What this model cannot do
The explorer runs a real connectome through the simplest neuron model that reproduces the fly's core reflexes. Here is what the field says such a model is missing, in its own words — and, for each, the flybench task that measures the gap and what a better model would need. None of this is a bug. It is the reason the benchmark exists.
1.A wiring diagram is not a program
Knowing every synapse of the C. elegans nervous system for decades did not by itself yield its function: the same anatomical circuit produces different behaviours depending on neuromodulatory state, and the connectome constrains but does not determine what a network computes.
Bargmann & Marder 2013, Nat Methods 10:483 ("From the connectome to brain function")
Measured: Every hard-tier task is a version of this question. The reference model reproduces the five core reflexes and fails most of the rest; the shuffled-wiring control says how much of each pass is the wiring at all.
Needs: Nothing fixes it in general. Each task names one missing ingredient.
2.Neuromodulation is missing, and it is not a detail
Neuromodulators reconfigure circuits: the same anatomical network can produce several different outputs depending on which modulators are present, so a connectome without its modulatory state is a set of possible circuits, not one.
Marder 2012, Neuron 76:1; Bargmann 2012, BioEssays 34:458
Measured: Dopamine, octopamine and serotonin neurons are in the graph as ordinary excitatory cells (the model's sign convention). Hunger, arousal and mating state — which gate feeding, halting and egg laying in the fly — have no representation. — halt_walk_off, egg_laying_ovidn
Needs: A modulatory state with literature constants, applied to a documented target set. Nobody has one for the whole fly.
3.Gap junctions are not in the diagram
The FlyWire and MaleCNS connectomes annotate chemical synapses; electrical synapses are largely invisible to the EM pipeline. The giant fiber's fastest output — to the jump motor neuron — is electrical.
Dorkenwald et al. 2024, Nature 634:124 (methods); Allen et al. 2006, Semin Cell Dev Biol 17:31 (the GF–TTMn gap junction)
Measured: Task 19 asks for the measured ≈ 0.6 ms giant-fiber-to-jump-muscle latency. A model with one chemical delay per synapse cannot get under ~1.8 ms. It fails, and is meant to: the task names a missing mechanism, not a wrong parameter. — gf_to_muscle_latency
Needs: An electrical-synapse table from an independent source (dye coupling, innexin expression) and a second edge type in the model.
4.Every neuron is the same point
Fly neurons are electrotonically extended; the location of a synapse on the arbor matters, and dendritic integration in Drosophila central neurons is not a sum of inputs at a point.
Gouwens & Wilson 2009, J Neurosci 29:6239; Lesser, Azevedo et al. 2024, Nature 631:369 (motor neuron size and input)
Measured: Task 23 ramps a leg's premotor pool and asks for the size principle. The wiring gives large motor neurons proportionally more synapses, so a uniform point model recruits them first — the opposite of the fly. The paper's own inference is that intrinsic properties must compensate; the model has none. — leg_mn_size_principle
Needs: Per-type intrinsic parameters (input resistance, threshold) measured, not fitted — and a rule for which types get them.
5.Synapse count is not synaptic strength
The number of synaptic contacts between two neurons correlates with functional strength but does not determine it; transmitter identity itself is a prediction from EM images with a stated error rate.
Shiu et al. 2024, Nature 634:210 (the weight convention); Eckstein et al. 2024, Cell 187:2574 (transmitter prediction)
Measured: Two tasks turned on a single transmitter label: LN23 is glutamatergic in MaleCNS and unlabelled (so excitatory) in FlyWire; ExR6 the reverse. The same circuit fails in opposite directions on the two brains. — co2_pathway_specificity, epg_ring_attractor
Needs: A documented transmitter-override list as a dataset correction, with its evidence, and physiology for the handful of edges a task hinges on.
6.One animal, and animals vary
Neurons and circuits reach the same function with widely different parameters from animal to animal; a single specimen's numbers are one solution among many.
Marder & Goaillard 2006, Nat Rev Neurosci 7:563; Schlegel et al. 2024, Nature 634:139 (hemibrain vs FlyWire)
Measured: Task 14 reruns the core reflexes with every synapse count jittered ±25 %. At the working gain, escape survives and taste is a coin flip; there is no gain at which both are robust and the brain stays sparse. — wiring_robustness
Needs: Homeostatic rules with measured targets — or a second connectome of the same sex to bound the variation directly.
7.It has no eyes, no nose, no body
Behaviour is closed-loop: the fly's own movement changes its sensory input within tens of milliseconds, and a brain scored open-loop on hand-made stimuli is being asked a different question than the animal answers.
Lobato-Rios et al. 2022, Nat Methods 19:620 (NeuroMechFly); the digital-sphinx result, bioRxiv 2026.03.20.713233
Measured: Every stimulus here is "the neurons this stimulus would drive, at 100 Hz": a convention the ratio checks cancel and the rate checks do not. Optic flow is HS and H2 driven directly; an odour is one whole glomerulus. — optic_flow_rotation, da1_sparseness
Needs: A sensory front end (flyvis for vision) and, eventually, the closed-loop track — with the shuffled-wiring control run alongside, because a body controller can produce realistic behaviour from a worm's connectome.
8.Nothing switches off
Real neurons adapt, synapses depress, and circuits return to rest within a second of a stimulus; a network of memoryless point neurons with recurrent excitation has no state that outlives its membrane time constant.
Benda & Herz 2003, Neural Comput 15:2523 (adaptation); Shiu et al. 2024 (the reference model has none)
Measured: After a half-second taste of sugar, ~8 % of the brain fires at a constant rate forever. Adding spike-frequency adaptation with two literature constants fixes it and moves the hard tier by +0.15 — and costs the tasks that needed a sustained response. — return_to_rest
Needs: The mechanism exists; what is missing is the per-type constants and the second mechanism (synaptic depression) that the compass and egg-laying tasks say adaptation alone gets wrong.
Every wrong prediction the benchmark has made is in docs/FINDINGS.md; every task's pre-registered prediction and outcome in docs/rfcs. If a critique here misquotes its source, open an issue.