Muscle-resolved VNC readout, and the rhythm question answered

[abgnydn]

Three things: instrumentation, a builder fix that is inert until the blob is rebuilt, and the analysis that answers whether this connectome can walk.

Muscle-resolved motor readout

main.ts averaged each leg's ~60 motor neurons into one scalar, which cancels every flexor against its extensor by construction. vnc.meta.json already partitions the 369 leg motor neurons into 6 legs × up to 20 named muscle targets, and motor_by_target was loaded and never read.

The readout now resolves per muscle and computes antagonist ratios for the three joints with a true pair in all six legs — tibia, femur/trochanter, coxa_twist — pooled per side and normalised, since the pools are lopsided (14–15 flexor MNs against 2 extensor at the tibia). tergotrochanter is excluded: it is the escape-jump muscle and does not belong in a walking drive.

Instrumentation only. driveLegs is untouched and nothing consumes the numbers yet. Measured live under a DNa01 stim: per-muscle values vary over time and differ between legs.

Glutamate sign

build_vnc.py signed glutamate −1 for every neuron. Correct centrally (fly glutamate is inhibitory via GluCl), wrong at the neuromuscular junction where it is excitatory. Motor neurons are now gated to +1. build_csr.py deliberately unchanged — the brain has no NMJ.

Scope, measured rather than assumed. Of 709 motor neurons, 580 (81.8%) carry sign −1 — but that sign is read in exactly one place in src/, a boot log line, and the weights motor neurons receive come from their presynaptic partners and are untouched. Motor neurons emit 0.126% of total synaptic mass. The maximal counterfactual moves the structural result 0.67% and the rhythm sweep not at all. This is a labelling and biophysics correction, not a dynamics fix.

Inert until rebuilt: data/manc is not in the checkout, so vnc.bin was not regenerated. Re-run download_manc.sh → build_vnc.py → write_manifest.py → upload_to_r2.sh, and re-verify walking afterwards — the drive constants in main.ts were tuned against the current blob.

Can this connectome walk? No — and now with a proof

tools/vnc_rhythm.py, reproducible against the shipped bin:

• It does oscillate — 5–12 Hz, spectral prominence 58× to 1.8e5× over background. An earlier claim that it never oscillates was wrong.
• But it is global synchrony, not gait. All six legs peak at one frequency in phase. Kuramoto R_sync 0.475–0.910 against R_tripod 0.049–0.331, and across every probed setting the best-fitting 3/3 split of the legs was never the true tripod.
• Tripod alternation is unreachable by tuning, and this is a proof rather than a failed search. The excitatory fraction is 0.517–0.632 in all 36 leg-pair couplings, so nulling across-tripod coupling needs an inhibition multiplier above 1.716 while keeping within-tripod positive needs it below 1.326. The interval is empty.
• 600-setting sweep (100× weight, 20× drive, 16× E/I): 0/600 whole-leg rhythm, 0/600 muscle-resolved under the strict criterion. Resolving antagonists does not reveal a hidden rhythm — that hypothesis is dead.
• The tripod signal in the wiring is fully explained by anatomy (segment distance and ipsi/contra); residual after removing it is identically zero. The connectome is organised by proximity, not by gait. Pool-level averaging is the wrong lens; any future attempt needs per-neuron structure.

Verification

tsc clean · unit 5/5 · full e2e 34/34 on an idle machine.

🤖 Generated with Claude Code