Independent evidence exposed persistent bias.
V3 reduced mixed unsafe touchdowns from 84.2% to 2.4%; under occlusion, from 34.6% to 1.4%.
Read the frozen results ↗
AEGISLAND · RESEARCH ARCHITECTURE
Model VASE
Vision · Auxiliary evidence · State reliability · Escalation
Can it recognize a bad estimate before the landing decision?
Answer so far
Yes in planar simulation. Not yet through real camera input.
Independent estimates and selective fallback reduced unsafe outcomes in frozen simulation. The camera and flight results do not yet demonstrate the complete system end to end.
One system. Clear evidence boundaries.
Only the V3 simulation path is connected in the executable supervisor today.
- 01
Vision
Produces landing estimates and detections. Synthetic-image and KIOS detector results stay attached to their own tests.
IMAGE TRACK · SEPARATE - 02
Auxiliary evidence
A noisy, lower-rate reference estimate exposes persistent visual bias in the V3 simulator.
V3 PATH · CONNECTED - 03
State reliability
Disagreement, freshness, confidence, and calibrated uncertainty describe what may be trusted.
PHASE 12 · NOT YET CONNECTED - 04
Escalation
The frozen V3 supervisor returns
V3 PATH · CONNECTEDPROCEED,HOLD, orABORTto the simulator.
Model VASE is an integrated research architecture, not a newly trained checkpoint. KIOS boxes and Phase 12 intervals are not silently fed into the V3 supervisor.
Evidence, by track.
Switch views. Each result stays with its original model, data, and protocol.
2.4%V3 mixed-stress unsafe touchdowns
10,000 simulated episodes
10,000 simulated episodes
1.4%V3 occlusion-stress unsafe touchdowns
Baseline: 34.6%
Baseline: 34.6%
Selective fallback helped under stress.
Mixed-condition success was 99%. Low-light timeouts were 3%; mixed lateral bad-estimate rejection remained weak.
Read the held-out study ↗Calibration improved, but is not in the controller.
Phase 10R and 11 missed protected gates. Phase 12 passed its frozen gates: 95.53% lateral coverage and a 2.230× tail-width ratio.
Read the Phase 12 report ↗Transfer evidence stops at simulation.
Phase 8 recorded a PX4/Gazebo diagnostic mismatch. Phase 22 passed 10 locked simulation gates; neither establishes flight performance.
Read the Phase 22 record ↗+3.4 ppPhase 24 six-condition mAP50
Published KIOS aggregates
Published KIOS aggregates
−13.3 ppOcclusion + mixed mAP50
Phase 23 detector regression
Phase 23 detector regression
Average gain hid harder-condition losses.
On the same 86 KIOS frames, the six-condition mAP50 average rose while occlusion and mixed results fell.
Read the aggregate audit ↗Frame transitions measured across 516 paired views.
On 86 frames × 6 conditions (516 paired views), Phase 25 recorded 49 recovered, 65 regressed, 223 both-pass, and 179 both-fail transitions. A fresh replay matched all 24 published Phase 23 metric cells exactly and all four prediction tables byte-for-byte. These within-KIOS comparisons do not establish a causal training mechanism or independent generalization.
Inspect the Failure Atlas ↗Next: test whether the failure pattern transfers.
The admission method is hash-locked, but no independent dataset is admitted. Both KIOS sessions and all 422 real frames are excluded; a screening pass still requires documented provenance, ontology, annotation and rights review.
Read the Phase 26 protocol ↗Next decisive evidence
Admit independent data. Evaluate offline signals.
The Phase 23 checkpoint is recovered and paired. The next test is to admit session-screened independent imagery under Phase 26's frozen protocol and evaluate whether Phase 25 reliability signals can inform the VASE supervisory layer offline.