C.17:4.3 - Surprise: unexpected under which model?

Surprise reports how improbable one declared sample of the bearer is under one generative model. For a discrete probability, a common raw result is -log p(sample) in bits or nats. State the modeled sample unit and encoding and how bearer size is handled. Compare bearers only under a justified common extent, a declared per-unit or code-length normalization, or another calibrated rule suited to the model. For a continuous model, identify the measure as well as the representation; a density value alone is not representation-independent. Otherwise keep the raw model result within its exact basis and do not treat it as a comparable Surprise coordinate. Also identify the model episteme and edition, training basis, preprocessing, fit and out-of-distribution checks, calibration, refresh condition, and limits.

Novelty and Surprise answer different questions. A bearer may be unlike the selected corpus yet unsurprising under a broad model, or close to known examples yet surprising under a narrow model. Keep both results visible when both matter.