# Engineered visual-to-spiking bridge and direct-drive diagnostic This is an open-loop neural interface experiment, not a reconstructed animal or body controller. The first interface failed its prespecified direction criterion. The later diagnostic directly stimulated upstream visual neurons and passed; it does not repair the visual interface or establish autonomous behavior. ## Sources and layout - Original spiking model: https://github.com/philshiu/Drosophila_brain_model/tree/91bdd1e7dcf193f3e7ca5a8933497fcef63b7960 - Visual model: https://github.com/TuragaLab/flyvis/tree/0a69fa2fc8e1bfe1d837d722c252ab5f5d6fcbfa - Original annotation: https://github.com/flyconnectome/flywire_annotations/tree/df6bb136f5b3d91c3992df4e8de2642329e2a384 - Later annotation: https://github.com/flyconnectome/flywire_annotations/tree/ebd66db2596fcc39c6950fb54ea3efa00f7fe8a0 - HS/H2 circuit study: https://doi.org/10.1038/s41593-025-01948-9 - T4 subtype directions: https://doi.org/10.7554/eLife.40025 The public `run.py` and `diagnose.py` copies replace only the private absolute baseline location with a portable relative path. `script-normalization.json` records executed-original and public-copy SHA-256 values. Original execution summaries retain the original hashes; the public auditors accept either an exact runner hash or this explicit normalization mapping. Private executed scripts and all numerical data remain unchanged. Recreate the documented Linux runs layout; record a new script hash for any further edits. Source/data Git blobs are checked before and after execution. Place v630 `annotations-v1.1.0.tsv` at `runs/2026-10-28/steering-v01/sources/annotations-v1.1.0.tsv`, and the later `annotations-v2.1.0.tsv` at `runs/2026-10-28/visual-steering-v01/sources/annotations-v2.1.0.tsv`. Their source commits and hashes are recorded in `source-provenance.json` and `structure-v01/summary.json`. The later subtype labels actually select the candidate inputs after exact original-ID and v630-side filtering; this is an annotation-transfer assumption. The baseline source path is `runs/2026-10-26/2026-10-26_fly-baseline_v01/source`; its seven original files and source manifest must come from the pinned model. The two dependencies are isolated: `requirements-visual.txt` for Flyvis and `requirements-spiking.txt` for Brian2. The visual source snapshot is installed as `0.0+g0a69fa2` using setuptools-scm's explicit snapshot version; this is not an official release number. The official pretrained ZIP has SHA-256 `71c78d4070556a536b13b23ee3139cd2788aa2a9d07d430a223b4edead281db1`; the selected checkpoint SHA-256 is `d0e42857e738d0315897c2d50fde9eb3fb1a3fb1f071d55dfc13d53d72bccc3f`. ## Execution sequence 1. With the original full source and pinned annotation files present, run `prepare.py` under the spiking environment. This generates `structure-v01`. 2. Freeze `protocol.json` before observing any motor simulation. 3. Run `project_visual.py` under the visual environment. It copies the fixed checkpoint into a new data directory, saves all 721 columns of each T4a/b type for two moving movies and a uniform-grey control, then computes the population projection and original-model identity/side eligibility. 4. Run `run.py` under the spiking environment. It records a local-clock preflight hash receipt, generates common per-cell random numbers, and executes 30 fresh full-network trials into `results-v01`. Only the specified eight chemical rows are zeroed in route cuts; fresh restores reinstate their original weights. 5. Run `validate.py --source-root ` in the spiking environment to reconstruct the means/rates, random candidates, driver events, all pathway counts, direction criteria and complete spike-event restorations. 6. After the bridge's failure, freeze `diagnostic-plan.json` and run `diagnose.py`. This executes nine fresh full-network trials with new seeds and matching four-channel direct HS/H2 candidates. Run `validate_diagnostic.py` afterward. 7. Run `analyze-v02.py` to produce `analysis-v02` from saved data only. Use fresh output folders; the simulation scripts refuse to overwrite runs. The recorded Linux invocations used one CPU thread and nice15, a 30-minute timeout for the main run and 15 minutes for the diagnostic: ```sh OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMBA_NUM_THREADS=1 nice -n 15 timeout --kill-after=30s 1800 -u run.py OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MPLBACKEND=Agg validate.py --source-root OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 nice -n 15 timeout --kill-after=30s 900 -u diagnose.py OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 validate_diagnostic.py ``` ## Scope of the evidence The projection uses **positive changes in population-mean model activity**. It is not Flyvis's native per-neuron rectified synaptic-output function, and does not preserve individual retinotopy, signed deviations below baseline, biological firing rates, electrical synapses or graded-cell dynamics. Stable root membership plus matching annotation sides is a lookup check, not independent biological identity or spatial validation. The gain is an engineering assumption fixed before motor trials, not calibrated using desired steering. `results-v01` includes the actual driver monitors and whole-network spikes. The diagnostic archived candidate schedules and actual whole-network spikes; its actual driver equality was asserted inside the runner, while separate per-trial diagnostic driver-monitor files were not archived. Audits verify saved data and source hashes, not a new neural inference or biological validity. The first post hoc analysis used original weights as reference values even on cut rows, with an inactive-edge flag. It remains private as `analysis-v01`. `analyze-v02.py` records zero actual weighted arrivals for cut rows. Intact curves, primary counts and criteria were unaffected. Nominal arrivals are model weight times arriving spike count before decay, integration and refractoriness; they are not membrane voltage, measured current or a biological causal fraction. The public manifest covers exported files. Logs have private run roots redacted and newline formatting normalized; original raw logs remain in the Linux and private local run. No model weights were trained or tuned after the results.