Visual input factors: subtype, timing and equal-eye controls

Question

The previous rectified visual bridge still produced right-dominant DNp15 output for A, even though its anatomical criterion expected left dominance. We tested whether the assigned T4 subtype or temporal shape changes this output, and what happens when both eyes receive the same movie.

Three new input-noise seeds, nine conditions each: 27 fresh full-network trials. Every condition used the same original model weights and shared per-cell uniform random draws within a seed. Counts below are DNp15 left / right spikes from 200–1000 ms; they are anatomical model labels, not calibrated physical turn directions.

A and B were both right-dominant at baseline. Exchanging subtype rates reversed B to left-dominant in all three seeds, while A remained right-dominant. Flattening the temporal waveform did not reverse either pattern in these trials. The prior criterion expected A left > right and B right > left; subtype exchange did not satisfy both. These are exploratory within-model comparisons, not a verified controller or a uniquely identified failure cause.

Equal-eye controls

Both movies produced right-side DNp15 output in this specific bridge and model. This is not a claim about all descending outputs: DNa02 was left-only in four of these six trials and silent in two. This exposes an asymmetry worth investigating, but does not identify whether it originates in the circuit, membership transfer, rate conversion or their interaction. Group sizes differ by side, and sampled events differ even with shared random draws.

Controls and limitations

Subtype swap exchanges the two aggregate T4a/T4b rate traces within each eye, then divides by the receiving group’s size. The expected total per eye is unchanged at every time step. The flat condition replaces each group’s 200–1000 ms waveform by its mean over that interval, preserving its expected time-integrated input while leaving pre-200 ms unchanged. Neither guarantees identical realized events; the actual driver and candidate records are included. Both eyes remain within the previous maximum aggregate rate. Runtime assertions checked all 27 weight arrays against the original pinned hash and recorded seven source blobs before and after; quiet produced zero whole-network spikes on all three seeds. An independent saved-data review checked the recorded hashes and five small source files; the original completeness and connectivity files were not independently reread for that review. As in the previous bridge, all stimulated T4 group neurons use a zero-millisecond refractory setting; other neurons retain the model default. The frozen protocol did not separately restate that carried-over setting; see the reproduction note.

The mapping broadcasts population signals to annotated groups. Swapping their rates deliberately tests an alternative assignment; it does not prove either biological identity or connectivity is wrong. The visual model was pretrained, and no new learning occurred. Individual spatial mapping, measured firing rates and physical left/right turn mapping are unvalidated. No live video, body, closed-loop steering or driving was tested. The three seeds are input-noise realizations of one model, not three animals.

Next: investigate circuit-side asymmetry with matched controls and only then consider a spatial mapping supported by corresponding visual coordinates. Not run.

Update record

• 2026-10-29: Completed 27 Linux trials, saved all event data and checked the expected-rate invariants and model hashes. No parameters or output sign were changed after viewing results.

• 2026-10-29: Follow-up: matched left and right T4 stimulation. Right-only input produced more ipsilateral DNp15 firing in the selected subsets; the pairing is not validated homology.