Seventeenth check: the right-eye relay reading with 10 matched controls

The relays: 5 more matched controls on the right side

Our own test · pre-registered · brain only (Shiu et al. LIF on FlyWire v783) · one model · not a real fly

What we did. A looming shadow on the right eye drives the turn neuron on the far side (left DNa02). On 13 Oct we compared silencing the relay neurons (E2, and PVLP141 alone, P) with silencing other neurons matched one by one on activity, in 5 draws. Today we added draws 5-9 with the same matching rule and the same reading rule, on the right eye only; the left eye's reading cannot change with more draws under this rule, so no left trials were run.

Checks before reading (both passed). The draw maker reproduced the 13 Oct right-side draws 0-4 exactly (the same neurons, seeds and tries). A re-run of the 13 Oct right-eye seed-0 trial matched it on the four registered criteria (input file hash, 43 far-side DNa02 spikes, 27,446 whole-brain spikes and the set of active neurons); only timing and memory fields differ. We read "bit-equal" in the plan as these four criteria. The 5 new draws were each accepted on the first try, and no trial was cut.

Band of the 10 controls: 38-47 Hz (5 on 13 Oct: 40-43 Hz). 0 of 10 controls at or below E2, and 0 of 10 at or below PVLP141 alone, so both right-eye readings stay "fell more than matched activity". The left eye keeps its 13 Oct values and reading ("fell, like matched activity").

Descriptive only. One new control (draw 7, 38 Hz) is below the lower edge of the band with nothing silenced (40-49 Hz), so the controls now overlap the bottom of that band. 2 of the 10 right-side draws (draws 1 and 4, both from 13 Oct) include a neuron that synapses directly onto the far-side DNa02; none of the new draws does.

8 new entries and 5 new ledger rows

Project claims are the projects' own; ledger numbers are the authors', scored by our fixed rules and not recomputed by us.

• FlyScore A: a wiring-specificity benchmark on MaleCNS v1.0. Sugar to MN9: real 1.0 (40 trials) vs 0.0 for 5 weight-shuffled and 5 rewired copies; ledger: helps.
• Flybrain (prove-ai) A: seven MaleCNS v1.0 projects; the grade rests on its connectome language model (bits per byte, lower is better): fly 2.6388, degree-rewired 2.6286 (15 graphs), random graph 2.5827, a parameter-matched GRU 2.5575; ledger: mixed. AUW160150/flybrain has the same commit (f42f586) and is named in the entry as a mirror; neither GitHub page says "forked from".
• FlyLab pharmacology A: predicted imidacloprid effect on a 5k-cell MaleCNS v1.0 cut, real -0.678 Hz vs rewired -0.267, weight permutation -1.102, random -0.974, sign permutation -0.379 Hz (100 graphs each); ledger: not scored, because the metric has no better or worse direction.
• MaleCNS_training A: held-out tactical accuracy on 12x12 gomoku, fly 0.787 vs rewired 0.785, RNN 0.788, GRU 0.790; ledger: no difference. We did not read its rewiring code or graph count.
• fly-connectome-study A: a pre-registered foraging test over 200 worlds, real 44.0 vs shuffled 44.36 (labelled degree-preserving), random 50.905, linear learner 46.17, floor 0.485; ledger: mixed.
• flylab-demo C: a local NeuroMechFly/FlyGym lab with a four-neuron FlyWire v783 fragment; its "no edges" run is an ablation, not a wiring null.
• discord-fly-brain C: a Discord fly agent on hand-designed weights; no wiring control.
• Roblox avatar claim U: a recap video says fruit-fly brains control Roblox avatars. The builder is unknown; we did not watch the video and found no code. It is a separate claim from fruitfly-roblox and fly-jjs-fight (verdict row).

The ledger now has 58 studies, 49 with a wiring null: helps 22, no difference 11, worse 4, mixed 10, not yet scored 2 (Does fly wiring help?).

Re-graded from code: Pokémon on a Raspberry Pi 5

The repository ianwes92/fly-brain-plays-pokemon (MIT, published 14 Oct 2026) names the Twitch stream in its README, so we graded the stream's verdict from the code: U to B, borderline C (our judgement call; a stricter reading could give C).

• Wiring and model: FlyWire FAFB v783, connections of five or more synapses, run as a whole-graph leaky integrate-and-fire network.
• Input: the game's memory (RAM), not the screen: walls, people and doors are mapped onto visual neurons (LPLC1, LC10, LPLC4).
• Output: descending-neuron rates pass through a hand-written priority arbiter that picks the key.
• Chat: chat keys are added to the fly's keys, and a paid override replaces the fly's keys for a while.
• Learning: only Kenyon-cell to MBON synapses change (chat !pain / !pleasure and code-side rewards).
• Controls: no shuffled-wiring control. We did not watch the stream.

Tracking and freshness

• Repositories: 130 checked: 123 unchanged, 6 with new commits (fly-ai, help-the-fly-escape, navis, flywire-connectome-scz, malecns-keratitis-pilot, neuropest), 1 still gone (flydoom, record kept).
• Links: 328 checked after one re-check: 284 answered, 16 redirected, 19 blocked our checker, 6 client errors, 0 server errors, 3 could not be reached. Both Zenodo paper links that timed out on 14 Oct (fly-brain by erojas, flybrain-connectome-benchmark) resolve again. fruitfly.live still has a TLS certificate error.
• Verdict rows: the repositories of 14 of the 22 verdict rows re-checked; none has new code since grading. One row is new (the Roblox avatar claim).
• Videos: 56 of 56 available on 15 Oct 2026; no video added or removed.
• Lead scan: 56 code repositories (11 new) and the weekly video searches reviewed; nothing new met the bar for a catalogue entry beyond the two follow-ups above.

What we could not resolve

• The rule's direction. Our rule compares the relays with the lowest matched control, so extra controls can only make the relays look less special, never more. So 10 controls that all stay above the relays are a narrow check: they show that this reading did not change, not why the relays lower the rate.
• Small samples. 5 seeds per side; 5 matched draws on the left and 10 on the right; the relay arms have 3 seeds each; one brain-only model with an input we set; a simulated rate, not fly behaviour.
• E2 is mostly PVLP141. PVLP141 accounts for most of the activity removed by silencing E2 on each side, so the test says little about the other relay neurons.
• Grades. Several grading calls were made from the files named above and not all were re-read line by line; Flybrain's A rests on one sub-project; the Pokémon B could be C under a stricter reading.
• Roblox claim. Who built the "fly" account is unknown; the earlier NiftRex video and the Roblox user profile are the next things to check.
• NeuroPest. Its last commit shows "n/a" in our data, and its v0.7.0 release subject was not verified.

Data behind this report: projects.json, controls.json, videos.json. The 13 Oct design and every earlier trial: report of 13 Oct 2026.