LanG 2.0 preprint reports security detection gains with lower energy use

A LanG 2.0 preprint reports that small local AI models improved security investigations while using less energy than Claude Opus 5.5 in its tests. Campaign correlation precision rose from 18.75% to 51.81% against the system's predecessor, while rule acceptance increased from at most 60.00% to 98.33%. The system links evidence to entities, requires human approval for rule changes and keeps a ledger designed to expose tampering. Tests covered eight flow datasets from five networks, real XDR alerts and multistage campaigns. On unseen networks, it reported 73.42% F1, a measure combining detection precision and recall, for at least 650 times less energy than Claude Opus 5.5. These results concern the tested workloads.