Baseline
Loading resultsLoading source image…
- Matched
- —
- Unmatched
- —
- Missed
- —
The evidence.
In plain sight.
Aerial landing research, with the failures left in. Compare original detector predictions, explore camera stress, and see what the separate simulation actually established.
Original-model replay verified. Retrospective evidence, not flight certification.
Same view. Two models.
Full explorer →Phase 23
Loading resultsLoading source image…
- Matched
- —
- Unmatched
- —
- Missed
- —
AnnotationMatched boxUnmatched box
Loading the saved original-model record…
Illustrative reused KIOS frame, not a representative sample. Metrics count every prediction; overlays show all matched boxes and up to two unmatched boxes. Target matching permits false positives and does not mean landing success.
Original replay, exact match. 24 / 24 published metric cells matched exactly; four prediction tables matched byte-for-byte.
Check the receipt →The average improves.
The hard tail doesn't.
Phase 23 improves mAP50 in four of six camera conditions. Occlusion and mixed stress expose regressions that the average hides.
−13.3 pp mAP50 on the equally weighted occlusion + mixed-stress pair · Phase 24 reanalysis
Baseline measured · Phase 23 paired
Read the paired detector study →All six conditions. No hidden tail.
BaselinePhase 23
Published condition-level mAP50, not the selected-frame match above. Same 86 reused frames from two dependent sequences.
Chart unavailable. Read the original condition table.
86 protected frames516 cross-model views860 within-model stress pairsInput audit ↗Original model record →
A signal in simulation.
A separate question.
Can the system know when its landing estimate is unreliable? Model VASE · answer so far: a simulation found a recovery signal. The camera results remain separate.
These are different experiments. They have not been joined into one tested model.
How Model VASE fits together →97.6%
V3 landing success under mixed stress.
Frozen simulation, not a camera-to-control result.
Phase 22: transfer without refitting.
10 / 10 gates passed on the final synthetic 32-cell holdout.
- RMSE-surface R²
- 0.8319
- MAE-surface R²
- 0.7744
- Stable-sign accuracy
- 100%
Nothing swept out of view.
Reproduce the work →Controlled occlusion53 to 48 matched frames. Graded degradation, not a universal breakpoint.Original-model replayAuthenticated checkpoint. 24 metric cells. Four byte-identical tables.29 research recordsThe frozen passes, failed gates and later detector studies, together.
Independent data: no dataset admitted. Phase 26 remains a separate prospective gate. The original raw-source occlusion sweep remains INCONCLUSIVE after a failed clean gate; the later Q95 result is a separate saved-artifact replay.
See every phase →View all 13 frozen Phase 12–22 verdicts · 6 PASS / 7 FAIL
Phase 12 → Phase 226 PASS / 7 FAIL
Simulation and benchmark results only. No flight-safety claim, certification claim, or controller-tuning claim.
_(ACD24-0180-035).jpg){kind=link}
