Digital Drosophila Archive
3.5 KiB
3.5 KiB
Phase 4R — KC–US timing audit
Status: COMPLETED-DIAGNOSTIC-NO-CANDIDATE. No training, new timing selection, weight update or next phase.
Fixed-event matched/random ratio
| Seed split | Window (ms) | Random [0,120) | Random [30,110] |
|---|---|---|---|
| calibration | 5 | 0.017351 | 0.131271 |
| calibration | 10 | 0.040701 | 0.230398 |
| calibration | 20 | 0.067120 | 0.167925 |
| calibration | 50 | 0.058699 | 0.070346 |
| calibration | 100 | 0.494515 | 0.457529 |
| calibration | 250 | 0.902178 | 0.763949 |
| heldout | 5 | 0.342145 | 2.648988 |
| heldout | 10 | 0.225614 | 1.337659 |
| heldout | 20 | 0.149717 | 0.392041 |
| heldout | 50 | 0.080961 | 0.099113 |
| heldout | 100 | 0.495901 | 0.463649 |
| heldout | 250 | 0.898125 | 0.765088 |
Actual target-compartment KC timing
- {"split": "calibration", "label": "CAL_A", "peak_bin_start_ms": 20.0, "peak_spikes_per_5ms": 42.8, "spikes_85_90ms": 0.0, "fraction_events_first30ms": 0.9137931034482758, "median_last_spike_age_at_US_ms_edge_weighted": 69.05000000000001, "p10_last_spike_age_ms": 65.0, "fraction_edges_recent_20ms": 0.000474158368895211}
- {"split": "calibration", "label": "CAL_B", "peak_bin_start_ms": 15.0, "peak_spikes_per_5ms": 36.8, "spikes_85_90ms": 0.4, "fraction_events_first30ms": 0.9011627906976745, "median_last_spike_age_at_US_ms_edge_weighted": 71.3, "p10_last_spike_age_ms": 67.39, "fraction_edges_recent_20ms": 0.000474158368895211}
- {"split": "heldout", "label": "CAL_A", "peak_bin_start_ms": 20.0, "peak_spikes_per_5ms": 46.4, "spikes_85_90ms": 0.4, "fraction_events_first30ms": 0.9108910891089108, "median_last_spike_age_at_US_ms_edge_weighted": 69.0, "p10_last_spike_age_ms": 65.4, "fraction_edges_recent_20ms": 0.000474158368895211}
- {"split": "heldout", "label": "CAL_B", "peak_bin_start_ms": 15.0, "peak_spikes_per_5ms": 37.2, "spikes_85_90ms": 0.8, "fraction_events_first30ms": 0.8902821316614421, "median_last_spike_age_at_US_ms_edge_weighted": 72.60000000000001, "p10_last_spike_age_ms": 67.8, "fraction_edges_recent_20ms": 0.000474158368895211}
Verification
- Phase 4Q events and credit reproduced: True
- Ten full-runtime repeat runs exact: True
- Five frozen trials, two repeats each: True
- Equal delay on both pathways preserves relative timing: True
Interpretation and limits
- Phase 4Q failure reproduced only for its declared timing protocol; it is not evidence against all local credit rules.
- Early target-KC activity and late 90ms reinforcement are measured associations. No APL ablation was authorized, so APL causation remains unresolved.
- The exact uniform-time expectations remove random-US sampling noise on the same KC event stream.
- US=0..120ms versus 30..110ms is a protocol difference, not an event-energy mismatch. Both use one forced spike per selected DAN.
- The same 1.8ms delay on KC and DAN leaves relative timing unchanged. Counting only KC arrival is a different gate convention.
- Pre-existing Phase 4Q controls were offline algebraic tests. This audit adds recorded events, disk checkpoint restoration, real full-runtime repeats and frozen reference comparisons.
- No timing/eligibility candidate was selected. Independent heldout seeds replicate diagnosis; heldout odor generalization was not tested.
- Heldout 5ms/10ms credit exceeds the narrower random-N expectation (2.649x/1.338x), but calibration does not replicate this and all ratios remain below the prior 3x gate.
Final integrity verification is recorded separately in audit/frozen-integrity.json and the root checksums.sha256.