Sixteenth check: the looming turn in degrees, at each input strength

The looming turn in degrees, at each input strength

The question: when the brain model's turn command appears at about 30 spikes per second per looming neuron (our 8 Oct brain series), does the simulated body turn away at the same strength, and by how much? We fixed the rule and the possible sentences before the first body run: a run turns away when its turn away from the shadow, minus the turn of the same body with no brain input, is more than 10°.

How the runs went

• Two checks before reading. Every one of the 12 brain drives matched its saved fingerprint (sha256), and the no-brain floor run repeated our 8 Oct floor exactly (0 of 700 trajectory values differ; heading after onset −1.49°). Both passed.
• Runs. 12 new body runs plus the floor, 63.6–78.6 s each (median 71.7 s); none was cut by the clock or stopped. The body software was installed fresh with the same locked package list as on 8 Oct.
• Which runs are new. Seed 2 at 30 and 80 spikes/s walked on 8 Oct; the other walked points are new. Drives that are identical to the no-brain drive (left eye 10 and seed 0 at 20; right eye 10 and seed 2 at 20) take the floor's walk by determinism and are marked "same as no brain".
• Saturation. None of the 12 drives was clipped by our mapping (clipped share 0.0 in all).

Deviations, disclosed

• The plan first said that the 7 brain drives of 6 Oct still waited for body runs. They had walked on 8 Oct; this run walked the 12 drives of 8 Oct that had never walked instead. The 6 Oct table on Is it real? now shows those 8 Oct walks in its Body column.
• The pre-registered left-eye relay sentence reads "about 37 deg instead of 37 deg"; we keep it as registered and show the unrounded values (37.3° vs 37.2°) beside it.

Under the relay headline: body runs with relays silenced

The 13 Oct brain-only test of the relays stays the headline, in its pre-registered words (report of 13 Oct):

The two sides disagree: on one side silencing the relays lowered the turn neuron's rate more than matched activity elsewhere, on the other it did not. We do not draw a conclusion.

The body lines describe single runs at 80 spikes/s against the intact 80 spikes/s run of 8 Oct (for the right eye that is the 6 Oct drive, walked on 8 Oct). They do not answer the relay question.

4 new entries, no new ledger rows

From the backlog of 21 candidate repositories, each was checked against one file rule: committed code with a result file or a wiring comparison at the checked commit. 4 were graded (the cap for this run), 10 were excluded with today's date and 8 are carried to the next check. Project claims are the projects' own, not our findings.

• B fly-cx-navigation-servo (miwamasa): a MaleCNS v1.0 central-complex rate model (2,308 neurons) with two in-repository papers on heading, navigation and servo control; analysis outputs committed as JSON.
• B NeuroPest (IrohAmca): a desktop fly pet driven by a pruned FlyWire v783 spiking network with hand-made inputs and outputs and engineered rules; raw probe tables committed. Linked by its source repository only (we did not download its Windows build).
• B Fly Garden (carsonshicks2016-beep): a local 3D lab running the full FlyWire v783 network in Brian2 with a NeuroMechFly body; its README calls the live model an engineered hybrid, and many measured reports are committed.
• n/a FastBrain (ArcadiaE): a GPU speed-up fork of a whole-brain model benchmark; graded n/a like its upstream, because it is a simulator speed-up, not a behaviour test.

None of the four compares the real wiring with a scrambled or rewired copy, so the controls ledger is unchanged: 53 studies, 44 with a wiring null (helps 21, no difference 10, worse 4, mixed 8, not yet scored 1).

Tracking, freshness and corrections

• Repositories: 126 checked: 119 unchanged, 6 with new commits (brainlab, fly-ai, flyvis, flywire-connectome-scz, flight-test-the-fly, help-the-fly-escape), 1 still gone (flydoom, record kept). The commit dates of four catalogue entries were updated.
• Links: 308 checked: 263 answered, 16 redirected, 19 blocked our checker, 6 answered with the same client errors as before, 1 server error answered on a second try, and 3 could not be reached. Two of those are the Zenodo paper links of fly-brain by erojas and flybrain-connectome-benchmark: the connection to zenodo.org timed out twice on 14 Oct, which we do not read as broken; they will be checked again. fruitfly.live still answers with a certificate error.
• Verdict rows: the repositories of 14 of the 21 verdict rows re-checked; none has new code since grading. The other 7 rows are carried.
• Videos: 56 of 56 available on 14 Oct 2026; no video added or removed.
• Lead scan: 70 code repositories and 18 video leads reviewed; nothing met the bar for a verdict or a catalogue entry this time.
• Correction: fly64's causality citation now points to scripts/validate_causality.py lines 37-47 (the last two checks of the block were lines 46-47).

What we could not resolve

• One run per drive. Each point is one body run; the right-eye reading turns on a single run 0.8° under the line, and another seed could read differently.
• Our mapping. The degrees come from our own hand-made rule that turns the turning neurons' rates into steering, with a walking drive we hold constant. They restate the brain result in the body; they do not add a test of the wiring, and no scrambled brains were run at these strengths.
• Not a steady rise. The turn did not grow at every step (left eye 11.1, 27.2, 26.9 and 37.2° at 30, 40, 60 and 80 spikes/s; right eye 16.8, 33.8, 62.2 and 55.4°). We did not test why.
• Relay body lines. One run per arm and side, read descriptively; they do not change the 13 Oct relay headline.
• Zenodo links. The two time-outs were not verified as real problems.

Data behind this report: projects.json, controls.json, videos.json. Earlier steps of the looming test: How strong must the shadow be? (8 Oct) and We ran the map (5 Oct).