Standard nulls create a smell-to-motor shortcut
A connectome null model is a randomised copy of the wiring used as a baseline: if evolved agents built on the real wiring outperform agents built on the copy, the measured wiring is credited with mattering. The paper's first move is to measure what two common randomisations actually do to a compressed adult Drosophila connectome (FlyWire v783, 512 cell-type groups plus 1,000 Kenyon cells).
Both the column shuffle (N1) and degree-preserving edge swaps (N2) route 10.6 to 10.7% of olfactory output directly onto descending motor groups. In the connectome itself that share is 0.012%, and 52 of 53 descending groups sit two or more synapses from olfactory input. The median path from olfactory receptors to descending groups falls from three synapses to one. The author notes this is not a bug in the nulls; it is what randomising the boundary does.
A direct route changes what the comparison measures
The agents forage in a unit square with 24 odour-marked patches, an energy budget that forces foraging, and four ecologies crossing a stationary or pursuing predator with all-nutritious or all-toxic patches. Each wiring condition evolves as its own island of 512 agents for 600 generations, with ten seeds per condition. Because a short reactive route from smell to movement is useful in this task, a null that hands the control agents that route is no longer a neutral baseline for the connectome's own wiring.
This is the paper's framing claim rather than a separate measurement: the verdict of a null comparison is a statement about the properties the randomisation changed. The experiments below test that claim by holding the boundary fixed instead.
The connectome loses to shortcut-carrying nulls
Against N1 and N2, the connectome is behind at the last common-garden probe, a fixed evaluation of evolved agents in a shared set of worlds. On seed means the deficits are −0.22 and −0.20 fitness units, with p_holm = 0.029 for N1 and 0.016 for N2. The registered primary endpoint, fitness averaged over the whole run, detected no difference: seed-mean medians were −0.03 for N1 and −0.02 for N2, and the paper reports this as no detected difference rather than as equivalence.
The strategy measures point the same way. Blocking olfaction costs the connectome 0.03 fitness units, against 0.68 for N1 and 0.64 for N2 (p_holm = 0.008), so the controls lean on smell far more heavily than the connectome does. All quantities are dimensionless simulation units.
Boundary-preserving nulls close the gap
The corrected grid replaces the third standard null with two boundary-preserving controls, N4 and N5. These apply the same column-shuffle or degree-preserving operations to interior edges only, where the source is not a sensory group and the target is not a descending group, and the engine aborts unless every sensory-output and motor-input edge is bit-identical to the connectome. About nine in ten interior edges are replaced.
On seed means the connectome-minus-control difference is +0.002 for N4 and −0.074 for N5, both inside the pre-registered ±0.10 equivalence bound. A sensitivity analysis shows the N4 verdict holds for any bound down to 0.045, while the N5 verdict holds at 0.10 but not at 0.075, so the paper calls it marginal. A re-run under a calibration that also matches the spread of activity returned the same verdicts. The olfaction-blocking cost differences against N4 and N5 (−0.07 and −0.11) are not significant after correction and are five- to ninefold smaller than the gaps against N1 and N2.
One ecology's lead did not replicate
The equivalence claim is made on seed means across four ecologies, not per ecology. In the safe-food ecology N5 was ahead of the connectome by 0.26 at ten seeds (p_holm = 0.029). Ten further pre-registered seeds in the two all-nutritious ecologies did not reproduce it: seeds 10 to 19 gave +0.031 with a 90% interval of −0.068 to +0.152, and all 20 seeds gave −0.104 (−0.208 to +0.016, p_holm = 0.27).
At 20 seeds the per-ecology intervals for N4 and N5 lie inside ±0.10 in one ecology (predator plus nutritious patches) but not in the other, and N1 stays ahead in both. The paper is explicit that per-ecology heterogeneity is shown to make it visible, not as a claim of equivalence.
Moving the shortcut into the connectome raises fitness
The intervention experiments test the mechanism directly. Adding the shortcut to the connectome by degree-preserving target swaps, until its olfactory shortcut share reaches N2's 0.106, raised fitness by 0.44 in 10 of 10 seeds and raised the fitness cost of blocking olfaction from 0.15 to 0.99. A graded dose series raising the share to at least 1, 3, 5 and 10% raised fitness in step, from 2.37 to 2.93 on seed means over both ecologies, with a positive per-seed rank correlation in 9 of 10 seeds. At generation 0 the doses are indistinguishable, so the trend is a product of evolution rather than of the starting wiring.
The swap-matched comparison separates amount of rewiring from what the rewiring creates. The full dose took 68 to 129 swaps (mean 98.5) against a sham's fixed 101, yet was ahead of the sham by 0.534 on seed means (10 of 10 seeds). In the seven seeds where the full dose used fewer swaps than the sham it was still ahead in all seven. A boundary-targeted sham that rewired the same boundary edges without creating shortcuts matched the connectome (+0.007, inside the ±0.10 reference), while the full dose was ahead of it by 0.600.
What the sham result does not settle
The registered prediction that the original sham would match the connectome failed: sham minus connectome was +0.034 with a 90% interval of −0.062 to +0.123, crossing the bound, and the two ecologies disagreed in sign. The paper therefore reports that some small part of the advantage follows from degree-preserving rewiring in general, next to the 0.534 attributable to the shortcut at the same swap count, and says it cannot tell from these runs what that part consists of. An unregistered analysis of final-generation in-run fitness does meet the equivalence criterion, and the paper reports it while noting it does not govern the registered test.
A separate check supports that the interior still matters: zeroing interior weights costs 18 to 23% of fitness and randomly permuting them costs 46 to 55%, in every condition, with no detected difference between the connectome and any control. The agents did not solve the task on the boundary alone.
What to check before trusting a null comparison
The practical priority this paper establishes is a diagnostic, not a verdict about the fly brain. Before reading a connectome-versus-null fitness gap as evidence about measured wiring, check the null's sensory-to-motor path statistics alongside the degree statistics it is said to preserve. Here the shortcut share separated the outcomes: 10.6 to 10.7% for the standard nulls against 0.012% in the connectome, and roughly 0.01% for the boundary-preserving nulls.
The paper's own boundary is that these are evolved embodied agents with a compressed connectome in a foraging task, and the equivalence verdict is a seed-mean statement across four ecologies at ten seeds, with the N5 result marginal at the 0.10 bound. Within that scope, the comparison that matters most is the one at matched swap counts: rewiring that creates a sensory-to-motor shortcut was ahead of rewiring that does not by 0.53 to 0.60, while the connectome and the no-shortcut sham were indistinguishable.
Sources used in this answer
Null-model treatment of the sensory-motor boundary changes an evolutionary connectome comparison
Standard connectome nulls route 10.6–10.7% of olfactory output directly to motor groups versus 0.012% in the connectome; boundary-preserving nulls put the connectome's deficit inside a ±0.10 equivalence bound.