Does the real fly wiring search alloy space better than a shuffled copy of itself?

PREREGISTER_wholebrain.md · lines 1–60

Pre-registration: does the measured wiring search better than a shuffle?

Written before the end-to-end comparison finished, so that the conditions under which the hypothesis fails are on record rather than chosen afterwards.

The hypothesis, narrowed

Not "a biological brain is a better generator". That is not falsifiable and there is no principled reason to expect a fly's wiring to suit alloy search, which has no relation to anything a fly does. The claim under test is:

Under a declared distribution of interfaces, starting compositions and objectives, the measured connectivity induces proposal statistics that make the search reach a lower mixing energy per evaluation than degree-preserving shuffles of the same graph.

What is held identical across arms

Objective, evaluation budget, starting compositions, step size, minimum fraction, restart patience, the readout projection seed, and the search seed. The common-mode calibration is performed separately for every graph from the same objective-free reference compositions, so that a control does not inherit the measured brain's centring and acquire a constant drift.

Declared failure conditions

1. The measured brain loses or ties end-to-end under this protocol. The practical superiority hypothesis fails for this protocol. It is reported as a failure, not reframed.
2. The advantage holds for only one interface. Element-to-glomerulus assignment and the motor projection are both arbitrary. If the result survives only one of them, the claim is restricted to that composite algorithm and may not be stated as a property of the connectome.
3. A 7x8 affine map reproduces it. Row-normalisation makes the dynamics contractive for gain below one, so if the rectification pattern is stable over the compositions visited, the whole brain collapses to d(x) = A log(1 + x/0.05) + b. If that small map reproduces both the headings and the search performance, the claim that whole-brain complexity is necessary is retired, whatever the end-to-end numbers say.
4. A matched random proposal field reproduces it. If a synthetic field matched for magnitude, covariance and spatial correlation searches as well, the finding is about proposal geometry and not about this connectome.
5. The decode gap is explained by path length. Shuffling can shorten the route from receptors to descending neurons. If the measured and shuffled graphs differ in sensory-to-descending path length, the 50 ms decode difference is partly a measurement of route length and must be reported as such.

What may not be claimed

No significance from three null graphs: a randomisation test over three draws bounds the one-sided p at 1/4, regardless of how many compositions are decoded. More null graphs, not more compositions.

A trained probe recovering energy from the population code shows that energy is accessible to that probe. It does not show the fixed motor projection exploits it. The two are reported separately.

The objective is a cluster expansion fitted to MACE-MPA-0. It is a physically motivated model, and a search method can preferentially find its errors, because different proposal fields visit different parts of its domain. Finalists from every arm - including the comparable losers - are re-evaluated with a stronger evaluator before any materials claim is made.