# Reproduction and audit

Verify every manifest SHA-256. `protocol.json` and `run.py` were hashed before the run in `frozen-sha256.txt`; this is a local receipt, not independently timestamped registration. To rerun, reconstruct the Linux directory layout: this bundle at `runs/2026-10-29/paired-hemisphere-v01`, the pinned original spiking model at `runs/2026-10-26/2026-10-26_fly-baseline_v01`, prior bridge membership and pathway metadata at `runs/2026-10-28/visual-steering-v01`, and prior saved rates at `runs/2026-10-29/visual-factor-v01/results-v01/injected-rates.npz`. A copy of the last file is included in `sources/`; previous public evidence bundles contain the pinned inputs and source references. Use a fresh run directory: the runner refuses to overwrite `results-v01`.

The experiment uses the first 640/684 lowest numeric IDs per side for T4a/b, pairing by within-group rank. The paired left/right cells are not known homologues or retinotopic matches. There are 53 right-side exclusions relative to the earlier full population. `results-v01/pairs.csv` and `excluded.csv` record IDs. The left-derived aggregate from the previous rates file sets the common expected rate; original right-versus-left aggregates differ by at most 0.003758 pulses/s from floating-point rounding. All 2701 originally eligible inputs have zero refractory, as in the previous bridge, even though only 2648 cells are selected; other neurons use default settings.

`audit.py` checks saved spike/driver hashes, counts, selected pairs, and equality of driver events within each movie and seed. `verify.py` independently reconstructs the saved candidates and drivers from the recorded random seeds and rates; both are saved-data checks by the project, not a rerun of the neural simulation. Runtime code asserted exact Brian synapse input/output arrays against the intended target pairs and checked the full weight hash and seven source blobs; the actual synapse connection arrays were not separately exported. Membrane trajectories were not archived. The whole brain remains connected in unilateral trials, so the comparison does not isolate a physically independent hemisphere. Public log paths are normalized; private originals remain on Linux.

Annotation clarification, added after review: T4a/b subtype labels come from the later annotation commit ebd66db2596fcc39c6950fb54ea3efa00f7fe8a0, transferred only for matching original-model root IDs and unique v630 side (v630 commit df6bb136f5b3d91c3992df4e8de2642329e2a384). This transfer is an assumption, not independently validated identity. `check_readouts.py` was added after review to recount 84 output/side values from saved spikes over [200,1000) ms; it does not change the frozen experiment or rerun dynamics.