Research library
Every report includes its paper, citation, evidence package, and checksums. None are flight results or peer reviewed.
Nomos-TR-2026-04 · Physical Jetson VLM study
We put two open vision-language models on an 8 GB Jetson, showed each one 100 aerial images, and separated visual recognition from mission-policy behavior.
In plain EnglishThe larger model recognized every selected scene but failed half the direct priority requests. Letting simple code apply the rule after recognition restored 100/100 priorities.
100/100scene labels correct
6/12direct priorities correct
2.95×energy vs small model
Nomos-TR-2026-03 · Real telemetry + Jetson
An anomaly detector can flag a change without knowing why it happened. This report tests the missing evidence layer, then challenges four small models to keep their answers tied to proof.
In plain EnglishThe incident record preserved enough proof to explain all 15 controlled ground failures. None of the four small language models was reliable enough to diagnose them alone.
0.870held-out F1
15/15fault evidence retained
0/4models cleared grounding
Nomos-TR-2026-02 · Physical Jetson study
We trained two satellite-image models, ran them on an 8 GB Jetson Orin Nano, and measured the accuracy, speed, and energy tradeoff.
In plain EnglishThe larger model was a little more accurate and used more energy. Neither choice was always best, so an onboard scheduler may have a real decision to make.
+0.926accuracy points gained
26.9%energy increase
2useful model choices
Nomos-TR-2026-01 · Compute-HIL simulation
We tested whether a controller could switch model sizes as power and workload changed, without letting the processing queue grow out of control.
In plain EnglishWhen compute became overloaded and missing work was expensive, switching model size helped. When the system had spare capacity, it made no difference.
12/60cells won against baseline
48/60cells tied
0losses