Weight identity does not establish serving parity

A verified checkpoint hash and successful strict state loading establish tensor identity and compatibility, not parity with a training platform's deployed predictions. Preprocessing, resize policy, class namespaces, backend artifacts and postprocessing can still differ. When investigating a discrepancy, run the vendor-supported inference path on the same authorized examples, retain its native labels and predictions, and compare outcomes before adjusting confidence thresholds or claiming the model is fixed.

Do not infer a foreground label index solely from classifier-head shape. A raw training export and a packaged deployment can expose different class namespaces. Converting native coordinates for an application API is separate from changing native model semantics.

Applies when: a custom model loader disagrees with the training platform's deployed behavior. Keywords: model checkpoint deployment parity preprocessing class mapping inference runtime export