Does the whole fly brain compute anything a small table of 56 numbers cannot?

E47 — The whole brain is a 7x8 matrix at this operating point, and the shuffle wins by a shortcut

Three checks, each able to retire the architecture on its own. Two of them fire.

Settling. Row-normalisation bounds the row sums at one, so for gain below one the dynamics are contractive with a conservative decay time of tau_max/(1-g) - 200 ms at gain 0.9, four times the 50 ms the readout was taken at. Measured against an 800 ms reference, the descending direction is already there: cosine 0.976 at 25 ms, 0.990 at 50 ms, 0.999 at 100 ms. The magnitude is not - the mean rate is 1.45e-06 at 50 ms against 3.14e-06 converged. Since the heading is normalised, 50 ms costs about 1% of direction. Starting from a settled state rather than from rest also moves the direction by cosine 0.990, so the persisted state and the remaining settling are the same size.

The shuffle takes a shortcut. Receptors reach the descending neurons in a median of 3 hops through the measured wiring and 2 hops through the shuffle (1325 of 1326 descending neurons reached in both). A shorter route mixes the signal less, so part of the shuffle's decoding advantage in E46 is route length rather than wiring. The E46 comparison stands as a measurement; its interpretation does not, and must be stated as route length plus wiring rather than wiring alone.

The whole brain collapses to a 7x8 matrix. Fitting d(x) = A log1p(x/0.05) + b on 160 calibration compositions and testing on 80 held out, the affine map reproduces the descending-neuron heading at cosine +0.980 mean, +0.989 median. The reason is in the rectification: 84.3% of neurons never change their active/inactive state across compositions, and the descending population idles at a mean rate of 3e-06. At this operating point the network is not using its nonlinearity, so 163,972 neurons and 6.1M edges are computing what fifty-six numbers compute.

Pre-registered failure condition 3 has fired (PREREGISTER_wholebrain.md): the claim that whole-brain complexity is necessary is retired, whatever the search numbers say.

The end-to-end search, at this operating point. Cluster-expansion mixing energy at the random limit, 16 walkers x 60 steps, 8 seeds, identical objective, budget, starts and readout projection:

The brain searches 41 meV/atom worse than a random direction, and 24 worse than its own shuffle. The spreads are over eight search seeds, which is the right denominator for search noise; the shuffled arm is a single graph, so graph-to-graph variation is not in that figure and no significance is quoted for the measured-against-shuffled difference. This arm resets the membrane state between moves, which makes the heading a fixed function of composition, so a walker that has just failed proposes the identical step until the restart rescues it; that degeneracy is real and is corrected by persisting state, but it cannot explain an affine map fitting the headings at cosine 0.98.

What this establishes. Not that a connectome cannot generate a search. That the architecture as built operates the brain in a nearly linear, nearly silent regime where its wiring cannot express anything a small matrix could not, and that in that regime it searches worse than chance. The diagnosis names the fix - an operating point where the rectification pattern actually moves - and the fix is testable against the same pre-registered conditions.