digital-drosophila/models/male-cns-lif/pn-recruitment-audit-v1/README.md

Phase 4AI — PN recruitment and graded-APL source audit

Read-only, post-outcome conditional accounting of 80 already sealed Phase 4AG-r2 records. This is not a new whole-brain experiment, a learning trial, an independent brain oracle, or an APL-removal counterfactual.

Frozen contracts

• Reuse sealed Phase 4AH replay_core.py without modification; reconstruct actual ALPN targets (686 in the full graph).
• Keep the full frozen fast-APL effective-edge zeroing, delay, refractory, reset, threshold, ordering and float64 LIF coefficients. PN graded inhibition is exactly zero because the existing overlay applied only to KC.
• Separate accepted timed incoming-edge antecedents, non-feedback voltage lineage, and the legacy nonnegative graded-APL input map. The latter is not signed PN-to-APL LIF current.
• APL source groups: trial design ALPN, other ALPN, KC. Reject unaccounted nonzero input support. Compare summed mathematical float64 traces with stored float32 using prior analytic per-tick bounds, including grouping roundoff. Accumulated source shares remain descriptive; their robust majority is not proven.
• Strict timing: direct PN arrival at s+18 must precede the KC threshold tick; APL write at s follows the threshold and can affect only later ticks. Never-fired cells remain explicitly censored.
• Primary odor statistics retain fixed 686 coordinates; secondary statistics retain the fixed complement of six-odor design union. Technical repeats are excluded from odor comparisons.

Execution

Use the existing CPU environment, Python -B, and four-thread OMP/MKL/OpenBLAS/Arrow limits. No Torch import, GPU or parameter changes are needed.

1. prepare.py: synthetic tests, source/input hashes, prior integrity, fixed inventory, metadata schema and preregistration.
2. run_audit.py: exact PN conditional replay, event-stream provenance, APL split, twelve repeat checks.
3. finalize.py: independently re-read saved evidence, validate timed edge events, aggregate all fixed records, check prior integrity, seal model and experiment.
4. publication.py: verify both seals, then exclusively publish small evidence to a new registry. Bulk arrays stay on the server; Git is not their off-host backup.

Any failure is preserved. Do not edit a registered source or reuse a populated experiment directory. A new revision requires an explicit separate directory and record. Audit PASS never implies learning PASS, identified unique causality, or permission to tune frozen models.