Development of “Weave Back”: An Integrated System for Human Error Prevention

Abstract

Near-miss reports capture frontline signals that can prevent accidents, yet many organizations stop at recording and sharing them, and struggle to translate narratives into implementable actions. This study proposes WeaveChain, an integrated framework for converting near-miss narratives into actionable knowledge through three stages—Factors, Mode, and Action—and focuses on the design and evaluation of its action-generation module, WeaveBack. WeaveBack takes as input Performance Shaping Factors (PSFs) extracted upstream and one of 20 human-error modes and generates candidate countermeasures anchored in PSF–mode combinations. To prevent generic, context-insensitive outputs and vigilance-only advice, WeaveBack enforces a structured protocol that crosses ten intervention domains (L/T/H/P/E/C/R/M/O/X) with two intents (human-error prevention; frontline improvement), thereby ensuring 20 comparable candidates per case. We further curated two reference datasets (software-oriented and hardware-oriented, 100 factors each) and refined them through an iterative improvement loop that alternates KPI scoring (11 KPIs, 0–10) and human revision. Rubric-based evaluation conducted by the authors showed that, across all factors in both datasets, the mean scores for KPI1–KPI5 exceeded the threshold (≥5), while KPIs related to external value remained relatively lower. These results suggest that the proposed design can operationalize near-miss learning for small teams by stabilizing a structured candidate set that supports comparison, selection, implementation, and continuous refinement.

Keywords: Near-miss Reports, Human Error Prevention, human–AI Collaboration, Countermeasure Generation, KPI-based Evaluation

DOI: 10.54941/ahfe1007738