Preview Based View Selection and Cost Aware Stopping for Budgeted Inspection

Abstract

Budgeted inspection combines evidence acquisition with a decision about when to stop. This manuscript proposes AVID, a partially observed decision architecture, and evaluates a simpler committee-based realization on 1,797 handwritten-digit images. Classes {3,8,9} represent a proxy defect label, and six deterministic image transforms represent views. Across twenty repeated splits, the preview-based selector reports single-view AUC 0.952 versus 0.864 for random order. Its candidate scoring uses actual unacquired view features, giving it information unavailable to the non-preview baselines. Cost-aware stopping uses 1.65 acquired views per decision versus 3.90 for a posterior-odds threshold baseline, but has lower reported net utility: −0.058 versus +0.036 under correct=+1, wrong=−5, and acquisition cost 0.15. Always acquiring six views has accuracy 0.9392 versus 0.8650 with stopping. The result therefore documents a preview-assisted selection and accuracy–cost trade-off on a digit proxy, not validated industrial acquisition efficiency. Reinforcement learning, attention-based aggregation, asymmetric error costs, and formal Wald guarantees are not executed by this realization.