Tyler Barr, Embry-Riddle Aeronautical University; Kyah Adams, Embry-Riddle Aeronautical University; Di Wu, Embry- Riddle Aeronautical University
https://doi.org/10.64861/VQVF2441
Keywords: orbital debris, Space Situational Awareness, Monte Carlo simulation, debris environment modeling, policy compliance, quantitative governance, explainable AI, large language models, Code World Model, decision support
Abstract:
Orbital-debris governance requires traceability among policy requirements, bounded model settings, and quantified environmental outcomes. Catalyst links approved sources, two-line element (TLE)-derived scenarios, Embry-Riddle’s Python physical-model layer (E-MOCAT), and a Code World Model (CWM), while deterministic services and analyst controls retain numerical, evidentiary, and execution authority.
A 500-run object-resolved E-MOCAT campaign tested five ten-year conditions in 100 matched-seed blocks sharing a 19,085-object initial state. The study-defined business-as-usual (BAU) reference set post-mission disposal (PMD) to 0.50, annual rocket-body explosion probability to 0.20, and replayed launches. Integrated mitigation (PMD = 1.00; explosion probability = 0) produced the lowest tested means for collisions, generated fragment objects, and final population: 55.6%, 94.8%, and 55.6% below BAU, respectively, although final population remained 2.096 times the initial population.
A separate 800-case study compared Reference CWM (adapter unloaded) with Catalyst-Adapted CWM (adapter loaded) using Meta CWM-SFT, a 32-billion-parameter supervised-fine-tuned checkpoint, and produced 1,600 paired responses. Deterministic checks evaluated objective criteria, while human-in-the-loop validation applied five predefined criteria to cases requiring semantic judgment: evidence grounding, internal consistency, task relevance, appropriate limitation to model-visible evidence, and absence of material error.
The evaluation yielded 785/800 Successful Response labels (98.13%) for Catalyst-Adapted CWM and 660/800 (82.50%) for Reference CWM. In a distinct 720-response explainable artificial intelligence (XAI) study, Catalyst-Adapted CWM achieved the highest observed mean content score and faithful-complete rate among six tested configurations spanning CWM, Gemma, and Nemotron.
Together, these studies connect declared governance settings, quantified E-MOCAT outcomes, and evaluated CWM responses within a traceable workflow.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications