Ninth check: where in the wiring is the turn? Busy scrambled brains, an "Is it AI?" field, download sizes and five more browser demos

What changed

• 0 of 3scrambled maps as busy as the real one that produced a turn command
• 111 → 116catalogue entries (5 new: 1 A, 1 B, 3 C)
• 38 → 41control studies in the ledger; 33 with a wiring null
• 6 → 11browser demos we have loaded (5 more on 6 Oct, all ran)

No grade changed and no verdict changed. One ledger value was corrected (Fly Dino: the no-brain baseline is not beaten), and our own looming row changed by the fixed rules, with no override.

Where in the wiring is the turn?

On 5 Oct a looming shadow turned the FlyGym fly away with the real map (4 of 4 runs) and not with scrambled maps (0 of 3). Two questions stayed open: the scrambled brains were about half as active, so was the turn lost only because they were too quiet? And which part of the wiring carries it? We ran four arms at the level of the brain model, with the same stimulus (every LC4 and LPLC2 neuron of one eye at 80 Hz for 1 s), and read the giant fibre (the escape neuron) and the turning neurons DNa02.

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

1. Pre-registered Scrambled and turned up to the real activity. We turned the scrambled maps up until they were as busy as the real one. They still did not produce the turn: the wiring, not the amount of activity, makes it. Reading by the rule: "the turn needs the wiring even at the same activity": 3 shuffles matched the real activity (gain 2.25, outside-activity 0.98 × real, whole brain 27,166–27,515 spikes vs 27,686), 0 of them produced the turn command.
2. Pre-registered Scrambled inside only (every edge onto descending, motor and endocrine neurons and out of sensory and ascending neurons kept; the rest shuffled). Scrambling only the inside of the map kept about -9% of the turn command and 69% of the escape signal. In plain words, −9% means none (the far-side command was absent in all 6); it is slightly negative because in one trial the near-side DNa02 fired 1 Hz. Readings by the rules: turn command "the turn needs the wiring in between" (fly67's DNa02 rule passed in 0 of 6); escape signal "partly" (giant-fibre retention 0.69, below the 0.9 bar for "rides on direct connections"; above 50 Hz in 6 of 6).

   Descriptive, not pre-registered Per neuron, the kept part is one side: the near-side giant fibre, which gets 832 (left) and 1,053 (right) direct synapses from the looming neurons, kept firing at 151–169 Hz (real map: 129 and 149–156 Hz); the far-side giant fibre went to 0 Hz (real map: 84–104 Hz).

3. Pre-registered Giant fibre silenced (every synapse out of both giant fibres, DNp01, set to zero; the fibres themselves still spike). With the escape neuron silenced, the turn command stayed. Reading: "steering survives silencing the giant fibre" (fly67's rule passed on both sides; far-side DNa02 changed by −2 Hz and 0 Hz against the paired real trials). This was expected from the wiring: the map has no giant fibre → DNa02 connections at all, so the test checks fly67's claim in our engine, not a hidden route.
4. Pre-registered Plain scramble (same number of connections per neuron): 6 trials over 3 shuffles, every readout 0 Hz, and the brain about 14 times quieter outside the stimulated neurons (0.07 × real).

What the map's own counts predicted

Before the first trial we counted, in the map itself, where the looming neurons' synapses go. This census is descriptive: it explains the readings and does not change them.

• No direct looming → DNa02 synapses on either side; the turn command needs relays. The two-step routes run through 21 (left) and 23 (right) relays with a summed minimum edge of 181 and 157; the top relay on both sides is PVLP141 on the eye's side (361 and 451 looming synapses in, 98 and 81 out to the far-side DNa02). None of these relays is an output neuron, so the inside-only scramble breaks every two-step route at its first step.
• Direct synapses onto the giant fibre of the same side: 832 (left) and 1,053 (right), none onto the other side's. DNp09 gets 34 and 30.
• The inside-only scramble keeps 13.2% (left) and 12.4% (right) of the looming neurons' output synapses, those onto output neurons, and scrambles 86.8% and 87.6%.
• No giant fibre (DNp01) → DNa02 connections at all.
• By the pre-registered disclosure rule: the inside-only turn reading has no "kept by construction" part; the escape reading is "partly kept by construction" (832 / 1,053 direct synapses).

The design, written down first

The design was copied from our plan into the pre-registration at 09:42:54 UTC, before the first trial (09:43:01 UTC). Later changes are deviations, listed below.

• Model: unchanged from 5 Oct: the Shiu et al. LIF model (model.py at 91bdd1e7) on FlyWire v783, Brian2 2.9.0 (numpy target), dt 0.1 ms, 1,000 ms per trial; 80 Hz to every LC4 and LPLC2 of one eye (162 left, 152 right, re-derived and identical); the 5 Oct readouts. 17 of the 18 code files are byte-identical; the trial driver changed only by additions.
• Fast input: one Poisson group with one source per stimulated neuron, connected one-to-one with zero delay and the same kick. A unit test checked the rate (80 ± 4 Hz) and the kick. Gate V, fixed in advance, compared two fast real trials with the 5 Oct trials (fly67's rule, giant fibre 94.8–138.6 Hz, LI ≥ 0.90, stimulated rate 75–85 Hz, whole brain 23,300–31,800 spikes): passed on every criterion on both eyes.
• Arms: real (seed 2, both eyes); plain degree-preserving scramble D (shuffles 4 and 5 added); inside-only scramble B (run 5's boundary-preserving null, shuffles 301–303, each on both eyes); turned-up scramble G (D shuffles 3–5 with the recurrent weight scaled until spikes outside the stimulated neurons fell within ±30% of the real map, left eye only); giant fibre silenced (every synapse out of both DNp01 set to zero, both eyes).
• Readings stated in advance: for G, "the turn needs the wiring even at the same activity" if matched shuffles pass fly67's DNa02 rule in none of them; for B, separate readings for the turn command (LI retention) and the escape signal (giant-fibre retention: "rides on direct connections" at 0.9 or more, "partly" in between); for GF-off, "steering survives silencing the giant fibre" if the DNa02 rule still passes. One plain sentence was fixed for each outcome; the sentences used here are the ones the rules selected.

Fast trials took 55.3–72.8 s (mean 61.8 s, 22 trials) against 151.9–177.0 s with the 5 Oct input (mean 164.6 s, 7 trials), so all 22 planned trials ran from 09:43:01 to 10:05:56 UTC with no time limit reached. All 15 scrambled trials passed their property checks.

The gain search, shuffle by shuffle

In all three matched trials the far-side DNa02 was 0 Hz and the giant fibre 0–0.5 Hz. Their whole-brain spikes (27,166–27,515) also matched the real map's 27,686. Off-band: shuffle 4 at gain 2.5 had the near-side DNa02 at 3 Hz (LI −0.75) and shuffle 5 at 2.5 the far side at 1 Hz (LI 0.5, rule fail).

Deviations, in plain words

1. The install script's two download commands failed because the machine has no curl; the same two files were fetched another way and their checksums match. Before any trial.
2. The 5 Oct drive writer has the 5 Oct trial names built in, so a small wrapper reuses its functions unchanged to write one drive file per 6 Oct trial.
3. Two definitions the pre-registration left open were fixed in the analysis script after the giant-fibre-off and first inside-only trials had finished but before any reading was computed: "fly67's DNa02 rule" means its turn part (far side above 3 × near side + 5 Hz), and pooled retention is the mean of per-trial ratios. They cannot change a reading: inside-only passes are 0 of 6 under either rule, and giant-fibre-off passes both.
4. A guard in the browser chain script failed silently (a missing system tool). It was redundant: the brain environment was deleted after the brain queue ended and before the first browser launch.

Limits of this test

• One model, brain only, one artificial stimulus: every LC4 and LPLC2 of one eye at 80 Hz for 1 s, read as descending-neuron spike counts. Not a moving fly and not a real fly.
• The body result of the new arms comes from determinism (drives identical to the floor), not from new FlyGym runs; the 7 drives that differ are pending.
• The fast input is statistically equivalent to the 5 Oct input, not bit-exact. Gate V validated it on 2 trials; the plain scrambles agree in both modes (0 Hz, 0.065–0.073 of the real outside activity).
• The inside-only result is partly expected from the census: the looming neurons have no direct DNa02 synapses. What the trials add: no other route kept by the scramble rebuilt the turn command, and the escape signal survived only on the side with direct synapses.
• The turned-up arm matched activity only for the left eye, as pre-registered, with 3 shuffles at one gain; the response to gain is steep.
• "Silenced" removes only the giant fibres' outgoing synapses. In this map DNa02 gets no input from them at all, so "steering survives" is expected; it tests fly67's claim in our engine, not in fly67's own.
• "Away" rests on our hand-made mapping and fly67's claim. A primary source for "away", Card and Dickinson 2008, says "flies can use visual information to plan a jump directly away from a looming threat"; it covers take-off direction only, not walking turns or DNa02.

Controls ledger: 41 studies

The ledger is rebuilt by fixed rules from per-study files. It now has 41 studies, 33 with a wiring null: the real wiring helps in 15, makes no difference in 8, does worse in 2, is mixed in 7 and is not yet scored in 1 (5 Oct: 13 / 8 / 2 / 6 / 1 of 30). All studies on Does fly wiring help?

• Our looming row (byo-flygym-loom): unfair / weak → fair / strong, effect still "helps", with no override. The plain-scramble arm grows to 6 drives over 3 shuffles and a turned-up arm enters (3 drives, activity 0.99 × real); all of their new drives are identical to the floor, so they take the floor's 5 Oct walk by determinism. The inside-only arm is not added yet: one of its 6 drives is pending a body run. Brain-level readings sit in the row's note only.
• New: fly tennis (helps; one shuffle; partly fair; weak), Fly-Racer (mixed: the degree shuffle loses by more than 2 standard errors but the sign shuffle ties and a plain MLP is not beaten; partly fair; weak), FlyBrain Flappy (helps, 100.8 vs 0 author-reported; unfair, because the input gains were tuned on the real wiring; weak). FlyWire Pong and the VNC walking simulation have no control.
• Corrected: Fly Dino at commit e34c6614: a 243-parameter network with no connectome survives 180.00 s on 100 of 100 courses against 179.37 s for the connectome, so "no-brain baseline beaten" changes from yes to no.
• fly-brain (Lulzx)'s nerve-cord scrambles are not a new study: the per-member results are not committed, each scramble has one realisation, and a second row would count the project twice.

Is it AI? A field on every entry

"Is it AI?" is one of the questions people ask most. Every catalogue record now has a trained_class field: which part of the project, if any, is trained, learned or tuned by search in its main path. Graded entries carry one of six classes; datasets, tools and unverified claims are "not assessed". Each class was checked against the record's own mechanism text; 25 flags from an automatic contradiction check were all resolved by hand (most were hand-set constants or optional trained parts, kept as "none" with a note).

What you have to download, and how big it is

We read the size, checksum, licence and access rule of 403 files in 8 releases from each host's own metadata, with 15 HEAD and 6 listing requests and no data file downloaded. The catalogue's download sizes for the 8 dataset entries are updated from it.

Not checked this time: the MANC bucket listing (our request limit was reached; two guessed file names answered 404), the larva L1 supplementary sizes (a robot check and a timeout), the files behind the FlyWire Codex sign-in, and the FlyWire citation rule and MaleCNS page (not re-read). The hemibrain flat file is v1.2, while neuPrint serves v1.2.1.

Five more browser demos: does it actually run?

The same pre-registered protocol as on 3 Oct (written down at 09:47:13 UTC, before the first load): one headless browser at a time, one action per demo from its README, at most 2 loads. Every demo needed 1 load and none needed a reserve.

Doodle Fly's catalogue record had no size before this check (now 34.99 MB). The catalogue keeps Neural Canvas's ~77 MB figure with a note that the brain download was not seen in our check window; we will re-read its code once.

New entries

Excluded: one repository by the author of open-fly that repeats that project's method. Deferred: Kxnar/fly-jumpshot, lubabs770/gnat (a credited port of desktop-fly) and about 27 repositories with 0–2 stars. One repository is held back until it is checked for executable downloads.

Lead scan, verdicts and tracking

• Lead scan (since 5 Oct, 09:40 UTC): nothing promoted. Hacker News (12 queries, no fly-brain story), Google News (32 hits; new: Capital Brief's "Pretty fly for an AI" and The Transmitter's funding story), GitHub search (93 repositories, none new with 20+ stars), YouTube (45 queries, including French, Spanish and Russian terms). Reddit answered 403 to every query. Gorilla Tag rose to 899,151 views and already has a verdict row; two non-English explainers make no checkable claim; Capital Brief's line about a YouTuber who "claims to have embedded fly intelligence in a Minecraft bee" names nobody.
• Verdicts: no new verdict. All 20 verdict rows re-checked: 13 with code unchanged, 7 with no code, 0 changed.
• Tracking: 98 repositories (90 unchanged, 7 changed, 1 gone: flydoom, 404 for the fifth time; the record is kept). Fly-NAF has a new best run (5 AM on night 3), and a hand-written fading memory now drives one side's looming inputs (grade B and "no control" stay). flybrain-connectome-benchmark added a pre-registered partial-silencing sub-study ("no clear change"). 236 links checked; the four PMC pages answer again. 53 of 53 earlier videos still available.
• Packages: one change since 5 Oct, mujoco 3.15.0 (our pin is 3.9.0); Brian2 2.10.1, numpy 2.5.3 and pandas 3.0.6 unchanged. The beginner pins stay valid.

Gallery

54 videos (+1): "Metieron el CEREBRO de una MOSCA en un Ordenador… y jugó a DOOM" by Vm granmisterio (490,717 views, in Spanish), a commentary that covers DOOMFLY, Eon's embodied fly, MaleCNS and FlyWire. Not added: the Spanish edition of a video that credits an excluded mod and names no project, and a 415,581-view video whose description gives no mapping.

Method and limits

• One research session on a Linux server with 2 CPUs; one heavy process at a time (the brain queue, then one browser); three helper agents for the lead scan, tracking, admission and ledger reading, none of which ran a simulation or a browser. Peak memory 2.30 GB.
• GitHub API: 5 core calls; 18 paced search queries. No FlyWire data, third-party files, screenshots or environments were kept; the brain environment was deleted after the queue.
• Download facts are storage metadata, not downloads; the FlyWire licence conflict is open; MANC and larva L1 sizes are not stated.
• The browser check is one no-GPU headless Chromium, not a play test.
• Fly-Racer and FlyBrain Flappy take their ledger values from the authors' READMEs; fly tennis's grade A rests on one match per condition.

Next

• Body runs for the 7 pending drives (about 1 minute each): the two new real trials (real body n becomes 6), both giant-fibre-off trials, the pending inside-only trial (then the inside-only arm can enter the ledger) and, optionally, the two off-band gain trials; plus a repeat of the floor as a cross-session check.
• A planner decision: an inside-only scramble that also keeps the looming neurons' own outputs, or a targeted "PVLP141 silenced" arm, to test the relay directly.
• Carried: the MANC listing, the larva L1 sizes, the FlyWire citation rule and licence question (owner), the Neural Canvas code re-read. Fixed dates: browser re-checks of the 3 Oct set from 17 Oct; beginner re-test 19 Oct.

Sources

• Shiu et al. 2024, Nature, doi:10.1038/s41586-024-07763-9; model code philshiu/Drosophila_brain_model at 91bdd1e7.
• FlyWire FAFB v783 on Zenodo, doi:10.5281/zenodo.10676866; FlyWire guidelines, flywire.ai/guidelines; annotations, flyconnectome/flywire_annotations.
• fly67's committed looming numbers, validation/circuits.json at 45f62fd9.
• Card and Dickinson 2008, Current Biology 18(17), doi:10.1016/j.cub.2008.07.094.
• MaleCNS, male-cns.janelia.org; BANC, doi:10.7910/DVN/7WTH1N; hemibrain v1.2, compact summary.
• Our earlier looming test: report of 5 Oct 2026; our method for scrambled nulls: report of 1 Oct 2026.