Do T4 selection and event assignment explain the DNp15 difference?
Research record
- Question
- Does the previously observed DNp15 difference persist under two fixed changes to the right T4 population and event assignment?
- Evidence
- Saved matched-side trial outputs and driver events, pinned original spiking circuit, prior rectified Flyvis population-rate inputs and recorded T4 annotations.
- Experiment
- Twelve new Linux trials: two predefined right-side mapping variants, three reused noise seeds and two saved visual-rate patterns; previous left and original-right trials are reused as references.
- Results
- For all six seed/pattern combinations, right DNp15 remained above the saved ipsilateral left reference in both variants. Assignment reversal changed right counts by -2 to +1 spikes; rotation changed them by -1 to 0.
- Limitations
- One connectome, three reused seeds, two nonrepresentative mappings, no anatomical homology or calibrated retinotopy; rotation changes membership and assignments together. Runtime synapse routing was asserted but not independently exported. No causal pathway, body, steering or learning was established.
- Reproduction
- See the public evidence bundle for frozen protocols, runners, trial spike and driver files, comparison audit, source links, hashes and Linux layout. Prior reference trials and input files remain in their published bundles.
Sources
- Previous matched left/right experiment and reference trials
- Original spiking model at tested commit
- Prior visual factor experiment and saved input rates
Run evidence
The previous matched-input test found more ipsilateral DNp15 spikes on right-only stimulation than on left-only stimulation. Each side had 1,324 selected T4a/b cells, but the numeric-ID pairing was arbitrary and 53 eligible right-side cells were excluded. Can two different, fixed right-side mappings remove the observed difference?
Experiment
We kept the original network, saved E0/E180 visual-rate traces, stimulus amplitude, three noise seeds and 200–1000 ms readout window. We made two predetermined changes separately:
- Reverse assignment: retain the same selected right-side cells but reverse the order in which driver channels are assigned within each T4 subtype.
- Rotate selection: rotate the sorted right-side IDs by 27 within each subtype before choosing and assigning cells. This replaces 22 T4a and 27 T4b cells (49 of 1,324; approximately 3.7%) and also changes all right-side channel assignments.
Each variation has six fresh right-only trials: two input patterns × three seeds. The same three seeds and reference trials are reused across variants, and the E0/E180 patterns share random uniforms: these 12 comparisons are dependent, not 12 independent replications. The original right-only and left-only trials are saved references from the previous experiment, not newly repeated controls. Some descriptive fields copied into the frozen protocols still refer to the previous quiet and bilateral controls; only the explicit two right-only conditions were run in each variant. All 12 new right-side driver event tables exactly match their same-seed, same-pattern right reference. Left reference trials used the same candidate events in the preceding experiment. These are fixed synthetic population inputs derived from a pretrained visual model, not measured biological firing.
Results
DNp15 spikes on the stimulated side during 200–1000 ms. The first two numerical columns are saved reference trials; the last two are new trials.
| Pattern | Seed | Left reference | Right reference | Right reversed | Right rotated |
|---|---|---|---|---|---|
| E0 | 20261221 | 0 | 23 | 21 | 23 |
| E0 | 20261222 | 0 | 21 | 22 | 20 |
| E0 | 20261223 | 1 | 23 | 21 | 23 |
| E180 | 20261221 | 6 | 39 | 37 | 38 |
| E180 | 20261222 | 3 | 37 | 37 | 36 |
| E180 | 20261223 | 6 | 39 | 37 | 38 |
Right DNp15 remains higher than the saved left reference in all 12 comparisons. That pattern survived these two specific changes; the result does not prove it is robust to all cell selections, mappings, seeds or inputs. DNa02 left and right readouts remained zero in all 12 new right-only trials; this is a different output and is not evidence of physical steering.
Checks and limits
A separate saved-data audit re-counted all 48 new DNp15 and DNa02 left/right values from spike files and compared each new driver table with the previous right-side table. The audit checked frozen protocol and runner receipts, the exact selected IDs and indices, and the reused reference output files. The runner checked seven pinned source blobs before and after, checked unchanged native weight hashes and asserted its in-memory input-to-cell synapse indices during each run. The audit compared the saved source receipts; it did not independently reread the two large graph files. The latter indices were not independently exported, so saved driver tables alone do not independently establish delivery to the intended cells. The first audit script encountered a JSON serialization error after its checks; a corrected audit was run without rerunning neural simulations. The original failed audit is retained privately.
The two sides were not isolated, paired cells are not verified homologues, and the rate bridge lacks individual retinotopic and physical turn calibration. Reversing assignment and rotating membership are not independent tests of a particular biological pathway; the rotation changes membership and assignment together. The 2,701 originally eligible T4 inputs retained the prior 0 ms refractory modification. These results cannot establish the reason for the difference, real animal asymmetry, learning, body control or driving.
Next: select a specific candidate pathway using the recorded connectivity, freeze a matched-input cut/restore protocol, and test it. Not run.
Update record
- 2026-10-30: Executed 12 Linux trials and audited saved input events and output counts. Previous controls were reused; no body, training or driving was tested.