Probabilistic Static Analyses Emit Mutually Inconsistent Alarms — Conflict Extraction Finds Sets That Cannot All Be True

arXiv 2608.10755·low signalXBlueskyLinkedInCopy link

Probabilistic Datalog lets static analyses such as pointer analysis, data race detection and side-channel analysis rank alarms by likelihood, but introduces a failure mode absent from deterministic analysis: marginal probabilities do not guarantee joint satisfiability, so the output can contain alarms that are individually plausible yet mutually inconsistent. Developers then waste effort investigating alarm combinations that cannot co-occur. The paper contributes conflict extraction to surface those inconsistent sets — relevant to anyone building or triaging LLM-assisted static analysis pipelines that rank findings by confidence score.