A residual DNp15 input: DNb03 cut and restore
Research record
- Question
- Does a remaining direct input explain the B-pattern residual after the original eight-row route cut?
- Evidence
- Post hoc ranking of all 175 original DNp15-right incoming rows using saved B-cut spikes, followed by a frozen single-row intervention on three new input seeds.
- Experiment
- Linux: 15 fresh full-network simulations, three seeds by quiet, B intact, original eight-row cut, additional DNb03 cut, and fresh DNb03 restoration with the eight-row cut retained.
- Results
- Right DNp15 residuals 6, 7, 7 become 0, 0, 0 after the additional cut and return to 6, 7, 7 after fresh restoration. Full ordered network spikes, driver events and weight hashes match for all three restored pairs.
- Limitations
- Synthetic fixed inputs and one connectome model; three random input seeds are not independent animals. Discovery is post hoc, biological identity is annotation-based, and neither vision-guided steering nor driving was tested.
- Reproduction
- Pinned protocol, executed script, discovery tables, saved spikes, driver events, hashes and audit are linked in the evidence manifest.
Sources
Run evidence
Question and result
The prior paired-input experiment left right DNp15 active after eight HS/H2 connections were cut. Here a ninth connection, from a neuron annotated DNb03 right to DNp15 right, was tested without changing that original cut.
| Input seed | Intact B, left / right | Original eight cut | Additional DNb03 cut | DNb03 restored |
|---|---|---|---|---|
| 20261131 | 0 / 63 | 0 / 6 | 0 / 0 | 0 / 6 |
| 20261132 | 0 / 62 | 0 / 7 | 0 / 0 | 0 / 7 |
| 20261133 | 0 / 63 | 0 / 7 | 0 / 0 | 0 / 7 |
Counts cover 200–1000 ms. The new primary field specified a decrease of at least one right DNp15 spike in every new seed: 3/3 met. The frozen file also retained conflicting criteria from the prior A/B experiment; see the protocol clarification below. Zero residual activity was observed in all three, but was not the primary criterion. With no admitted input, the whole network produced zero spikes in each quiet trial.
How the candidate was chosen
Before inspecting the ranking, we specified an analysis of all 175 original incoming rows to DNp15 right using the previous three B-cut trials. For every residual output spike, presynaptic arrivals in the preceding 20 ms were counted using the model’s 1.8 ms delay and weighted by positive contact count. Positive uncut non-driver rows active in all three discovery seeds were ranked by total weighted arrivals, with row number breaking ties. Nine rows qualified. The highest was the DNb03-right → DNp15-right row, with 47 positive contacts in the pinned model (row 1113765; source root 720575940611016910; target root 720575940623431351). The v630 identity is an annotation, not an independent biological validation.
An initial discovery script used inherited table index labels as simulator row positions and erroneously included already-cut rows. Its output was rejected before the experiment. The corrected version resets the positional index, checks all four original right-side cut rows against their source IDs, and applies the same fixed ranking. Both versions and the correction are preserved in the manifest. Weighted arrivals are an association, not reconstructed current or proof by themselves; repeated arrivals can contribute to more than one target spike.
Protocol clarification
The frozen protocol inadvertently retained the earlier A/B experiment’s route_intervention, primary_direction and distinct_controls fields alongside the new B-only primary and controls fields. This conflict was identified after execution. The unchanged file is internally ambiguous, so the result should not be presented as an unambiguously pre-specified test. The executed conditions and new fields define the comparisons reported here. See the full erratum; no A condition was run in this experiment.
Intervention and controls
We froze the selected row, three new integer seeds, inputs and criterion before running. Each seed used five fresh full-network simulations: quiet, B intact, B with the original eight rows zeroed, B with those eight plus the selected row zeroed, and a fresh restoration in which the selected weight was restored before the run while the original eight remained zero. Runtime checks compared all 14,687,178 weights to the expected arrays; the seven pinned source files matched before and after execution.
Each active B condition received identical recorded driver events within a seed. All three restored runs exactly matched their original-eight-cut counterparts in the ordered whole-network spike events, driver events and full weight hash. Membrane trajectories were not recorded or compared. This is a fresh-run restoration, not recovery during a running simulation. The Linux job used one CPU and took about 64 seconds for 15 trials.
The 2,701 eligible T4 cells and the artificial pulse schedule match the previous paired experiment: two 50,000-candidate/s template streams, stimulus 200–700 ms, time step 0.1 ms, and no direct stimulation of DNp15. Input subtype assignment and pulse strengths are modeling assumptions. All seeds use the same model and are not separate animals. We did not test A-pattern generalization, new learning, a live visual bridge, a body, steering or driving. The result supports a causal effect of this model connection under these conditions, not biological necessity or a unique explanation for all residual activity.
Update record
-
2026-10-28: Completed the saved-data discovery and held-out intervention on Linux. Next: assess other inputs and a faithful visual bridge; those tests are not run.
-
2026-10-29: Follow-up: new native inputs and a rectified visual bridge; new-pattern results were inconclusive and A visual direction still failed.