# Reproduction and saved-data audit

Verify the SHA-256 values in `manifest.json`. The two separate `protocol.json` and `run.py` files were hashed before execution in each `frozen-sha256.txt`; these local receipts are not independently timestamped registrations. The inputs, seeds, inclusion rule and conditions were fixed before either run. The two right-only variants each contain six fresh trials; left-only and original-right controls are reused from the earlier matched-side bundle.

Reconstruct the Linux layout with this bundle's `reverse-assignment/` and `rotate-right-27/` folders at `runs/2026-10-30/paired-selection-reverse-assignment-v01/` and `runs/2026-10-30/paired-selection-rotate-right-27-v01/`, respectively. Restore the original model and environment at `runs/2026-10-26/2026-10-26_fly-baseline_v01/`, bridge metadata at `runs/2026-10-28/visual-steering-v01/`, previous matched-side evidence at `runs/2026-10-29/paired-hemisphere-v01/`, and visual-factor saved rates at `runs/2026-10-29/visual-factor-v01/results-v01/`. The frozen protocol condition lists are correct. The inherited `stimulus`, `outputs_rule` and `controls` descriptions also mention quiet, left-only and bilateral trials; those are stale descriptions from the prior study, not conditions executed in this study. Input equality and comparison to the previous saved trials were checked after the run. Prior published evidence contains source references and copies of required inputs. Rerun only in fresh folders: the runners refuse to overwrite `results-v01` or `cache-v01`. Use the same pinned model and dependency versions as the previous experiment.

`audit/audit-v06.py RUNS_ROOT NEW_AUDIT_OUTPUT_DIRECTORY` checks frozen hashes, original source-hash receipts before and after, recorded weights, selected IDs, matching saved driver/candidate tables, trial file hashes 48 new output values re-counted from spike events, and the reused reference output values re-counted from their saved spike files (96 checks including duplicates). It requires the two previous reference directories in the layout above. The first audit script failed at JSON serialization; the corrected script completed without re-running simulations. `audit/audit-v06.json` records comparisons and restrictions. Runtime code asserted in-memory driver-to-target synapse arrays, but did not export them for independent inspection. Neither audit reran neural dynamics or independently reread the two large original graph files.

The right-rotation condition changes both the 49 selected identities and assignments; it does not isolate either effect. The reversal changes only the mapping of event channels onto an unchanged selected set, not the driver event table. Neither set of pairs establishes retinotopic or anatomical correspondence. All other constraints of the original matched-side experiment apply.

## Prerequisites and commands

For saved-data auditing, restore the previous [matched-side results](https://dropkit-contents.pages.dev/research/fly-paired-hemisphere-v01/manifest.json) as `runs/2026-10-29/paired-hemisphere-v01/` and the pinned [source manifest](https://dropkit-contents.pages.dev/research/brain-only-feeding-v01/source-manifest.json) as `runs/2026-10-26/2026-10-26_fly-baseline_v01/source-manifest.json`. Put both new folders at the `runs/2026-10-30/` paths above. With Python and pandas/parquet support installed, run `python3 audit/audit-v06.py /path/to/runs /path/to/new-audit-output` from the extracted bundle. The output directory must not exist.

For a neural rerun, also reconstruct the original pinned model files listed by that source manifest and its [dependency lock](https://dropkit-contents.pages.dev/research/brain-only-feeding-v01/requirements.lock.txt); use Linux and fresh run directories. Restore [bridge membership](https://dropkit-contents.pages.dev/research/visual-neural-bridge-v01/visual-v01/bridge-neurons.csv), [pathway metadata](https://dropkit-contents.pages.dev/research/visual-neural-bridge-v01/structure-v01/pathway-neurons.csv), and [saved input rates](https://dropkit-contents.pages.dev/research/fly-visual-factor-v01/results-v01/injected-rates.npz) at the relative paths specified above. Run `python3 run.py` from each fresh variant folder. The saved-data audit does not rerun neural dynamics.