Quantifying Threshold Sensitivity in Vision-Based Ergonomic Assessment: Decision Stability of RULA and REBA in a Batik Workshop

Automated pose-estimation pipelines usually report one RULA or REBA score per frame, without indicating whether that decision would survive small errors in the estimated joint angles. Both instruments convert continuous angles into ordinal categories through fixed thresholds and non-additive lookup tables, so the same angular error may leave one decision unchanged while reversing another. Earlier vision-based studies report aggregate agreement with reference systems or experts; none attaches a robustness statement to the individual decision. This study proposes a reproducible protocol that does so. It combines an observability audit, conditional lower and upper bounds for decisions whose inputs cannot all be seen, controlled angle-level perturbation at magnitudes anchored to published single-camera angle errors (σ = 2, 5, 8, and 12°), and a set of metrics in which baseline decision retention (BDR) measures whether the reported decision survives and the Ergonomic Decision Stability Index (EDSI) measures the probability of the most frequent category after perturbation. The protocol was instantiated on one batik worker (four photographs, four videos). At the level of the RULA action level, the four photographed postures retained their lower-bound decision in at least 93% of replications at σ = 5°, although two of them lay only 0.1° and 0.2° from a component threshold, where distance-based predictors anticipated retention of only 21–52%. The same posture that was stable under RULA retained its REBA risk level in only 49% of replications. Unobserved inputs left the final action level undetermined between its bounds for every posture. Across 534 video frames, component-level stability fell from 0.88 at 2° to 0.40 at 12°. The results demonstrate the method on one case; they do not support population inference or a validated triage rule.