A port authority scores arriving container images against a fixed trained model all day without updating weights. What workload is in use?

A port authority scoring arriving container images against a fixed trained model all day without updating weights is running inference: applying a trained model for prediction without the training data path or an online weight-update loop. Supervised training that fits decision boundaries using labeled container photo batches would still rewrite weights from labeled batches and incur optimizer cost, not the stem’s fixed-model all-day scoring. Fine-tuning that adapts a pretrained backbone to new container defect types updates the backbone weights for those defect types; it is not the fixed, no-update scoring run described in the stem. Offline evaluation that measures model accuracy on labeled test images before deployment is a pre-deployment accuracy check, not the live all-day gate scoring of arriving containers with a pinned checkpoint. Exam caveat: the same municipal cluster may later schedule a training job when labeled container data accumulates; that future training job is separate from today’s inference scoring and must not be confused with it on capacity plans. Operational check: confirm serving loads a pinned model version, that no optimizer or weight-update path runs on the scoring service, that throughput meets vessel-arrival rates, and that operators do not mis-label the job as fine-tuning or offline evaluation work.