Two KIOS detectors
320px YOLO11n baseline versus 480px YOLO11n with image augmentations. The source’s “Phase 22 Baseline” is a detector, separate from the Phase 22 simulation model.
Phase 23 · KIOS protected detector benchmark
The detector gained recall.
Clean and blur improved. Occlusion and mixed stress got worse.
Detector comparison. 86 protected real-video frames; separate from the frozen Phase 1–22 simulation record.
59.3%Clean recall · +20.9 percentage points vs detector baseline
55.5%Clean mAP50 · +12.9 percentage points
13.8%Occlusion mAP50 · −21.0 percentage points
8.1%Mixed-stress recall · −10.5 percentage points
Mixed result: clean precision fell 79.6% → 71.7%; mAP50–95 fell 27.2% → 22.3%.
Six condition examples
Illustrative views from the published set. The aggregate table below summarizes all 86 protected frames.
These are condition examples, not detector-output overlays or six independent test sets. Published metrics are condition-level aggregates over the same 86 protected frames.
What changed by condition
Deltas are percentage points (pp).
Swipe sideways for every metric →
| Condition | Baseline mAP50 | Phase 23 mAP50 | mAP50 Δ | Baseline recall | Phase 23 recall | Recall Δ |
|---|---|---|---|---|---|---|
| Clean | 42.7% | 55.5% | +12.9 pp | 38.4% | 59.3% | +20.9 pp |
| Blur | 41.3% | 58.6% | +17.3 pp | 37.8% | 57.0% | +19.2 pp |
| Low light | 41.7% | 42.4% | +0.7 pp | 37.2% | 44.2% | +7.0 pp |
| Noise | 43.3% | 59.6% | +16.4 pp | 39.8% | 54.7% | +14.9 pp |
| Occlusion | 34.8% | 13.8% | −21.0 pp | 32.6% | 24.4% | −8.1 pp |
| Mixed stress | 18.5% | 12.9% | −5.5 pp | 18.6% | 8.1% | −10.5 pp |
The weak tail: both mAP50 and recall fell under occlusion and mixed stress. See the Phase 24 charts.
Method and boundary
Phase 23 trained the detector; Phase 24 reanalyzed the results.
Temporal split
252 train, 64 validation, 20 embargoed, and 86 protected test frames. Conditions reuse the test frames.
Original Phase 23 replay verified: 24 / 24 metric cells matched
The original best.pt checkpoint was recovered without substitution. SHA-256: 43240be969708c32b3e11340c9846b3a588baf361378398677c794d16a455310. Fresh inference matched all 24 published aggregate cells with maximum delta 0.0, then regenerated four prediction tables byte-for-byte. This was a pinned original-model replay, not a newly trained model or a complete clean-room reconstruction of every original dependency.
The recovered lock matched Python 3.13.2, Ultralytics 8.4.160, Torch 2.7.0+cu118, NumPy 2.2.3 and Pillow 11.0.0. Phase 23 retained CUDA device 0 and 480 px; the baseline retained its separate 320 px CPU settings.
Checkpoint packaging alone does not rerun inference. The historical bundler remains available as py scripts\bundle_phase23_checkpoint.py; use the current replay commands and receipt to check outcomes.
The 516 cross-model views and 860 within-model clean-versus-stressed pairs answer different questions. Neither is independent validation. Read the paired research record →