Sixth check: who does the walking? We gave the fly brain a body, refereed the Terraria fly and added 4 entries

What changed

• 100 → 104catalogue entries (4 new: 1 A, 2 B, 1 C)
• 51 → 52gallery videos (1 new, 0 removed)
• 32 → 33control studies; 25 with a wiring null
• 0grade changes; FlyDoom (eganeganegan) still "not found", record kept

Add a body: who does the walking?

Readers arrive from videos titled "the fly brain learned to walk". The honest answer has three parts: the legs, rhythm and balance come from a body controller; the brain sends a few numbers; and someone chose how neuron rates become those numbers. We measured each part with the most-used brain model and the most-used fly body. The rate-to-drive mapping between them is ours, hand-made and not fitted, everywhere it appears.

Replay any run 100 ms at a time, with what the brain sends and the two drive numbers it becomes:

The design, written down first

The design and the gain-check addendum were copied into the pre-registration at 09:35:33 UTC, before the first brain trial (09:37:57 UTC). Every trial file holds the spike counts of all 1,299 descending neurons in ten 100 ms bins.

Deviations, in plain words

• The turning sign of P9 (changed before any body run). Bidaye et al. 2020 report that P9 "drives forward walking with ipsilateral turning": a fly with only its left P9 active walks and turns left. A direction check showed that FlyGym turns a fly toward the side with the smaller drive (drive 1.0 left and 0.3 right turned it 129° to the right). Our pre-registered formula gave each side its own P9 rate, so a left-only P9 fly would have turned right, the wrong way. We used the other side's P9 rate as each side's forward term instead. With both P9s active this changes each drive by only 0.08–0.14 per bin; with left P9 alone the fly now spins left (+254° real, +210° scrambled), as the paper says. The pre-registered drives are kept in every drive file but were not run in the body.
• The first FlyGym install failed with "No space left on device" in /tmp: with caching off, the installer unpacks a large wheel in the temporary folder. Pointing TMPDIR at a folder with room fixed it (7 s).
• An analysis fix, not a design change: the body files store the heading wrapped to ±180°, which would have read the left spins as right turns. The analysis unwraps it from the 10 ms trajectory, and the body script is fixed for future runs.

Which neurons, and why

Install, sizes and times

• Brain environment (the beginner-path pins of Build your own): installed in 36 s, both data checksums matched. It was deleted before the body install began, so the two environments never coexisted on our 3 GiB machine.
• Body: uv 0.9.0 in a throwaway environment (10 s), CPython 3.12.11 through uv (2 s, 103 MB), flygym==2.1.0 with caching off (7 s on the second attempt; 47 packages including mujoco 3.9.0, numba 0.68.0, numpy 2.5.3). Environment 602 MB; scratch peak 805 MB.
• Memory: brain trials peaked at 0.77 GB each, body runs at 0.27 GB. A body run of 1.2 s body time took about 70 s on one CPU.
• Pins are in a lock file from uv pip freeze; the steps are on Build your own.

What the brain sends down

Gate: our first sugar trial gave MN9 82.0 Hz, bit-exact with our runs of 28 Sep and 1 Oct. All 15 brain trials of this check finished, and every shuffle passed its property checks (edges, synapse total, in- and out-degree).

• Sugar sends no walking command. With real wiring every candidate walking neuron is at 0 Hz while MN9 fires at 82 Hz; the 65 active descending neurons are other types (DNge031, DNge059, DNg103 and others). This reproduces the upstream authors' own committed sugar result (0 spikes in the same 14 neurons).
• P9 fires because we stimulate it. Its rates are identical in real seed 0 and in all three shuffles, which use the same trial seed.
• The real brain adds turning commands; the scrambled brains add nothing. With real wiring, P9 recruits DNa02, more on the right than the left in all three seeds, plus other descending types (DNa06, DNa11, DNg75, DNg97 and more). The scrambled brains recruit no mapped neuron. Left P9 alone recruits the left DNa02, the same side.
• No backward command (MDN) in any trial.

Readings, as pre-registered

• Retention (null − zero) / (real − zero) on forward distance: scrambled 1.016, brainless constant 1.025, brainless random 1.002.
• Scrambled retention ≥ 0.9: "the walk does not need the fly wiring: it comes from the stimulated command neurons and FlyGym's controller".
• Brainless-constant retention ≥ 0.9: "a brainless constant command walks as well as the brain".
• Sugar: "tasting sugar sends no walking command in this model".
• The brain does change the heading, through DNa02 and our 0.5 weight: −26°, −0.4° and −14° over the three real seeds, −14.6° for the scrambled brain (from the left–right difference of the stimulated P9 rates alone). One second of walking cannot tell a brain-made turn from noise of this size: the brainless random drives turned it by −7° to −27°.

What is FlyGym's, what is ours, what is the brain's

The new ledger row

byo-flygym-body (steering a body, open loop): real 13.93 ± 0.58 mm (3 seeds), floor 0 mm (zero drive), degree-preserving null 14.15 mm, brainless constant and random drives as no-brain baselines. By the fixed rules: No difference, retention 1.016, baseline retention 1.025, baseline not beaten, fair null, strong method. Brain activity of the scrambled brains is comparable to the real one (0.818 of its spikes).

Disclosure: the rules count 3 shuffle samples, but the three shuffles sent identical drives, so in effect this is one walk. We did not override the rules, because we never use the borderline override on our own rows. The study belongs with the Build your own record, like our 1 Oct test; the catalogue gets no new entry for it.

Limitations of the body test

• One stimulus pattern (P9 at 150 Hz on both sides), one model, one body, 1 s of walking, 3 seeds and 3 shuffles (one walk in effect).
• The mapping decides how much the wiring can matter. At 100 Hz per unit, the P9 drive sits at the 1.2 cap in most bins (90 of the 120 drive values of the six P9 walks), so forward speed can hardly differ between conditions. Another hand-made mapping could let the wiring matter more. We did not tune it, by design.
• Open loop: the body never feeds back to the brain. The brain trial runs first, and the body replays its drive.
• Not tested: the pre-registered P9 sign in the body, other P9 rates, other descending neuron types (DNa01 was at 0 Hz in every trial).

Open items from 1 Oct, settled

The gain check: not matched on 3 shuffles

On 1 Oct the degree-preserving brain of our fair test stayed silent at 1.5, 2 and 3 times stronger synapses, but no gain brought its activity within ±30% of the real brain's (9,680–17,976 spikes). This check tried ×2.5 as pre-registered. Shuffle 3 at ×2.5 fired 10,138 spikes, inside the band on the first try, so no second gain was needed. At the same gain, shuffles 4 and 5 fired 24,277 and 29,297 spikes, above the band. So the arm is not matched on 3 shuffles, and "tuned equally" stays false by the existing rule. MN9 stayed at 0 Hz in all six gain trials, from 0.37 to 2.12 times the real activity: the scrambled brains mis-route the sugar signal; they are not just too quiet. Different shuffles need different gains, which a per-shuffle gain search could settle (a candidate for a later check).

Notes fixed at the source

The controls data for our 1 Oct test said "Gain-matched arm: not reached (clock cut)" and "fewer shuffles than pre-registered (clock cut)". Both were stale: the gain arm ran, and all 5 pre-registered shuffles per arm ran. The data now gives the full gain bracket and says "all 5 pre-registered shuffles per arm ran". Our public copy no longer rewrites these notes.

fly-cartpole: updated to commit 077dcba

We read the author's commits between 0455e2e and 077dcba: same task and metric, the same degree-preserving null and 20 fresh seeds per arm, with both arms under the new training protocol (a posture-shaped reward and 3,000 tuning episodes). Judged like for like, the ledger now uses 077dcba: 392.5 ± 103.1 steps with the real mushroom-body wiring against 390.3 ± 96.4 shuffled (p = 0.48); frozen 18.7, a TD learner 500.0 (the ceiling), random 22.8. Retention 0.994; still no difference, fair and strong. The results file is unchanged at the current head (854fb1d).

Dropped as planned: a narrower boundary null (moved to a later check) and five more shuffles per arm (they could not change the readings).

No scoring rule changed; one display-only field added

Our 1 Oct test scores "mixed, unfair, weak" by the fixed rules, partly because the sign-shuffle brain ignites and so fires MN9 above the real rate. We did not change the effect, fairness or quality rules: only our row would change, a rule tuned now would be fitted to it, and "unfair" is the rule doing its job while the gain arm is unmatched.

Instead every ledger arm gained two display-only fields: activity_ratio (whole-brain activity, null ÷ real) and activity_class (reduced below 0.5, comparable 0.5–2, elevated 2–10, runaway above 10, or not reported). They never enter any score. Values: our fair test, degree-preserving 0.291 reduced, weight shuffle 0.388 reduced, sign shuffle 257.6 runaway, boundary-preserving 0.648 comparable; our body test, degree-preserving 0.818 comparable. All 95 other arms: not reported, because no other source reports network activity and we did no new hunting. An activity-aware rule could be adopted later only through a pre-registration made before the data it judges, or once at least 3 studies not made by us report network activity; it would then apply to every row, with before and after counts published.

What people argued about this week

Window: new items and new attention since 1 Oct 2026, 11:00 UTC. The promotion rule is unchanged: a specific, checkable claim about fly-brain-driven behaviour, new attention (for example 20,000 or more new views on a video), and no verdict yet; at most 2 claims.

Re-checks: Gorilla Tag 623,504 views (+95,748; verdict U, no code); the Roblox fight 141,595; "Fly's Brain Played Video Games" 267,918 (+53,607; an explainer with no single checkable claim, not promoted); "THEY TRAPPED A FLY'S BRAIN" 152,904; FlyLeno's stream 44,716; the Rainbow Six short 4,754. No widely seen "fly brain learned to walk" video appeared this week.

Verdicts

Terraria: B

"I put a fly's brain into Terraria. And tried to train it" (Russian title; George Ostrobrod, uploaded 29 Sep 2026; 28,244 views on 2 Oct). Code: gitlab.com/wdf.gost/terraria-wdf-fly-brain, MIT, read at commit 9052966d of 26 Sep 2026, three days older than the video.

• What runs: the MaleCNS v1.0 connectome (166,700 neurons, 25,088,107 signed connections; the input files' checksums are enforced) as a sparse spiking model with adaptation, with reward-modulated plasticity driven by the body's own energy, water and harm signals.
• What a person chose: which sensory neuron groups receive the game's light, smell, touch, pain and balance (the author calls the receptor assignment "approximate"), and how motor neuron groups are decoded into turning, walking, feeding, grip and adhesion. Grip and adhesion have documented fall-backs when the graph lacks those neurons.
• Result: we recomputed the author's 176-life log: mean lifespan 278.4 s and the quarter means match the author's report; the median rose from 102.8 s (first 44 lives) to 163.3 s (last 44; our one-sided permutation p = 0.037, Spearman ρ 0.11).
• Control: none, as the author's own report says ("no fixed-seed control organism with plasticity disabled"). So the "learning" is not separated from luck or from changes to the world during the run.

Row updates

• Gorilla Tag 623,504 and Roblox fight 141,595 views; both stay U.
• FlyDoom (eganeganegan) returned "not found" on 1 and 2 Oct; the record is kept with its last catalogued commit.
• The fly.ai kernel behind fly-plays-games gained NFT breeding and arena code only; no re-grade.
• Is it real? gained a "Who does the walking?" section near the embodied claims. It changes no grade.

Tracking

• All 100 entries re-checked: 88 repositories by git ls-remote (6 changed, 81 unchanged, 1 gone); every try, code and paper link checked (98 ok, 2 redirects, 3 blocked by rate limits, 1 not found: FlyDoom).
• Changed repositories: acamilo-flybrain, Brain Runners, Doodle Fly and fly-naf (commit subjects only, no result change seen); fly-cartpole (above); the fly.ai kernel (NFT code only).
• All 51 videos available. Changes: 4 new entries, 2 material and 7 minor updates, 0 removed.

New entries

Every new entry carries controls and wiring_effect fields (all "none" and "not tested"). Excluded: a Terraria mod by the same author with no fly brain. Deferred a second time for lack of time: three small repositories (a chess fly, a FlyLeno variant and a poker fly), plus about 30 new repositories with 0–3 stars from this week's sweep.

Gallery

51 videos re-checked, all available. Added 1: the Terraria fly, the creator's own upload (28,244 views, grade B from the catalogue). Not added: the creator's companion stream (3,072 views) and explainers that name no single entry. Thumbnails are unchanged while the owner decides between platform thumbnails and our own posters. We did not watch the new video (no browser this check).

Build your own and the package check

The beginner pins still install for Python 3.11: Brian2 2.9.0 pinned (latest 2.10.1), numpy 2.3.5 (latest 2.5.3), pandas 2.3.3 (latest 3.0.6); the lock installed cleanly in 36 s. FlyGym 2.1.0 is still the latest release and needs Python 3.12 to 3.14. The latest uv is 0.12.22; we used 0.9.0, the version of our first FlyGym test. The guide's deeper path "Add a body" is now tested, with the TMPDIR note and the FlyGym 2.0 interface warning.

Method

• Brain: Python 3.11 environment from the Build your own lock file; our fair-test worker extended with a stimulus choice and the descending-neuron recording (spike monitoring does not change the dynamics).
• Body: Python 3.12 FlyGym environment, MUJOCO_GL=disable; our drive maker (the mapping) and body runner, then a standard-library analysis script.
• Tracking and catalogue: git ls-remote, link checks run twice (a rate-limit answer now counts as blocked, not broken), video checks, GitHub metadata for the 3 new GitHub repositories, a catalogue diff and schema validation (projects 104, videos 52, controls: 0 errors each).
• Lead scan: news, GitHub and YouTube scans; view counts from the search pages because watch pages were rate-limited.
• GitHub API: 3 core calls (59 left at the start, 57 at the end) and 18 paced search calls.
• Memory: one worker at a time; brain trials peaked at 0.77 GB, body runs at 0.27 GB; the shared machine peaked at 1.94 GB; scratch peaked at 805 MB and was deleted at the end. No out-of-memory kill.
• Timing: one 48-minute session; every go/no-go passed and no time cut fired. Three optional extras were not done (flyvis numbers from the papers, a star refresh and a gap-filling gain trial).

Limitations

• The body result holds for one stimulus, one hand-made mapping and 1 s of open-loop walking (above).
• YouTube watch pages refused every request today, so video descriptions were not re-read; the Terraria claim quotes the description read on 1 Oct. X, TikTok and Reddit cannot be searched without a login.
• fly-walking-wiring's controls cannot be inspected because no results are public, so its B is provisional.
• The Terraria verdict rests on code older than the video.

Next

• Next check (no brain simulations planned): the flyvis and browser-demo checks, a re-read of fly-naf's new vision mapping, triage of the deferred repositories, and whether fly-walking-wiring's results appear (then it gets a ledger row).
• Later: a pre-registered per-shuffle gain search for the fair test; the narrower boundary null; and a second body test where wiring could matter, for example a mapping without the 1.2 cap or P9 at lower rates, plus a closed loop, written down before it runs.
• Gorilla Tag keeps growing (623,504 views) and still shows no code.

Main sources

• Brain model: philshiu/Drosophila_brain_model @ 91bdd1e7; Shiu et al. 2024, Nature 634, 210–219, doi:10.1038/s41586-024-07763-9.
• Body: NeLy-EPFL/flygym 2.1.0; Wang-Chen et al. 2024, Nature Methods, doi:10.1038/s41592-024-02497-y.
• Descending neurons: Bidaye et al. 2020, Neuron, doi:10.1016/j.neuron.2020.07.032; Bidaye et al. 2014, Science, doi:10.1126/science.1249964; Rayshubskiy et al. 2025, eLife, doi:10.7554/eLife.102230; von Reyn et al. 2014, Nature Neuroscience, doi:10.1038/nn.3741.
• Terraria fly: wdf.gost/terraria-wdf-fly-brain at 9052966d; the video.
• fly-cartpole: Curt-Park/fly-cartpole at 077dcba (results file unchanged at 854fb1d).
• New entries: arisliwind/flygym at 8a794e33; kosmoKwan/fly-connectome-wiring at c84104a; xiangdoz/fly-connectome-attractor at 07b01c8.
• Our data: controls.json (version 2026-10-02-run6), projects.json, videos.json; the chart as an image: who-does-the-walking.svg.
• Previous check: report of 1 Oct 2026.