A moving fly is an output to explain. The useful result is the relationship between input, neural activity, an intervention, and the assumptions that turn activity into movement.
Start with the frontal target, intact circuit, visual input on, motor output connected, and lift strength 1.0×. Complete or export that run before changing a setting. Then change one factor and repeat.
Record the question as well as the result. For example: “Does wing-power activity account for forward powered travel in this adapter?” is a testable question. “Does the fly look realistic?” does not identify which model component should be checked.
The table below is generated from the project’s recorded arena-validation file. Values belong to the default 12-model-second protocol and use arbitrary arena units.
| Condition | Maximum height above floor (au) | Final horizontal displacement (au) | First threshold launch (model s) |
|---|
Displacement is the straight-line distance from the starting point to the final horizontal position, not the total length of the trajectory. “None” means the assumed jump threshold did not trigger a grounded launch; it does not mean the body never left the floor.
Download the recorded settings, events, and full reference summaries.
DLM/DVM rates are clamped to zero, so the power transformation returns zero. TTMn can still exceed its launch threshold and give a short jump. This separates two modeled output functions: a launch impulse and powered lift/thrust.
The graph continues calculating activity, including motor readouts. The body adapter receives no powered commands. This is a useful check that the target’s animation is not directly pushing the body.
The resting model has no baseline spontaneous drive. Closing vision supplies no external input. A stationary target also produces no expansion for the resting body. Both conditions therefore remain silent and grounded under these equations.
The early threshold launch disappears in the recorded default run, but the body can still rise later through the wing-power adapter. This does not establish an experimentally verified alternative pathway. It tells us to inspect the remaining network routes and the rule allowing lift without a jump.
Both b2 rates become zero, so their right-minus-left difference and the adapter’s yaw command vanish. Other wing motor neurons and their mechanics are not substituted as steering outputs in this simplified adapter.
Run the motor-disconnection control and watch the neural readouts beside the resting body.
Open the disconnected setupSee all suggested comparisonsTwo arena runs begin with the same target motion and fly pose. After an intervention changes the fly’s movement, each fly experiences a different relative target position and velocity. Its later sensory drive can therefore differ.
A difference in a late neural trace can include both a direct network effect and an indirect feedback effect through movement. The recorded experiment does not automatically separate them. If your question requires the same input in both runs, begin with the fixed-input circuit experiment.
This is also why a greater maximum height need not produce greater horizontal displacement. In the reference runs, increasing the assumed lift strength changes the trajectory and subsequent sensory drive. The body and input histories evolve together.
The original 414-unit looming run and the full typed-CNS run use the same rate equation and prescribed input. They nevertheless produce different responses because the larger graph includes additional recurrent and indirect influences.
For right TTMn, the original subnetwork’s recorded peak is 0.034008, compared with 0.024912 in the full typed model, a roughly 36.5% overestimate relative to the full result. Those values belong to the original fixed-input experiment, not the arena. Inspect the full-CNS comparison.
Preserving each target’s original input denominator prevents one artificial source of strengthening after cropping. It cannot restore omitted activity. The smaller model is useful for interactive experiments, but its outputs need comparison with a larger model before quantitative conclusions are carried across scales.
The project includes checks for graph integrity, preserved original edges and input totals, zero-input silence, directed propagation, silencing, bounded activity, deterministic trajectories, and agreement between numerical implementations. Arena checks cover multiple scenarios and motor interventions.
These are implementation and structural checks. They do not measure agreement with neural recordings, muscle forces, real flight trajectories, or intervention experiments in animals. Browser visual and end-to-end testing were not performed during construction; numerical tests do not certify the appearance on every device.
You can run the checks from the downloaded package:
node --test tests/*.test.mjs
node tests/validate_arena.mjsThe model is most useful when a surprising output leads to a sharper question about the circuit or its assumptions. The user guide provides the controls for beginning that process.