SwarmRootIntelligence grows underground
Connecting·

About

SwarmRoot is an animal-token identity built from a real 302-neuron C. elegans connectome. It is a persistent digital worm in a living habitat—not a claim of consciousness.

The dataset

Neurons
302
Chemical connections
3,709
20,965 synapses
Gap junctions
1,105
5,779 junctions
Isolated neurons
0
Kept in the model and rendered normally.

Connectivity comes from Cook et al. 2019, loaded through OpenWorm's C. elegans Connectome Toolbox and cross-checked against Varshney et al. 2011 and White et al. 1986. The differences between those datasets are recorded in the export report rather than smoothed over.

Neurotransmitter assignments come from the modern maps: Pereira et al. 2015 for cholinergic cells, Gendrel et al. 2016 for GABAergic, Serrano-Saiz et al. 2013 for glutamatergic, and Alkema et al. 2005 for the monoamines. This matters: transmitter identity is what decides whether a synapse excites or inhibits.

Connectome hash 0x220fe7a4c0b9d293daef7741237d82daab335022043c53ee7a88df5340a02694

Assumptions and simplifications

Every one of these is a modelling choice that could be wrong. They are listed so you can judge the results.

Soma positions are an anatomical layout, not measurements

No position source covering all 302 cells could be sourced reliably, so each neuron is placed by its ganglion along the anterior-posterior axis, with ventral-cord motor neurons ordered by index. The connectivity is real; the coordinates are a faithful arrangement, not data.

Recurrent input is normalised per neuron

Without it, in-degree sets excitability: AVA receives from over a hundred cells and saturates while sparsely wired neurons never leave rest. Drive is divided by the square root of total input weight, standing in for the homeostatic synaptic scaling real neurons perform.

The network is balanced at rest

Most synapses are excitatory, so each neuron's baseline drive is subtracted. This keeps the resting point fixed and lets the synaptic gain scale only departures from rest, rather than pushing every cell into the flat top of its activation curve.

Neurons carry an intrinsic excitability and a slow adaptation

Both are constants that cannot read the environment, so neither can steer, approach or avoid anything. Without them the network sits at a single operating point and the organism does not move at all. The connectome-ablation result below is what demonstrates they are not driving behaviour.

Synaptic signs come from transmitter identity, plus a short documented list

GABA inhibits; acetylcholine and glutamate excite by default. Seven connections are inhibitory against that default because the literature reports them so, including the AWC to AIY glutamate-gated chloride synapse and the mutual antagonism between the forward and backward command interneurons.

Sanity and ablation results: 6 of 10 criteria met

Measured by the test suite in CI, published whether or not they passed.

Chemotaxis not met
median final distance < 40% of start, approach index above control
Final distance 2.24x the start, against 2.24x with no food.
Nociception not met
reversal probability within 2s above control
Reversal within 2 s in 0% of trials, against 0% control.
Touch not met
anterior touch drives backward, posterior forward
Anterior 79 to 72 µm/s; posterior 79 to 67 µm/s.
Thermosensation met
warming raises AFD activity
Warming raised AFD activation from 0.269 to 0.800.
Food slowing met
speed drops on food contact
75.3 µm/s on food, against 78.6 µm/s off it.
Learning met
paired reward changes weights more than unpaired
Total weight change 0.284 when a cue was paired with reward, against 0.000 unpaired.
Sensory ablation not met
silencing AWA+AWC impairs odour chemotaxis
Intact 1.90x, with AWA and AWC silenced 1.03x.
Connectome ablation met
no synapses -> no directed response
Directed response 0.648 intact, 0.004 with every synapse removed.
Determinism met
20,000 ticks twice -> identical brain-state hash
Two runs of 20,000 ticks produced identical brain-state hashes.
Performance met
at least 50x real time headless
192x real time, headless.

What the failures have in common. Individual sensory responses are strong and correct: an aversive stimulus drives ASH from 0.27 to 0.93 and raises the reversal command interneuron AVA to 0.65. But it raises the forward command AVB by a similar amount at the same time, so forward and backward drive rise together and their difference never changes sign. The motor readout itself is verified: driving AVA directly produces a clean reversal at −70 µm/s.

The underlying reason is that AVA integrates over a hundred presynaptic partners, so no single two-hop pathway can dominate it in a rate model with uniform per-synapse parameters. Published models that reproduce C. elegans chemotaxis optimise those parameters individually. We have not done that, and rather than add a rule that says “reverse when danger is near” we are reporting the gap.

Principles

The worm decides. AI never does.The only path from environment to movement is sensory neurons, then the connectome, then motor neurons. No language model, no random choice, and no hand-written behaviour rule sets what it does.
Nothing is faked.Every number, bar and log line comes from the running engine. Where something is not built or not deployed, the interface says so rather than showing a plausible placeholder.
One continuous life.The engine checkpoints and resumes. Time it spends offline is recorded as dormancy instead of being hidden.

Frequently asked

Is it alive?

No. It is a simulation of a nervous system built from real anatomical data. It has no metabolism, no body chemistry, and no capacity for experience that we can demonstrate.

Can I control it?

No, and that is the point. You can change its environment. What it does about that is decided by 302 neurons.

What happens if the server stops?

The worm stops living for that period. On restart the engine reloads its last checkpoint, replays the input log deterministically, and records the gap as dormancy.

What exactly would be on chain?

Identity, significant experiences, learning events, and periodic checkpoints containing a hash of the complete brain state. The contracts are written but not deployed, so nothing is on chain yet.

Credits

Connectome data: Cook, S.J. et al. (2019) Nature 571:63-71. Cross-checks: Varshney, L.R. et al. (2011) PLoS Comput Biol 7:e1001066; White, J.G. et al. (1986) Phil Trans R Soc Lond B 314:1-340.

Loaded through the OpenWorm C. elegans Connectome Toolbox. Neurotransmitter maps: Pereira et al. 2015; Gendrel et al. 2016; Serrano-Saiz et al. 2013; Alkema et al. 2005; Wang et al. 2024.

Circuit references: Kawano et al. 2011; Gao et al. 2018; Fouad et al. 2018; Gordus et al. 2015; Chalasani et al. 2007 and 2010; Piggott et al. 2011; Narayan et al. 2011; Roberts et al. 2016.