Accept, Conditional, or Refuse: Risk-Controlled Three-Way Dispositions for One-of-a-Kind Deposition Repair

Metal additive manufacturing (AM) is now used to repair damaged naval components instead of replacing them. However, a repair is made one part at a time, so there is no batch of identical parts from which an acceptance rate could be measured. This makes the accept-or-reject decision hard, because a machine learning (ML) model returns only a probability, and a probability does not tell the engineer how often the accepted parts will actually be wrong. In this study, we develop an acceptance rule that gives one of three answers, accept, conditional, or refuse, and that certifies the error among the accepted parts instead of merely estimating it. The two thresholds are set on a held-out group of earlier parts using a statistical confidence bound, so the promise still holds when only a few hundred parts are available. We compare our rule with three common alternatives on three public AM datasets that cover powder bed builds, melt-pool measurements, and wire-arc beads. Computational results show that the usual practice of accepting whenever the model is more than 50% confident breaks a 10% error target in almost every run, and that thresholds tuned on validation data break it about half of the time, while our rule keeps its promise in most runs. When the evidence is too thin to support the target, our rule accepts nothing and reports that plainly. We believe this gives engineers a defensible reason to ask for more inspection rather than trust a number that cannot be justified.