# Model evidence and correspondence

Research cutoff: 12 September 2026. This independent adaptation presents contact-driven feeding initiation in a tethered adult female fly. It preserves a published neural graph and makes one experimentally supported sensory-to-motor pathway visible. It does not reconstruct the complete behaving animal.

## Which fly and which reconstruction

The selected source is the **FAFB/FlyWire materialization v630 graph supplied with Shiu et al. (2024)**: 127,400 neurons, 14,687,178 directed connections representing 52,793,639 synaptic contacts. Neuron identifiers, edge direction, counts, and the authors' signed weights are retained. These are counts of the selected model files, not claims about every neuron or synapse in a living fly.

[Zheng et al. (2018)](https://doi.org/10.1016/j.cell.2018.06.019) produced the adult female brain EM volume. [Dorkenwald et al. (2024)](https://doi.org/10.1038/s41586-024-07558-y) published its reconstructed wiring diagram, with 139,255 neurons in the later public reconstruction. [Schlegel et al. (2024)](https://doi.org/10.1038/s41586-024-07686-5) supplied hierarchical annotations and cross-connectome cell typing. The public v783 snapshot dates to October 2023, with additional annotation updates described in [FlyWire's release guidance](https://flywire.ai/guidelines). It is not the v630 graph used here.

Anatomical coverage has since expanded. The [female BANC study, published June 2026](https://doi.org/10.1038/s41586-026-10735-w), connects brain and ventral nerve cord in one specimen. Its analysis emphasizes local sensory-effector loops coordinated by ascending and descending circuits, with higher brain regions supervising them. The [MaleCNS project, published September 2026](https://male-cns.janelia.org/), reconstructs a different, male specimen's entire CNS: [166,691 neurons and 11,691 types](https://research.google/pubs/sexual-dimorphism-in-the-complete-connectome-of-the-drosophila-male-central-nervous-system/). Earlier hemibrain, female VNC, and MANC datasets cover other specimens or partial regions. None is silently grafted onto this model.

The chosen brain volume includes the labellar gustatory inputs, intervening brain circuitry, and actual proboscis motor neurons needed for this experiment. Walking and flight require additional sensory, cord, muscle, and body interfaces. New anatomical datasets do not themselves supply their physiological dynamics.

## What the leading demonstrations establish

| Primary work | Actual computation and control | Consequence for this adaptation |
| --- | --- | --- |
| [Shiu et al., Nature 2024](https://doi.org/10.1038/s41586-024-07763-9) | Untrained, signed-connectome LIF model. Sensory or optogenetic-like Poisson stimulation predicts downstream firing. Feeding and antennal-grooming circuit predictions were compared with experiments. | Use the original graph and experimentally identified taste-to-motor interface. The reported 91% agreement concerns 164 circuit predictions, not a freely behaving fly's repertoire. |
| [Lappalainen et al., Nature 2024](https://doi.org/10.1038/s41586-024-07939-3) | Anatomically constrained visual network with dynamical parameters trained for optic-flow estimation; predicts previously measured visual responses. | Strong sensory-system evidence, but adding this network would introduce task-trained parameters and a separate motor-interface problem. Vision is excluded here. |
| [Wang-Chen et al., NeuroMechFly v2, 2024](https://doi.org/10.1038/s41592-024-02497-y) | MuJoCo body, sensory simulation, CPG/rule/hybrid walking controllers. Multimodal navigation uses RL. Fly-following combines connectome-constrained visual processing, object readout, descending steering, and hybrid locomotion. | A useful embodiment framework; its walking is not supplied directly by the reconstructed brain's motor circuitry. Its body/controllers are not included here. |
| [Vaxenburg et al., Nature 2025](https://doi.org/10.1038/s41586-025-09029-4) | Detailed fly physics with learned locomotion controllers; the [authors' repository](https://github.com/TuragaLab/flybody) supplies imitation environments and distributed RL training. | Biological body form and convincing movement do not establish a connectome-derived neural mechanism. |
| [Eon's technical account, March 2026](https://eon.systems/updates/embodied-brain-emulation) | Connectome model plus sparse, hand-selected descending readouts and imitation-trained body controllers; brain/body synchronization every 15 ms. | The authors explicitly acknowledge approximated brain-body mappings. This demonstration is not evidence that all visible actions follow directly from known synaptic anatomy. |
| [FlyGM, February 2026 preprint](https://arxiv.org/html/2602.17997v1) | Connectome graph policy with learned sensory encoder, neuron descriptors, update function, and motor decoder; imitation from an MLP expert followed by PPO. | Shows a structural prior for learned control. It does not identify the biological dynamics required here. |

[Brunton, Abe, Hu and Tuthill's *Digital Sphinx*](https://elifesciences.org/reviewed-preprints/111516), an eLife reviewed preprint available by August 2026, provides a direct methodological criticism: a worm connectome coupled through a learned decoder can produce realistic fly walking. Their result separates plausible output from meaningful neural reconstruction. It motivates three concrete constraints here: preserve identified sensory and motor interfaces; expose computed activity only; and evaluate causal perturbations, rather than judging fidelity from motion alone.

## Correspondence contract

These constraints define the implementation and its validation targets. They are not a claim that validation has already passed.

| Source property | Browser equivalent | Boundary or required check |
| --- | --- | --- |
| Individually reconstructed neurons | One dynamical state per original neuron and an index-to-FlyWire-ID mapping | No cell-type averaging or invented populations. |
| Directed, weighted chemical connectivity | Full sparse adjacency with original signed counts | No distance-based substitute, pruning, rewiring, or arbitrary steering connections. Recurrent edges remain available. |
| Local activity within a much larger graph | Sparse scheduling of active neurons and delayed events | This is a computational optimization, not a reduced connectome. Dormant neurons must become eligible when input arrives. |
| Inferred transmitter action | Preserve the published model's sign assignment | Signs are model assumptions based on transmitter predictions, not measured receptor-specific effects. |
| Membrane integration and synaptic latency | LIF state, synaptic decay, refractory periods, delayed spike delivery | Test the numerical implementation independently of the body. Analytic interval integration does not imply exact continuous-time spike timing. |
| Labellar taste reception | The source's 21 sugar and 21 bitter GRNs receive input only when their modeled sensory surface contacts the droplet | No odor field, sight, target coordinates, or taste at a distance enters the neural model. Contact-to-firing conversion is an assumption. |
| Proboscis motor output | Read spikes separately from the two source-identified MN9 neurons | Neutral labels MN9 1 and MN9 2 avoid inferring hemisphere from inconsistent source labels. No separate behavior selector or trajectory generator. |
| Rostrum protraction and coupled haustellum movement | Motor activation supplies torque to a damped, spring-return two-joint approximation | Muscle force, spring constants, coupling and joint geometry are phenomenological; they were not measured from this specimen. |
| Body-dependent sensation | Moving joints move the labellum; overlap with the same droplet is recomputed | Body movement can establish or break sensory contact. The droplet does not automatically follow the mouth. |
| Experimental intervention | Add bitter during sugar contact, block modeled taste input, or silence MN9 | Distinguish upstream sensory changes from downstream motor changes. Compare bitter suppression and loss of protraction with the cited experiments. |
| Observable neural causality | Display values read from running sensory, neural, and motor state | No decorative firing, unrelated pulses, or prerecorded traces. |

The source [model code](https://github.com/philshiu/Drosophila_brain_model/blob/main/model.py) specifies rest/reset −52 mV, threshold −45 mV, membrane time constant 20 ms, synaptic decay 5 ms, refractory period 2.2 ms, delay 1.8 ms, and signed connection count multiplied by 0.275 mV. The last quantity is a fitted global parameter. Voltage and synaptic drive reset on spikes; the code freezes both during refractoriness. Sensory drive is Poisson, so repeated trials can differ. Numerical equivalence, repeatability with a fixed seed, and long-run stability require separate tests.

## Why the visible movement is limited

[McKellar et al. (2020)](https://doi.org/10.7554/eLife.54978) mapped proboscis motor neurons to muscles. MN9 activates the rostrum protractor; manipulating it also affects the haustellum through joint coupling. Silencing it preserves the resting posture but prevents rostrum and haustellum extension during feeding. The three-part proboscis folds into the head and extends predominantly ventrally for surface feeding. This supports articulated mouthpart movement rather than a protruding straight tongue. [Gordon and Scott (2009)](https://pmc.ncbi.nlm.nih.gov/articles/PMC2650400/) describe retraction following stimulus removal.

The fly remains tethered; legs and wings are static. The model does not simulate locomotion, flight, grooming, ingestion, digestion, satiation, learning, vision, olfaction, or spontaneous exploration. Shiu's framework assumes zero basal firing and identical spiking units; it omits morphology, nonspiking neurons, electrical synapses, receptor-specific dynamics, plasticity, physiological internal state, and extrasynaptic neuromodulation. Absolute firing rates and the MN9-to-extension relationship are not established biological measurements. Preserving the full source topology does not make this a biologically complete whole-brain emulation.

## Sources and reuse

The medium follows [Bret Victor's concrete, manipulable simulation practice](https://worrydream.com/SimulationAsAPracticalTool/): a physical stimulus acts on the visible specimen, and a small adjoining readout exposes the causal path. This is a design inference, not biological evidence. [Wang and Brown's NN/g evaluation of AI-generated interfaces](https://www.nngroup.com/articles/ai-prototyping/), updated August 2026, identifies generic component-library layouts and misplaced information hierarchy. Accordingly, the specimen occupies the principal view; no overview cards, decorative network or achievement system was added. Scientific labels and numbers are tied to actual state. The warm cuticle, dark abdominal margins, red compound eye, translucent venated wings, six legs, antenna and folding mouthparts follow the lateral fly and anatomy in McKellar's Figure 1. They are original simplified drawing geometry, not a traced or measured reconstruction.

[MDN's Canvas guidance](https://developer.mozilla.org/en-US/docs/Web/API/Canvas_API/Tutorial/Optimizing_canvas) supports requestAnimationFrame scheduling and explicit display-resolution handling. The runtime uses a Canvas drawing and sparse typed arrays in a worker because measurements show these suffice on the target M3. Introducing a GPU solver, general physics engine or frontend framework would add implementation burden without a demonstrated requirement. Raster marks map to recorded spike times; no reconstructed 3D neural locations are implied. The main view omits sound because no modeled acoustic or biomechanical sound source is present.

The authors' [Edmond archive for these exact v630 model input files](https://doi.org/10.17617/3.CZODIW) records an MIT license. [FlyWire's general public-data guidelines](https://flywire.ai/guidelines) specify CC BY-NC 4.0. This package retains both notices and treats underlying connectome reuse as noncommercial; the archive notice is not used to override upstream terms. Repacking and indexing do not erase attribution or restrictions. The [Shiu reference code is MIT licensed](https://github.com/philshiu/Drosophila_brain_model/blob/main/LICENSE), separately from the anatomical data's provenance. McKellar's article is CC BY; NeuroMechFly's Apache-2.0 software license does not license its journal figures. Comparative papers and models are cited, not bundled. Original application work is separate from third-party material and does not imply affiliation with or endorsement by any source institution.