EviRoute: Auditable Fail-Closed Routing for Sensor-Based Edge AI

Abstract

Author's own manuscript submitted to FAIR 2026 (EDAS paper #1571348046): EviRoute: Auditable Fail-Closed Routing for Sensor-Based Edge AI. This PDF is the author's submitted version, not the publisher's version of record. Abstract. We present EviRoute, a contract-based architecture for input integrity, predictive risk, action feasibility, and recovery in sensor-based edge inference. On eight transforms of 6,000 WISDM test windows, its oracle-conditioned policy reduces wrong accepts from 40,016 to 13,320 (66.71%; 67.27% ± 0.74% over five seeds) at 47.45% local coverage under an equal-fault prior. Four forced-abstention faults dominate this pooled result; energy rejection further reduces routing-only wrong accepts by 8.01% (14,480 to 13,320). Accepted faulted outputs are only 41.5% correct. EviRoute has the lowest tested cost for ρ = c_a/c_w ≲ 0.95, when a wrong accept costs over 1.05× an abstention. Three exact-transform detectors achieve 100% recall yet average 0.15 recall and 0.68 residual accepted risk over a 36-condition graded battery. Replay is deterministic; host serialization/hashing and durable append cost 0.15 and 1.87 ms p50. On one Galaxy Note20 5G, LIGHT is 1.7× faster with approximately 76% lower BatteryStats CPU-energy estimate than FULL. This weak-model case study does not establish natural-fault detection, trustworthy accepted predictions, durable-before-release evidence, recovery efficacy, or deployment assurance. The accompanying code and experimental-result reproducibility artifact is archived separately: version DOI 10.5281/zenodo.22151450 (concept DOI 10.5281/zenodo.22151449).