# Phase 4W — late KC feedback audit

{"CPU_replay_exact": true, "candidate": false, "full_event_repeat_exact": true, "learning": false, "paths": 36, "phase": "4W", "prior_V_exact": true, "reference_pass": true, "rooted_time_supported_paths": 36, "status": "COMPLETE-CONDITIONAL-MECHANISM-AUDIT", "trials": 6, "whole_brain_causal_necessity_proven": false, "integrity_pass": true}

No learning or model changes. Inputs replay six old Phase4V offset trials. All frozen source artifacts remain read-only.
CPU single-cell omissions hold all other recorded arrivals and the applied APL overlay fixed. They are conditional input tests, not whole-brain causal interventions.
Ancestry uses real edges and actual event timing. Ranking the largest positive decayed latest-spike contribution is a heuristic, not a unique causal explanation. All inhibitory arrivals and graded APL are retained in the separate drive audit.

| neuron_idx | replay mode | mean late spikes (6 odors) |
|---:|---|---:|
| 59832 | base | 16.333 |
| 59832 | omit_APL | 16.333 |
| 59832 | omit_APL_overlay | 143.833 |
| 59832 | omit_DAN | 0.000 |
| 59832 | omit_KC | 15.167 |
| 59832 | omit_MBON | 28.667 |
| 59832 | omit_PN_olfactory | 0.000 |
| 59832 | omit_other | 0.000 |
| 78836 | base | 39.667 |
| 78836 | omit_APL | 39.667 |
| 78836 | omit_APL_overlay | 39.667 |
| 78836 | omit_DAN | 0.000 |
| 78836 | omit_KC | 39.000 |
| 78836 | omit_MBON | 46.667 |
| 78836 | omit_PN_olfactory | 39.667 |
| 78836 | omit_other | 49.667 |
| 82312 | base | 60.500 |
| 82312 | omit_APL | 60.500 |
| 82312 | omit_APL_overlay | 115.000 |
| 82312 | omit_DAN | 59.000 |
| 82312 | omit_KC | 60.500 |
| 82312 | omit_MBON | 60.500 |
| 82312 | omit_PN_olfactory | 0.000 |
| 82312 | omit_other | 65.667 |

The six reserved novel test odors remain untouched. No candidate, parameter selection, threshold/gain change, eligibility modification, training or game integration.
Full evidence: drive/summary.parquet, replay/results.parquet, ancestry/event-supported-edges.parquet, regression/comparisons.json.