From Model Disagreement to Reusable Evidence: Human-in-the-Loop Visual Adjudication for Multi-Model Timing-Diagram Parsing

Abstract

Timing diagrams combine waveform structures, textual annotations, and protocol assumptions, making their semantics difficult for multimodal models to interpret consistently. Models may disagree on sampling conditions, signal roles, and validity intervals, while conventional aggregation hides the evidence behind these differences. We present a human-in-the-loop visual adjudication workflow that structures model outputs as comparable claims, reveals competing interpretations and their evidence, and supports expert review. Adjudicated cases are further summarized as reusable accepted and rejected reasoning patterns. A case study with two cryptographic hardware specifications illustrates claim-level disagreement resolution and selective cross-document experience reuse.