The fly's goal circuit normalizes; it does not choose

What the work claims

This is a connectome-driven circuit analysis, a structural inference from a fully reconstructed wiring diagram, disciplined about which conclusions the wiring earns and which it only offers. Nanni and Lee start from a published functional observation: FC2 neurons carry the fly's goal direction as an activity bump, and they inhibit one another over distance, which has been proposed to enforce a single goal. They ask what circuit actually produces that inhibition, and whether it lets FC2 actively choose one goal among competitors, a winner-take-all, or merely keeps a goal set elsewhere as one clean bump.¹

The central claims are five. First, tracing every route in the FlyWire whole-brain connectome, the feedback inhibition onto FC2 decomposes into three parts: a uniform component, delivered by the four FB5A tangential cells, which reach every one of the 85 FC2 neurons; a smaller anti-local component from hDelta interneurons, which grows with bearing distance and so slightly suppresses flanking activity; and a negligible direct FC2-to-FC2 component. Second, a winner-take-all built as a ring attractor requires local recurrent excitation, and the FC2 wiring contains none, so the geometry alone cannot build one. Third, across a family of dynamical models parameterized by the real connectome, including a committed leaky integrate-and-fire network of all 85 FC2 neurons, no version of the circuit locks onto a winner at the connectome-scaled coupling. Fourth, the same structural facts replicate in the independently reconstructed hemibrain connectome: FB5A reaches all 88 FC2 cells there, the disynaptic loop is flat to 1.2 percent modulation, and the hDelta anti-local far-to-near ratio of 2.43 closely matches FlyWire's 2.50. Fifth, the connectome nominates where the goal is actually set: an upstream hDelta C-led recurrent network, while ruling out the leading published alternative, an hDelta K-PFG attractor whose neurons supply under 0.2 percent of FC2's input.¹

The authors are unusually explicit about what remains open. FB5A is predicted to be GABAergic, inhibitory, by the connectome's transmitter classifier at about 0.79 confidence, a value the paper calls a low-confidence prediction, notes is experimentally unverified, and argues is likely wrong, since FB5A is probably not GABAergic at all. And a different selection mechanism, mutual inhibition between two competing goals supplied by the hDelta route, could in principle select at very strong coupling; the authors bound this rather than exclude it.¹

How it works

The distinction being tested is between two canonical circuit computations. In a winner-take-all, competition makes one candidate win and suppresses the rest; in divisive normalization, a global inhibitory signal proportional to total population activity rescales every neuron, sharpening a pattern without deciding its content. A ring-attractor winner-take-all, the architecture of the fly's heading compass, additionally needs local recurrent excitation that amplifies the winning location, and that ingredient is precisely what the FC2 wiring lacks: direct FC2-to-FC2 connection is negligible, so all effective inhibition arrives through the FB5A pool and the hDelta interneurons.¹

The measurements are clean. If FB5A sculpted the bump spatially, its per-cell weight onto FC2 neurons should vary with bearing distance from the bump; it does not. The four FB5A cells reach all 85 of 85 FC2 neurons, the per-cell weight correlates with preferred bearing at essentially zero (plus 0.02 and minus 0.04 against the two bearing components), and the disynaptic FC2-to-FB5A-to-FC2 loop is flat across bearing distance with about 5 percent modulation. FB5A supplies roughly 95 percent of the total inhibitory synapse mass, uniformly; the hDelta route contributes a modest anti-local rise, far-to-near ratio 2.50 in FlyWire, the only distance-dependent term. FB5A, in short, scales the whole population without sculpting it, the same computational motif as the APL neuron, the global normalizer of the fly mushroom body.¹

Because wiring alone cannot settle dynamics, the authors test whether any plausible dynamical model of the real circuit nonetheless latches a winner. The bistability test is operational: drive the model with two competing goals, seed the state toward each candidate, and ask whether the outcome depends on the seed. Across rate models, a spiking leaky integrate-and-fire network built from the actual 85-cell connectivity, divisive normalization with swept sharpening exponent, and the hDelta mutual-inhibition matrix, no model is bistable at the reference coupling, so there is no within-FC2 winner-take-all; feedforward max-selection is excluded as well. A two-cue competition test with goals 120 degrees apart at a 1-to-0.7 amplitude ratio reproduces the behavioral signature reported from FC2 imaging. The authors close by proposing the experiment that would settle the account: silence FB5A while imaging FC2, predicting the goal bump should broaden rather than multiply.

Where a skeptic should push

The most load-bearing assumption is that synaptic connection counts, scaled to a reference coupling, are an adequate proxy for functional inhibition. The authors flag this themselves: synapse count is a structural proxy for functional inhibitory strength, and the proxy is weakest precisely where the conclusion depends on it, because the inhibitory identity of FB5A rests on a transmitter-classifier prediction at 0.79 confidence that nobody has verified physiologically and that the authors argue is likely incorrect. The structural claims, uniformity, flatness, absence of local excitation, replication across two connectomes, are robust and transmitter-independent; the functional interpretation, that FB5A acts divisively to normalize, is offered as a falsifiable candidate, and the paper's own earned-versus-offered framing should be preserved by any reader quoting it.¹

Second, the no-winner-take-all result is a bound, not a proof: it holds at connectome-scaled coupling, and the hDelta mutual-inhibition route could select between two competing goals at very strong coupling, which the authors bound rather than exclude. Whether biological operating points ever reach that regime is an empirical question the connectome cannot answer. Third, the nomination of the upstream hDelta C-led network as the goal-setter is tracing-based inference, with no functional evidence yet; ruling out the hDelta K-PFG alternative at under 0.2 percent of input is decisive for that candidate but does not positively demonstrate the nominated source. The proposed silencing experiment is cheap, direct, and has not been done.

Organoids need gain control, not winner-take-all

For organoid intelligence, the non-obvious implication cuts against the field's default mental model. Anyone who wants to steer living neural tissue eventually wants what the fly wants: hold one goal state, cleanly, against competing inputs. The attractor literature makes that sound like a topology problem, sculpt recurrent excitation and local inhibition into a ring or bump circuit, and disordered organoid tissue cannot be sculpted that way. The fly's answer is that the circuit holding the goal does not solve the topology problem at all. FC2 keeps its bump clean with one of the dumbest mechanisms available, a global gain control from four cells that inhibit everything equally, while whatever selection happens happens elsewhere. Division of labor, not a clever local circuit.¹

The opportunity follows directly: global normalization is the one steering primitive external hardware can emulate cheaply on tissue that cannot be patterned. A closed loop that reads total population activity from a multielectrode array and feeds back stimulation in proportion to it implements divisive normalization without knowing anything about the culture's wiring. On this paper's account, that primitive, plus a tonic drive representing the goal, may suffice to hold a single mode even in disordered networks, because a real brain, with all its evolutionary tuning, uses the same trick. It reframes the engineering goal from building structure into the tissue to supplying the two signals the tissue is missing.

Two cautions come attached. First, a benchmarking trap: a normalizer preserves whatever it is driven with, so apparent state-holding in a stimulated organoid can be content-free. If a culture's dynamics behave like FC2, the tissue is displaying a goal the stimulator set, not computing one; experiments that only read the bump cannot distinguish the two, and only perturbation of the drive, the equivalent of silencing FB5A, can. Second, governance: a substrate that holds whatever goal it is given concentrates control in whoever sets the drive; normalization makes living tissue obedient, which is precisely the property that makes biological computing useful and the property that makes its dual-use and ethics questions non-optional. And there is a quieter epistemic lesson worth importing: this paper models how to argue mechanism from structure, bounding the competing hypothesis instead of excluding it and publishing the confidence of the classifier your conclusion leans on. Organoid circuit claims are routinely asserted from far weaker proxies.

The bottom line

Established, and replicated across two independently reconstructed connectomes: the feedback inhibition onto the fly's goal-holding FC2 population is dominated by a spatially uniform FB5A component reaching every FC2 cell, the FC2 ring lacks the local recurrent excitation a winner-take-all requires, and no connectome-parameterized dynamical model, spiking included, selects a winner at reference coupling; the leading alternative goal-source is ruled out at under 0.2 percent of FC2 input. Not established: that FB5A is inhibitory (a 0.79-confidence transmitter prediction, unverified, possibly wrong), that the nominated upstream hDelta C network sets the goal, or that strong-coupling mutual inhibition could never select. Confirmation is the authors' own proposed experiment, silencing FB5A while imaging FC2, which should broaden the goal bump. For organoid intelligence the durable contribution is a reframing: if you want one clean state held in living tissue, the evidence from a real circuit says build the normalizer, not the selector, and then take seriously the question of who is holding the drive.