Experimental model & evidence

The open lab is an illustrative environment. Local flight behavior is procedural. Live activity requires the separate authenticated neural service. Neither motion nor a successful connection establishes biological validity.

Historical numerical evidence and source provenance are included in the complete project’s docs folder. Full MaleCNS execution and CUDA have not been retested for this visual edition.

> Habitat Upgrade update: see [HABITAT-UPGRADE.md](HABITAT-UPGRADE.md) for current implementation and validation. The evidence below retains its original scope. Typed food uses assumed strengths; new soma/population telemetry does not validate biology. Final browser acceptance and CUDA remain unverified.

# Experimental model card

**Dataset:** MaleCNS v1.0, official HHMI Janelia/Google Research/collaborators.
**Importer scope:** non-null typed annotated neurons; graph restricted to those IDs.
**Original source:** https://male-cns.janelia.org/download/
**Research context:** https://blog.google/innovation-and-ai/technology/research/male-fruit-fly-brain-map/

The neural implementation is original, not copied from NeuroCraft or the Eon
FlyWire simulator. Anatomical counts constrain topology, not physiological
parameters. The current model is not evidence of a complete emulated animal.

## Explicit assumptions

- Dimensionless LIF voltage, threshold 1, reset 0, membrane time constant 20 ms.
- Default integration step 1 ms; decaying recurrent current with 5 ms time constant.
- Two-step delayed spikes and 3 ms refractory interval. No learned synapses.
- Raw synapse counts supply connection magnitudes. Presynaptic consensus GABA, histamine and
  glutamate predictions are assigned negative signs; other/unknown transmitters
  positive signs. This simplified sign assignment is not universally biological.
- Incoming rows are divided by the square root of summed absolute incoming counts.
  Recurrent contribution gain 0.18. These are stabilizing engineering choices.
- Independent baseline noise at 0.5 events/neuron/second, adding 0.15 voltage.
- Odor stimulation targets typed ORNs, separated by somaSide or rootSide; looming
  directly stimulates LC4 populations; light directly stimulates R1–R6 populations.
  These are modeled population injections, not validated compound-eye or receptor
  transduction. Homogeneous food odor does not model receptor-specific chemistry.
- Smoothed DNa01/DNa02/DNg13 readouts form left/right motor commands; DNp01 readout
  contributes to the escape primitive. Gains and body mechanics are engineered.
- Missing populations give zero input/output and remain visible in mapping counts.
- Co-residents affect body collisions; no neural conspecific-vision mapping exists.
- No sex variants, aging, development, biological learning, hunger metabolism,
  fertility, social behavior or realistic aerodynamics.

`mapping.json` holds actual retained graph indices; `neuron_ids.npy` and
`neurons.csv` map them back to original research IDs and annotations. Raw and
processed hashes are recorded. Integrity hashes identify retrieved files; they do
not constitute upstream signatures or biological validation.

## Validation priorities

1. Verify selected pathways and stimulation/readout mappings against the primary
   publications and available physiological data.
2. Compare intact, no-input, silenced and shuffled controls with a fixed decoder.
3. Test integration-step sensitivity and model-parameter sensitivity.
4. Calibrate food and looming responses before describing movement as natural.
5. Optimize the GPU implementation only after output parity checks.
6. Add compound vision/physical bodies as separate measured integrations.

A successful graph import or moving 3D fly is not proof that these priorities have
been satisfied. The visible preview must retain its mode labels.

## v0.2 corrections and evidence

See `SCIENCE-AUDIT.md` for the numerical corrections, graph audit and causal
experiments. Photoreceptors are still modeled as spiking units although their
biological signaling is graded; sign correction alone does not fix this. Use the
v0.2 importer and a fresh checkpoint directory. CPU active-column recurrence is
compared against CSR; CUDA parity remains untested.