# Saved steering-input attribution · 2026-10-28 This is analysis of the earlier 48 full-network sensory trials. It does not add 48 independent brain experiments. Two output cells are replayed with upstream spike trains clamped to each original full-network recording. Their input traces may include the original circuit's feedback, but the replay does not recompute that feedback. This is not a validated two-cell substitute, a causal intervention, learning, movement or driving. ## Data and audit Download every file listed in this manifest.json, retaining the directories. Verify all lengths and SHA-256 values. Also download the earlier brain-only-steering-v01 release from https://dropkit-contents.pages.dev/research/brain-only-steering-v01/manifest.json with its broad/ and dm1/ layout. All 48 original sensory spike files, summaries, protocols, runners and index maps are preserved there. Earlier direct controls remain separate and are not replayed in this analysis. Use NumPy, Pandas and PyArrow from the saved environment locks; the independent audit imports neither Brian nor the analyzer. ```sh python validate.py . --saved-root ../brain-only-steering-v01 --receipt audit-repeat-v01.json ``` Choose a new receipt filename to repeat: existing receipts are never overwritten. The audit recomputes 71,664 connection/trial contribution rows, timing bins, state summary metrics and threshold crossings against all 48 original target spike-time sequences. This is a saved-data consistency audit, not a second numerical integrator or independent full-brain rerun. The files in results-v02/ include: - incoming-synapses.parquet: all 1,493 original rows entering the two outputs, with original IDs/indices, contact counts and model signs. Each ID/index pair was checked against the original model index. - presynaptic-annotations.csv: all 1,411 distinct source cells, from pinned FlyWire 630 annotations v1.1.0. Selected sources have unique annotation roots;82 duplicate rows elsewhere in the full annotation table are outside this selection. - source-contributions.parquet: every incoming row for each saved trial. inputStateJumpSum_mV is the nominal signed weight times arriving source-spike count. Interpret it as a weighted-arrival sum before decay, temporal integration and refractory handling, not a measured voltage/current or causal effect size. - arrival-bins.csv: positive and negative weighted arrivals per 10 ms, including the original 1.8 ms synaptic delay. Count arrivals in 200–1000 ms; source events whose delayed arrival is outside the recording are excluded. - metrics.csv: 96 output/trial records, including reconstructed g/v statistics. The model-g state has voltage units; it is not a physical conductance or current. - 48 trace Parquets: step 0–9999 at 0.1 ms, reconstructed left/right v and g before threshold testing and their refractory eligibility. Original full-network membrane traces were not recorded; these are derived traces. Target spike times were matched at the original 0.1 ms resolution. - top-contributors.csv: ten retrospective positive/negative contributors per output in each of 18 intact conditions across six seeds. Three conditions within a seed are paired, not independent observations. - summary.json and independent-validation receipts: source hashes, original7-file before/after Git-blob checks and per-trial target-time results. An execution/audit pass is separate from the earlier failed steering criterion. facts.json describes the largest model-negative contributor to right DNa02,630-annotated AOTU019-left 720575940627100584, and six right-only input cases. Its 44–48% share refers only to nominal negative weighted arrivals, not a share of a biological steering mechanism or a demonstrated causal importance. ## Fresh conditioned replay on Linux The analyzer is executable at top level; do not import it. Use a fresh directory and preserve existing outputs. It reads author commit 91bdd1e7dcf193f3e7ca5a8933497fcef63b7960, the source/data files in source-manifest.json and saved parent directories steering-v01/ and steering-dm1-v01/. Download the 21 MB annotation table from https://raw.githubusercontent.com/flyconnectome/flywire_annotations/df6bb136f5b3d91c3992df4e8de2642329e2a384/supplemental_files/Supplemental_file1_annotations.tsv into steering-v01/sources/annotations-v1.1.0.tsv. Its expected SHA-256 is 55c99c61eecf8db6cc36f1a684b35e4c4208afbab02197d753ccc8dc2a6e2e76. The plan's source_root points to the existing verified Linux source environment. Change that path only in a separately recorded plan/version; the digest then changes. Recreate the original source files and preserve the old saved-run runner/protocol bytes. The full original model is not executed by this analyzer; its default equations, synaptic weights, delay, reset and refractory settings are reused for the two conditioned outputs. No direct sensory voltage injection targets those outputs in these 48 trials. ```sh env OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 NUMEXPR_NUM_THREADS=1 nice -n 15 python analyze-v02.py --out results-v02 > analysis.log 2>&1 python validate.py . --saved-root .. --receipt independent-validation.json ``` The first implementation stopped before replay because it demanded unique IDs across the entire annotation table. Its incomplete results-v01 and log remain on Linux. Version02 checks uniqueness only within the selected 1,411 sources; no original event, source model or selection was altered. plan.json retains its initial prepared-not-run state; completed timestamps are in results-v02/summary.json. It is a locally fixed analysis plan, not external study preregistration. Cross-hardware/dependency reproduction has not been demonstrated. ## Remaining causal test · not run In the complete original network, selectively zero only the model-negative connection AOTU019-left 720575940627100584 → right DNa02 720575940604737708, then restore it before a fresh paired run. Preserve every other connection, test both left and right inputs using newly fixed heldout seeds, and retain the original direction criterion. More right-output spikes alone would not demonstrate correct steering. No such intervention is part of this release.