Matched left and right T4 stimulation

Question

The previous equal-eye test produced right-only DNp15 output despite equal expected input per eye. Here the same sampled events were routed to a fixed left subset, a fixed right subset, or both. This tests whether the difference persists under matched event times; it does not isolate an independent hemisphere or prove why the response differs.

T4a/b subtype labels were transferred from the later v2.1 annotation only for original model root IDs with unique matching v630 side; this identity transfer is an assumption. For each subtype, we sorted original eligible IDs numerically on each side and paired by rank, selecting 640 T4a and 684 T4b cells per side. The previous right population had 22 additional T4a and 31 additional T4b cells, excluded here. Those pairs are a reproducible sampling rule, not validated cell homologues or spatial matches. We reused the previous E0 and E180 visual-rate traces, took the left-derived aggregate as the common rate for both sides, and used the same per-pair event times and amplitudes. The previous left and right aggregate arrays differed by at most 0.003758 events/s due to rounding. Per-cell rates on the originally larger right population therefore differ from the earlier full-population test.

Three new input-noise seeds × two visual-rate traces × left/right/both plus quiet: 21 trials. All network connections remained present in every condition. Counts below are DNp15 left / right spikes from 200–1000 ms.

For both traces and every seed, right-only stimulation produced more ipsilateral DNp15 spikes than left-only stimulation, despite identical driver event tables. Simultaneous stimulation was not the sum of the two unilateral outputs in these trials. This is a within-model response difference under the selected subsets; it does not establish that a single pathway or the biological right hemisphere caused it.

DNa02 differed from DNp15: under E180 left-only input it recorded left/right 8/1, 5/0 and 9/1, whereas E180 right-only yielded 0/0 in all three. Both outputs are included in the archived trial records; no output was chosen after the run to declare successful steering.

Controls and limitations

An additional saved-data check independently recounted all 84 DNp15/DNa02 output values in the 200–1000 ms window. The saved candidate and driver tables match exactly across left, right and bilateral routes for all six seed/trace pairs. Bilateral routing sends each event to both members of its pair, so its total intended target deliveries are twice those of a unilateral trial. The program also asserted the actual driver-to-target synapse arrays against the selected left/right IDs at runtime; those internal arrays were not independently exported. Runtime checks matched full weight arrays to the original pinned hash in all 21 trials and seven pinned source files before and after. A separate read-only review verified saved records and five small source files; the two large original graph files and in-memory weight arrays were not independently re-read for that review. Quiet had zero whole-network spikes on all three seeds. The 2,701 original eligible T4 cells retained the prior bridge’s 0 ms refractory setting, including the 53 excluded from direct drive; other cells retained model defaults.

These are synthetic population inputs, not measured biological firing. Numeric-ID pairs have no demonstrated anatomical equivalence, and left/right groups differ in omitted cells and connectivity. Contralateral connections remain active. Thus the result narrows the explanation beyond unequal event counts, but does not separate cell selection, wiring and network interaction, or establish biological or physical turn direction. No new learning, live vision, body, closed-loop control or driving was tested.

Next: use fixed alternative matched subsets and targeted circuit interventions to distinguish selection effects from specific pathways. Not run.

Update record

• 2026-10-29: Executed 21 Linux trials, recomputed candidate events from seeds, and checked saved output counts and source/weight records. No neural dynamics were rerun during saved-data verification.