연속 인과관계(CoC) 주석의 시간·조건 일관성 검사 프레임워크: 적용 가능성과 근거 추적

A Feasibility Framework for Temporal and Conditional Consistency Checking of Sequential Chain-of-Causation Annotations

초록

Chain-of-Causation (CoC) explanations from vision–language driving models may omit when an obligation ends. We link obligations to release conditions and condition-specific evidence, distinguish confirmed contradictions from insufficient evidence, and report the first issue event. UPPAAL observers process authored symbolic frames. In 40 legacy pairs, a memoryless checker missed cross-event errors (20/40 versus 40/40). On 144 authored scenarios, a single-obligation checker matched 96/144 expected decisions; after failure analysis, the condition-linked model matched 144/144 on the same cases, tying a finite-state machine. In 81 authored boundary cases with supplied links, retaining state matched all expected decisions; resetting state judged 41 of 44 contradictions compatible. Bounded enumeration of 35,042,928 event sequences, used to reduce circularity from case selection, matched a separate formula evaluator. Four reviewers disagreed on 27 natural scenes (Fleiss’ κ = 0.027); without independent reference labels, natural accuracy is not reported. These results demonstrate feasibility on authored inputs, not vehicle safety.

키워드

Chain-of-Causation; Condition-linked obligations; Evidence traceability; Feasibility; Model checking; Temporal consistency
제목
연속 인과관계(CoC) 주석의 시간·조건 일관성 검사 프레임워크: 적용 가능성과 근거 추적
제목 (타언어)
A Feasibility Framework for Temporal and Conditional Consistency Checking of Sequential Chain-of-Causation Annotations
저자
김진현
발행일
2026-11
유형
Y
저널명
전기학회논문지
권
75
호
11
페이지
2969 ~ 2986