digital-drosophila/registry/catalog-20260929/evidence/phase4x-transmitter-provenance/REPORT.md

Phase 4X — transmitter provenance and audit-role repair

Status: AUDIT-CLASSIFICATION-REPAIR-PASS. No frozen brain/sign/raw changes; no learning candidate.

All original annotation and neurotransmitter columns match raw by body_id. Runtime labels equal raw consensus_nt; reconstructed sign and effective weights match frozen runtime. DPM has raw predicted/consensus dopamine with no ground_truth. Its confidence is a supplied field, not established calibration probability. Repair: separate transmitter metadata, exact DAN annotation class, exact DPM annotation type and frozen compartment-mapping evidence. Do not infer reward valence or DAN identity from transmitter/name alone.

KC neuron_idx	conditional replay	late spikes range (6 known odors)
59832	base	15–17
59832	omit_DAN_annotation	0–0
59832	omit_DPM_annotation	0–0
59832	omit_dopamine_without_DAN_or_DPM	15–17
59832	omit_legacy_metadata_dopamine	0–0
78836	base	39–40
78836	omit_DAN_annotation	0–0
78836	omit_DPM_annotation	14–16
78836	omit_dopamine_without_DAN_or_DPM	39–40
78836	omit_legacy_metadata_dopamine	0–0
82312	base	60–61
82312	omit_DAN_annotation	59–59
82312	omit_DPM_annotation	60–61
82312	omit_dopamine_without_DAN_or_DPM	60–61
82312	omit_legacy_metadata_dopamine	59–59

Six old odors, one fixed seed per odor; no new blind generalization claim. Other recorded arrivals and APL remain fixed in isolated replay. The last APL subtraction is not applied in CPU comparisons because stored V/g are post-state, pre-APL; it cannot affect any recorded event/state and no final post-APL g comparison is claimed. Consensus/ground_truth agreement is a provenance check, not independent prediction accuracy. No confidence threshold, label correction, sign change or gain change is inferred.