Evidence-Sufficiency Curves: How Much of Each Evidence Type Is Needed for a Deposition Acceptance Decision

An acceptance decision for a repaired part is built from several kinds of evidence, such as the machine settings, the description of the alloy, and measurements taken on the part itself. Each kind costs time and money, so an engineer needs to know how much of each is enough before the work begins. However, studies in additive manufacturing (AM) usually report one accuracy value at one data size, which hides the fact that richer evidence may only pay off once far more parts are available. In this study, we measure how the prediction error falls as more parts and richer evidence are added, using three public AM datasets that cover powder bed builds, melt-pool measurements, and wire-arc beads. For each kind of evidence we report the number of parts needed before the error stops improving, and we test whether a curve fitted on a small sample can predict the error at a larger one. For the wire-arc beads we also vary how many measurements are taken on each bead. Computational results show that richer evidence can lower the error by more than 40%, but most of that benefit appears only after the training set grows to hundreds of articles. A forecast made from a small sample is reliable on two of the three tasks and badly wrong on the third. Taking twelve measurements per bead instead of seven cuts the share of faulty beads wrongly passed by about five times. These curves provide a simple way for researchers and practitioners to plan what evidence to collect first.