The lie is the verb.
Its own public code shows a human-authored rig supplied the token details, pair, tax, launch clicks, confirmation, and wallet signature. The connectome moved a cursor inside that ceremony. FLYTOWN tested the deeper computational claim—and the measured wiring did not beat its controls.
Scope: the launch story overstates agency, and our tested computational advantage was null. Neither finding says the published connectome or the external project’s telemetry is fabricated.
A real token exists. A real connectome ran. Neither fact makes the launch autonomous.
Audit the causal chain$FLYBRAIN is an on-chain token. Its official pitch says it was “launched by a fruit fly connectome.” But the project’s public source records a different division of labor: the connectome produced cursor signals while conventional automation supplied the goal, text, market settings, irreversible clicks, and signature.
0x4eb990547bce4a982432ca88cf5fae7eed1a2d35
Official claim
The connection graph comes from microscopy. The running “brain” does not. The code gives every neuron identical leaky-integrate-and-fire parameters, turns synapse counts into one fixed voltage multiplier, and leaves a trainable gain for each cell type. Those are consequential engineering choices—not measurements of the living fly.
Read the modelThe connectome emitted motion and a stop signal from four selected descending-neuron readouts. A script inspected the page’s DOM, chose the nearest unfinished field, and typed a prewritten token value when the stop signal landed inside its box. The fly did not invent the name, ticker, description, target field, or objective.
Inspect the loopAutomation accepted the terms, uploaded the image, filled anything the fly missed, selected GOOGL, set the creator tax, moved to the launch control, pressed Launch, and pressed Confirm. The project’s own README admits the fly did not complete the form or choose the pair, tax, X handle, or confirmation.
Read its admissionA Python-injected wallet signed and broadcast the transaction after scripted confirmation. The receipt proves that a transaction succeeded. It does not prove that the simulated fly understood a token, selected an economic action, consented to terms, or controlled the irreversible step.
The lie is the word “launched.” A connectome-derived cursor participated in a launch orchestrated by human-written software. Every choice that gave the token meaning—and every irreversible action that made it real—came from outside the connectome.
The public repository now discloses many of these limits. That candor is useful. It does not make the headline causal claim true.
FLYTOWN did not reproduce their browser rig, so we do not claim their neuron telemetry is fake. We tested a separate, broader inference: whether measured fly wiring added useful task computation. It did not beat shuffled or rewired controls in our implemented planning and memory roles. That is a result about these roles, not a universal claim about biology. Real anatomy is not, by itself, evidence of agency or advantage.
Use the wiring diagram of a real fruit-fly nervous system as the fixed middle of an engineered planner. A task goes in; activity moves through the graph; action scores become a plan for AI workers.
The hypothesis lives here: if biological wiring contributes useful computation, the measured graph should beat an appropriately scrambled version when everything else stays fixed.
These diagrams come from the connectome artifacts used by FLYTOWN—not from stock network art. They establish that the test used measured biological wiring. They do not establish that the wiring helped the engineered task.
Every region is present. The diagram surfaces the 110 strongest cross-region edges so the overall routing structure stays readable.
All 2,952 neurons remain in the counts. For this view, they are aggregated into 14 annotated classes and the 54 strongest cross-class flows are shown.
Measured Node identities, counts, edges, directions, and synapse weights come from the checked-in datasets.
Engineering choice Circular order, aggregation, edge filtering, scale, and color are explanatory design choices.
Self-connections are omitted from these two views. The full, unfiltered artifacts remain in the repository. Inspect sources and citations
A strange result is only persuasive if the comparison is ordinary. The strongest tests altered the graph while keeping the task, seed, budget, worker setup, and action-to-plan compiler matched.
Edges from the published connectome artifact.
Labels scrambled or edges degree-preserved.
If the real graph does not beat this null, the wiring has not shown that it is doing the claimed job.
The implementation improved between attempts. The evidential standard did not: measured wiring still had to beat a relevant scrambled control.
Null result
p = 0.060 — not distinguishable under the declared p < 0.05 threshold.
The measured graph repeatedly fell into the same anatomical attractors. Those patterns were interpretable, but they were not task-dependent routing.
Preregistered null
p = 1.000 — the primary comparison was exactly null.
Every Kenyon cell was active on every task. With a dense code, the same synapses were eligible for depression each time; association-specific learning was impossible in this configuration.
Post-hoc null
p = 1.000 — the primary test was null again.
Two ablations moved in a biology-consistent direction at uncorrected p = 0.014 and 0.017, but neither survived correction for nine tests and both involved a constant-policy comparator. Suggestive is not evidence.
Different role, same null
0.594–0.6490.680–0.740The real circuit separated categories worse at every setting and held-out valence stayed near chance. It did achieve 100% collision-free codes, pointing toward a narrower novelty-detection hypothesis—not evidence for similarity retrieval.
The early runs used fixture suite v1, including private repositories. The public, pinned v2 suite exists, but the log records no v2 result yet. This limits reproducibility of those numbers and is part of the finding, not a footnote to hide.
The adult projectome did not merely underperform. A diagnostic measurement showed the mechanism: tasks became less distinguishable as activity moved through the real graph.
The real region-level graph reduced task information about eight times more than the shuffled copy at the graph-output stage.
Failure becomes research when it reduces uncertainty, exposes mechanism, and changes what gets built next.
Repeated nulls make continued tuning against the same suite hard to justify.
Task washout, dense eligibility, compressed readouts, and compiler-inert actions are inspectable failure mechanisms.
Constant-policy warnings, task-sensitivity measures, action-effect labels, public fixtures, and provider preflight came from failed runs.
Novelty detection and a larger adult mushroom body are specific, preregisterable ideas—not another vague promise to “use a connectome.”
The full live comparison found that a hand-written rules planner matched the LLM planner’s quality on completable tasks while using much less computation—and knew when not to start work.
p < 0.0001
p = 0.015p = 0.89 · dead even20 fixtures × 2 seedsThe LLM planner always tried to spawn work. Rules could express “blocked,” “needs approval,” and “already done.” FLYTOWN now uses rules as its default planner.
The strongest argument is not that the study was perfect. It is that the project records the places where it was not.
FLYTOWN is MIT-licensed code with documented connectome provenance, experiment reports, null controls, and a local inspection console.
git clone https://github.com/water-bear86/flytown.git