Stage-Wise Threshold Calibration for Early Defect Warning in Multi-Stage Manufacturing: A Cost-Sensitive Single-Line Automotive Case Study

In multi-stage manufacturing, conformity is confirmed only at the final test, after the cost has been incurred. On an automotive assembly line (16,500 workpieces), a regression model per checkpoint predicts the final measurement; a threshold turns it into a Go/NoGo decision. On the holdout, the models ranked the defects (area under the curve, AUC, up to 0.81), yet at the inherited final-station limit they raised no alarm and missed all 15. The threshold was therefore calibrated per checkpoint on an earlier period under a cost-weighted objective with an explicit defect rate, and evaluated on a later holdout. Carried over as a value, it failed: 31% of production flagged during calibration became 89% on the holdout. Carried over as an alarm budget read off recent predictions and set from inspection capacity, it recovered 14 of 15 defects at a 36% false-alarm rate, below the cost of all nine baselines on this holdout; the cost objective at the defect rate the holdout later showed, a retrospective sensitivity scenario, gave the same count. Under 20 other seeds, the advantage over the closest rules did not hold, and a flag on the ramped-up variant, available beforehand, would have been cheaper. The evidence is a retrospective, single-line case study with 15 holdout defects; the rule is a candidate for shadow-mode evaluation, not a validated deployment policy.