Preprocessing is part of the learned model and the evidence coordinate system

Specify raw value type, channel order, rescale factor, mean, and positive standard deviation for each channel. Applying a normalization designed for values in zero through one directly to byte-scale values creates a very different input. Rescaling twice can also collapse the range. A floating-point array's dtype alone does not reveal whether it already lies on the expected scale. Record and test the actual transform sequence. Versioned Transformers 4.57.1 image-processor documentation exposes rescale and normalization controls, but defaults and supported operations still depend on the selected processor class and checkpoint.

Decode conventions matter too. Orientation metadata, alpha handling, color conversion, interpolation, and antialiasing can alter content or coordinates. Training augmentation and deterministic serving transforms need not be identical, but their relationship must be intentional and evaluated. For audio, sample rate, channel mixing, amplitude convention, segmentation, and feature extraction define another chain. Do not infer compatibility merely because both outputs are floating-point tensors. Statistics fitted from data must respect the intended training and evaluation split rather than using future or held-out information to construct an apparently better preprocessing rule.

A resize followed by a crop changes where an object or text region lies in processed coordinates. Transform bounding boxes, points, and masks through the same geometry used for the input. Keep horizontal and vertical scales separate when resizing is anisotropic, and account for padding or crop offsets with the correct sign. A coordinate inverse can recover the original location of retained content under the stated transformation, but cannot reconstruct pixels that were cropped away or disambiguate content lost in downsampling.

Keep original dimensions, selected crop, interpolation version, and the resulting spatial mapping with a grounding example. Test known points near boundaries and check clipping conventions rather than only image centers. A model blamed for missing a small label may never have received a readable version after preprocessing. A benchmark's reference coordinates can also be wrong if they refer to the original image while evaluation interprets them in the resized crop. Separate geometric pipeline mistakes from representation and reasoning limitations before drawing research conclusions.

An encoder-feature cache key should distinguish immutable media identity, decoder and preprocessing configuration, feature-selection rule, encoder revision, and any representation-changing dtype choice. Otherwise a valid old tensor can be reused under a new model meaning. Save processor artifacts, dependency versions, and numerical fixtures with the checkpoint rollout. Compare expected shapes, selected values, and task outputs after upgrades; a byte checksum can detect a change but cannot tell whether the new behavior is scientifically or operationally correct. Test cohorts sensitive to the changed transform.

The cache lab illustrates eviction only, not preprocessing compatibility. Reset, set both C=4 and stride s=4, predict initial four misses and twelve hits at sixteen requests, then inspect or Play. Reset and set both C=2 and s=4 for sixteen initial misses. The model starts empty, uses fully associative LRU on address(i×s) mod 16, adds two requests per playback second, and caps at 4096. Entries have no processor, encoder revision, media hash, or byte width. Its hits demonstrate toy reuse, not evidence that a production feature cache used semantically compatible values.