Scoring the Near Miss: Two Engineering-Derived Hypotheses for AI Incident Escalation Design

Recent work has moved AI incident escalation criteria from proposal to empirical test: Gomez et al. (2026) develop an eight-criterion framework, including a near-miss gate, and stress-test it against ten documented AI incidents. They find that escalation triggers depend on a three-layer infrastructure of data, definitions, and thresholds, and identify design patterns that cause systematic under-detection. Two of these patterns are structurally identical to problems quality engineering has developed mature, quantitative methods to address: the binary, rather than graduated, treatment of near-miss signals, and the absence of a mechanism for aggregating individually sub-threshold harms across categories. This paper asks whether two such methods — FMEA-style severity–likelihood– detectability scoring, and statistical-process-control-style tolerance monitoring — can generate testable design hypotheses for these specific, named, unresolved limitations. A distributed manufacturing defect-detection system (ImmuneLine) motivates the inquiry but does not carry the argument. The paper does not claim these methods transfer directly, and treats the institutional, competitive, and legal differences between a single-owner factory and a multi-actor AI ecosystem as a first-class limitation.