AMEF: AN EVIDENCE-CENTRIC FRAMEWORK FOR MULTI-SOURCE REASONING AND SAFE ABSTENTION IN SMART PARKING

Abstract

Smart parking systems increasingly combine heterogeneous information sources, including computer vision, human reports, parking officers, and external systems. These sources may differ in reliability, freshness, coverage, uncertainty, and provenance, making direct aggregation insufficient for operational reasoning. This paper presents the Adaptive Multi-source Evidence Framework (AMEF), an evidence-centric framework that separates Observation, Evidence, Estimate, and Knowledge, while assigning distinct responsibilities to Qualification, Evaluation, Reasoning, and Learning. The framework explicitly supports Safe Abstention when available evidence is insufficient for a substantive estimate. AMEF is specialized into a ParkEase domain profile and implemented as a reference system support ing multi-source observation intake, evidence qualification and evaluation, reasoning episodes, Safe Abstention, and traceability. The implementation is evaluated using an external seeded conformance test suite comprising 29 scenario definitions across five scenario families and 290 replay runs. All runs satisfy the specified policy assertions. The final batch produces 160 Estimates, 120 Safe Abstentions, and 10 expected configuration failures. Among the 160 Estimates, 120 agree with the syn thetic ground truth and 40 do not, yielding 75% diagnostic correctness. These results demonstrate behavioral conformance of the evaluated implementation while also exposing limitations in the current reasoning and provenance method. The diagnostic result is limited to the constructed synthetic scenarios and is not interpreted as field accuracy or superiority over alternative fusion methods.