Factory Shift and End-of-Window Censoring in Rework-Effort Prediction for Turbine Non-Conformance Reports: A Competition Case Study

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We describe our solution to a three-part prediction competition (MAPNA TUGA × Sakkoo, 10th of 63 active teams) on non-conformance reports (NCRs) from a gas-turbine manufacturer: classifying items as short or long rework (C1), assigning them to four effort bins (C2), and regressing booked rework minutes under a clipped (MAPE+WMAPE)/2 metric (C3). The main difficulty was not model capacity but two structural differences between training and test data: the test year comes almost entirely (88%) from a factory system that appears in training only in 2024, and the test window is right-censored — items created in its final month carry essentially no booked work. We show how both were diagnosed, why time-based cross-validation built on the other factory gave misleading guidance, and how a one-change-per-submission protocol with exact reproduction checks made leaderboard feedback attributable. The final pipeline (LightGBM with recency weighting, kNN target features on 4096-dimensional text embeddings, a hierarchical C2 model, a grid classifier with an expected-loss decision for C3, and explicit censoring corrections) reached balanced accuracies of 0.833 (C1) and 0.650 (C2) and a C3 score of 43.01. Code: https://github.com/Rashiin/nullpoint-ncr-rework