Epistemic Gatekeeper: Training-Free One-Shot Gating for Open-World Industrial Anomaly Detection

Industrial anomaly detection (AD) models are typically built for a fixed inspection task and assume that test inputs belong to that task. In open-world deployment, however, contamination from unseen inspection tasks can enter the test stream. Existing AD models still assign anomaly scores to such out-of-scope inputs, producing scores that mix defect evidence with task mismatch. We cast this as a pre-score gating problem and propose Epistemic Gatekeeper, a training-free front-end gate for open-world industrial AD. The gate constructs an inspection-scope representation from a single normal reference image per category and rejects unseen-task inputs before the AD model is invoked. Across MVTec and VisA, the proposed method filters both near and far out-of-distribution (OOD) inputs in the 1-shot setting while preserving same-task anomalies and keeping ID-FDR near zero.