Eleventh check: the relay neurons read by firing rate, and what the map leaves out, in numbers

What changed

• Is it real?: the relay panel is now the firing-rate chart, led by its summary sentence; the 8 Oct sentences stay only with the label "by the left/right contrast (8 Oct)". The "Is it a brain or just a map?" answer links to the new table.
• How it works: a new section, What the map leaves out (8 rows, checked 9 Oct 2026), and the same relay chart.
• Projects, Does fly wiring help?, Video gallery and For AI agents: rebuilt from this check's data (version 2026-10-09-run11).

Which relays carry the turn? Read by firing rate

The map has no direct synapses from the looming neurons (LC4 and LPLC2) to the turning neurons DNa02. Its two-step routes to the far-side DNa02 run through 21 relay neurons (left eye) and 23 (right eye); 17 / 20 of them excite DNa02 in the model and 4 / 3 inhibit it. On 8 Oct we silenced PVLP141 alone and then all of them together, which mixed both kinds.

The design, written down first

• Model: the Shiu et al. leaky integrate-and-fire model (code 91bdd1e7) on FlyWire v783, brain only, 1 s per run, every LC4 and LPLC2 of one eye driven at 80 spikes per second; Brian2 2.9.0, numpy 2.3.5, Python 3.11; the data files matched our earlier runs byte for byte.
• Turn command = the far-side DNa02 firing rate during the 1 s shadow. Normal range = minimum to maximum of three runs per eye with nothing silenced (seeds 0, 1, 2). % of normal = mean silenced rate ÷ mean normal rate, rounded to the nearest 5.
• Readings per arm and eye, the first that holds: "gone" (20% or less), "fell" (every run below the normal range), "rose" (every run above it), "within normal range" (every run inside it), otherwise "unclear". An arm's reading is the eyes' common reading, or "sides differ".
• Arms: PVLP141 (1 cell per side); the excitatory two-step relays (17 / 20); the inhibitory ones (4 / 3); random active neurons (17 / 20, drawn with a fixed seed from neurons that fired in the run with nothing silenced, excluding the looming neurons, all 1,299 descending neurons and all relays); and, read by the same rules, the 8 Oct all-relays runs.
• Summary sentence: fixed in advance for four cases. The random-neuron control comes first: if it is not "within normal range" on either eye, the summary is "not specific to the relay neurons", whatever the relays did.
• "Silenced" removes the neuron's output synapses only; a unit test checked, for every list, that only the listed neurons' outgoing synapses changed.

Deviations, in plain words

• The unit test of the random-neuron lists ran after the random-neuron runs (the lists are written automatically during the queue); it passed.
• Numbers are filled in per eye ("17 (left eye) / 20 (right eye)"); the words are unchanged.
• The summary rule was read as "the random control is not within its normal range on at least one eye".
• The environment check run (left eye, seed 2) repeated our 6 and 8 Oct run exactly, so the planned seed-2 runs from 6 and 8 Oct were reused rather than run again.

Limits of this test

• One model, brain only. The input is a rate we set, not a measured response to a real shadow; not a real fly.
• The normal range comes from 3 runs per eye, and the left range is only 1 Hz wide. One random run at 26 Hz, 1 Hz above it, set the summary, while every excitatory-relay run (14-19 Hz left, 25-26 Hz right) sits below every run with nothing silenced.
• The random-neuron control is one draw and one seed per eye. A fairer control would match the silenced neurons' activity and use several draws and seeds.
• Described, never read: of the silenced relays, 4 of 17 (left) and 3 of 20 (right) excitatory, 2 of 4 and 2 of 3 inhibitory, and PVLP141 fired in the runs with nothing silenced; the excitatory relays without PVLP141 (one run per eye) gave 30 and 41 Hz.

What the map leaves out

People ask whether a fly brain simulation is "a brain or just a map". We put one number and one source on each part of biology that the map-based model leaves out or simplifies. The in-model counts come from our own copy of the model: by predicted transmitter, 86,025 acetylcholine, 24,858 glutamate, 19,147 GABA, 5,905 dopamine, 2,201 serotonin and 210 octopamine neurons (279 without a prediction); 9,059,302 excitatory and 6,032,681 inhibitory connections.

Short answer: The model is one fly's wiring map: 138,639 neurons and 15,091,983 connections, run with one simple neuron rule. It leaves out electrical synapses (gap junctions), neuropeptides, learning, the different behaviour of different kinds of neurons (including neurons that do not spike), the differences between flies and over a fly's life, and the nerve cord and body. So it can show what the wiring alone does, not everything a fly's brain does.

• 138,639 neurons
• 15,091,983 connections (neuron pairs)
• 54,492,922 synapses
• 1 hand-set strength (0.275 mV per synapse)

Quotes behind each number: the neuropeptide count is a review's statement; the innexin page answered a bot check, so its full text was read through NCBI's E-utilities. Nature pages answered a bot check to scripts, so the PubMed Central copies were read. Method: report of 9 Oct 2026.

flyvis: ready for a wiring test

flyvis (TuragaLab) is a trained model of the fly's motion-vision circuit built on its connectome. We read its code (commit 92b3845c) to see whether a scrambled-wiring test can run on a normal computer. Every item passed; the test is planned for the next check.

Lead scan and verdicts

• Nothing promoted. The only new item above 20,000 views retells Eon's March 2026 announcement, which our Eon verdict (U) already covers: a There's An AI For That video, approximately 103,764 views (watch page, 9 Oct 2026).
• Row updates: the "learns to hack" video (dzuma) reached approximately 54,805 views (YouTube search page, 9 Oct 2026, video page) and still links no code; the grade stays U. Idor's Spanish explainer reached approximately 567,371 views (search page, 9 Oct 2026) and joins the gallery.
• Verdict freshness: all 21 verdicts re-checked on 9 Oct: 13 with code unchanged since grading, 8 with no code to re-check, 0 changed.
• Coverage: YouTube views-this-week and upload-date searches (28 of 28 pages, in English, French, Spanish and Russian), Hacker News (no fly-brain stories), Google News (nothing new), dataset pages (no new release), Fly Brain Hub and flybrain.info (unchanged). Reddit answered HTTP 403, so it was not covered.

Four new entries and the ledger

• FlyVL A: a frozen male-CNS connectome as a motion encoder. In the author's committed result files, real wiring scores 90.7% vs 59.8% for a rewired copy (+0.309, 95% CI 0.287-0.332, 3 seeds), but a classical Reichardt motion detector scores 99.8%, and the author's stricter controls remove the gap. It joins the controls ledger as "helps" by the fixed rules (fair, strong; the no-brain baseline is not beaten).
• Flappy Fly B: a trained readout of the frozen MaleCNS connectome plays a Flappy-style game; it matched a no-connectome controller (7.00 pipes each) only in its sixth version, after hand-tuning.
• FlyWire Playground B: browser scenarios on the FlyWire brain model; it drops connections under 5 synapses and has adaptation on by default, so it is not the unchanged Shiu et al. model.
• Let It Flyppen C: music from fly-brain circuits through trained readouts; the brain receives only its position in the song, and the result is in the README only.

Controls ledger: 44 studies, 35 with a wiring null: the real wiring helps in 17, makes no difference in 8, does worse in 2, gives mixed results in 7 and is not yet scored in 1. Our own looming row is unchanged; its note now carries this check's rate readings, which never enter the effect.

Tracking and checks

• 111 code repositories: 107 unchanged, 3 changed, 1 gone (flydoom, record kept). fly-with-me added commits up to v2.0.0; its training document describes a branch with 9,978 trained synapses while saying the main branch never changes the connectome, so its grade (B) and "no trained part" stay until a re-check.
• 255 links: no new failures; 3 now answer bot checks.
• 55 videos re-checked, all available; 1 added (Idor's explainer, linked to the projects it discusses).
• Python packages: pyarrow 26.0.0 is out (we pin 25.0.1); Brian2 2.10.1 and numpy 2.5.3 are newer than our pins (2.9.0, 2.3.5), which stay until a test has been run with them.
• MANC's public bucket still holds only v1.0 (19 files, about 18.5 GB).

Method and limits

• One simulation process at a time, each started below a pre-set memory limit: 25 new brain runs, plus 4 reused from 6 Oct (bit-exact) and the 2 all-relays runs of 8 Oct. Every run has its own result file; the catalogue and ledger builds were run twice with identical outputs.
• Every number in What the map leaves out comes from a page read on 9 Oct 2026 or from our own model files. Pages that answered a bot check are named in the table notes. There is no single "share missing" figure: the rows measure different things.
• New grades rest on the authors' result files as read by us, not recomputed. The scan covers two search pages per term at one moment; TikTok, Instagram, X and Reddit are not covered.
• The run finished within its plan (about 36 of 45 minutes); no planned cut fired.

Next

• flyvis: a scrambled-wiring test on a normal computer.
• The relays: an activity-matched random control with more draws and seeds (for example 5 × 3 per eye), and more runs with nothing silenced for the normal range.
• A coupled brain-and-body looming test, in which the fly's movement changes what the eye sees, is planned; it is not yet tested.
• Four repositories queued for grading, and the fly-with-me re-check.