Toward Offline Satellite Security Repair: A Preliminary Host Study

Local vulnerability repair could reduce dependenceon ground contact for selected spacecraft applications, but aplausible patch does not establish that it should be activated.We evaluate separate generation and acceptance stages us-ing two quantized Qwen2.5-Coder models on three syntheticC predicates, with three seeds and up to two attempts pertrial. The 0.5B model produces no evaluator-valid repair innine trials; the 1.5B model produces two in nine. Across 34attempts, development-only and additional-acceptance policieschoose identical candidates, providing no evidence of an extra-gate advantage. Inference denies network syscalls, and network-filtered compilation/validation replay reproduces all outcomes. Aseparate measured host workload records 26-27 deadline misseswith immediate artificial validation and zero with phase-awaredeferral, while approximately doubling validation completiontime. That workload excludes inference and actual patch valida-tion. These findings provide preliminary reproducible evidenceand expose evaluation gaps. They establish neither a new repairalgorithm nor satellite safety. Independently sourced defects,protected build/supervisor isolation, and integrated target-devicemeasurements remain necessary.