned on the real wiring), weak.
• Synaptera (new, grade B): the landing neuron responds to expanding flow no more than in the author's shuffled connectome. Its null type is "other" by the fixed rules, so the row counts as baselines only (not tested for wiring). From the author's report only.
• Our looming row (byo-flygym-loom): real arm n 6 (49.9°), a new inside-only arm (n 6, −0.24° beyond the floor), still "helps", fair, strong, baseline beaten. Its one-line: "Our test: a looming shadow turned the fly away in 6 of 6 real-brain runs, 0 of 3 scrambled body runs (+3 floor-identical); our mapping." The giant-fibre-off result and the brain-only results are in its note only.
• Counts: 43 studies, 34 with a wiring null: helps 16, no difference 8, worse 2, mixed 7, not yet scored 1. No scoring rule changed.
• A convention, no schema change: extraction: "files" with quality.results_in_repo_files: false means the numbers come only from a README. Documented on Does fly wiring help? and For AI agents.

Lead scan and verdicts

Promoted: one claim. "Can A Conscious Fly Brain Learn how to Hack?" by dzuma (video page, uploaded 7 Oct 2026, approximately 37,310 views on 8 Oct 2026). The description links papers, the FlyWire Codex, Brian 2 and two deliberately vulnerable practice targets, but no code or data; we did not watch the video. Grade U; verdict row.

• Not promoted: Idor, "Internet está torturando una mosca" (video page, in Spanish, approximately 422,364 views on 8 Oct 2026; an explainer of the meme wave with no claim of its own); Garett, "Someone Made a Digital Circus for Fruit Flies" (video page, approximately 466,745 views on a search page, 8 Oct 2026; uploaded 19 Sep, older than our 14-day rule); two commentary videos already judged on 6 Oct.
• Row updates (approximate, 8 Oct 2026): Gorilla Tag about 928,414 views, still no code; the uncatalogued Minecraft video about 1,584,514 views (read from search pages).
• Other sources: Hacker News had no fly-brain story; Google News showed a real-fly memory study (not a digital fly); Reddit refused our request (unavailable); dataset pages showed no new release; GitHub search found 88 repositories, none with 20 or more stars.
• GitHub API: 5 of the 10 requests the owner allows per run, all answered; no rate-limit stop. Not full social coverage: TikTok, Instagram and X were not searched.

Five new entries

"Is it AI?" now: 48 of 87 graded projects have a trained, learning or search-tuned part. Deferred for grading: four cloned repositories (among them a pre-registered study with rewired controls). No name-copy download lure was found or linked.

Use in class

A 20-minute plan in six steps on the tested beginner path (sugar neurons → MN9): one live sugar trial, the scrambled-map control and the looming result shown from our tests, and three discussion questions from our plain answers, with browser demos and play.drosophila.io as no-install alternatives. Built only from results we already had; every number has a source. The plan on Build your own

Tracking and checks

• Code: 103 repositories checked without the GitHub API: 99 unchanged, 3 changed (Brain Runners: write-up commits, its result files untouched; navis: two fixes; one excluded repository), 1 gone again (flydoom; record kept). Nothing material; no re-grade due.
• Links: 243 links: 211 answered, 13 redirected, 15 refused automated requests, 4 client errors; no new link problem.
• Videos: all 54 still available; the dzuma claim video added (55).
• Verdict freshness: all 21 rows checked on 8 Oct 2026: 13 with code unchanged since graded, 8 with no code to check.
• Carried: MANC's size is still unknown; Neural Canvas's code downloads about 77 MB after Start, while our 6 Oct check saw 5.87 MB; larval L1 sizes not read (the source refused).
• PyPI (8 Oct 2026): Brian2 2.10.1 (our pin 2.9.0), numpy 2.5.3 (beginner pin 2.3.5), pandas 3.0.6 (pin 2.3.3); FlyGym 2.1.0 is the latest. FlyGym and its scipy need Python 3.12.

Method and limits

• One simulation process at a time; every brain trial started below the pre-set memory guard. 19 brain trials (1 environment check, 10 strength, 4 relay, 4 seed checks), 7 body runs plus the floor, and 2 descriptive threshold body runs. Every trial has its own result file and log line; the catalogue and ledger builds ran twice with identical outputs.
• The run finished within its 60-minute plan (about 48 minutes); no planned cut fired.
• Brain-level results never enter the ledger's effect. The body test uses our hand-made mapping and a constant external walking drive.
• Synaptera's ledger class follows the fixed rules; we did not inspect its shuffle method.
• YouTube coverage is three search pages per term at one moment; some watch-page counts of older videos cannot be read without a cookie.

Next

• Which longer paths carry the turn: a three-step census, silencing by depth or cell type, and a pre-registered far-side rate measure next to the ratio.
• A measured real-fly LC4 rate from the electrophysiology literature, to place the 30-per-second threshold.
• A coupled brain-and-body test, in which the fly's movement changes what the eye sees, is planned; it is not yet tested.
• Grade the four deferred repositories; map the two new explainer videos.