New T4 combinations and a rectified visual bridge

1. Does the residual connection generalize?

The previous test found that cutting a DNb03-right → DNp15-right model row removed residual firing under B-pattern input, with the earlier eight HS/H2 rows already cut. We kept those exact rows and compared four combinations of T4a/b input. For each of three new seeds, each pattern had intact, original-eight-cut, additional-DNb03-cut and fresh candidate-restored trials, plus quiet: 51 trials.

Each triplet follows seeds 20261201, 20261202, 20261203; counts cover 200–1000 ms. The prospectively fixed rule treated a zero baseline as indeterminate, not a pass or failure. C and D therefore do not establish generalization or absence of an effect. In each pattern and seed, the actual driver events matched across its four conditions. All 12 fresh DNb03 restores exactly reproduced the original-eight-cut condition’s ordered whole-network spikes, input events and full weight hash. Quiet had zero whole-network spikes in all three seeds. Input groups, two expected 50,000-candidate/s streams, refractory convention, pulse weight and delay were unchanged from the prior paired native-input experiment.

2. Does source-inspired rectification fix the visual bridge?

The visual-model source applies ReLU to individual neuron activity before synaptic weighting. The earlier bridge averaged activity first. We reused its saved Flyvis T4a/b movie responses and compared the earlier transform with a candidate that applies per-neuron ReLU before averaging, then applies the same baseline subtraction, positive clipping, 1e-6 threshold, denominator and input scaling. The first saved-array reconstruction omitted the 1e-6 threshold and failed its assertion; it was retained. The corrected comparison reconstructed the old encoder and its candidate remained below the prior 50,000 Hz per-eye aggregate limit on the saved frames. The planned temporal integrals were added from saved arrays after review, without rerunning either model. This transformation is still a population-level engineered bridge: it is not the complete Flyvis synaptic current or a biologically calibrated rate conversion.

We then ran 30 fresh spiking-circuit trials, using the three seeds and saved movies from the previous bridge to compare methods. These are matched-randomness comparisons, not held-out seeds. Different rates produced different actual input events; A/B trials gained roughly 4,100–4,300 events and lost 119–131 relative to the old encoder, despite shared random draws. Within each new pattern, cut and restore used identical driver events.

The fixed anatomical criterion required A left > right and B right > left in each seed. A failed 3/3; B passed 3/3, as with the earlier bridge. Rectification alone did not resolve the direction failure. All six fresh eight-row restores exactly matched the intact ordered network spike sequence, driver events and full weight hash; blocked-input DNp15 and DNa02 outputs were zero. In the grey condition each full network produced one spike, while both DNp15 outputs were zero. The protocol mentioned a separate quiet trial, but the executed condition list did not include one; input-block and grey are separate observations. Seven pinned original source blobs matched before and after both jobs. See the protocol deviations.

Limits and next step

The same model and input-noise seeds are not separate animals. Flyvis is a pretrained visual subsystem; no new training occurred. The mapping broadcasts population activity onto annotated T4 groups without verified individual retinotopy, real firing-rate calibration or physical left/right turn calibration. No live video, body, closed-loop steering or driving was tested. These observations do not establish why the direction fails or which candidate change would fix it.

Next, isolate spatial mapping, source identity and rate conversion with fixed controls before adding a body. Not run.

Update record

• 2026-10-29: Executed both Linux experiments and checked saved results. Preserved the failed initial saved-array reconstruction; recorded new-pattern indeterminacy and A-direction failure.

• 2026-10-29: Follow-up: visual input factors; subtype exchange reversed B, while equal-eye tests showed DNp15-specific right output.