Preprocessing belongs in the release boundary
A model’s behavior depends on more than its weight files.
Research note · November 14, 2026Research question
Which preprocessing inputs must be retained to explain a changed inference result?
Method
- Freeze a small model and vary one preprocessing dependency, such as a normalization constant or tokenizer vocabulary. Preserve the input before and after preprocessing for comparison.
- Create a manifest that distinguishes the model artifact, preprocessing code, configuration and data assets. Record each independently so a change can be localized without inventing a single opaque version label.
- Compare results under the original and revised pipelines using the same raw input. Identify whether the difference originated before model execution, during execution or in output interpretation.
Records to retain
- Raw and transformed inputs
- Preprocessing code revision
- Vocabulary or normalization assets
- Model identity and output comparison
Source basis
SLSA provenance distinguishes build inputs and dependencies, while ONNX defines a graph representation with explicit operator semantics. A surrounding application can still perform transformations outside that graph. Those transformations need their own identity boundary.
Interpretation & limits
This proposed experiment does not establish complete preprocessing capture in Vraelis. Equal preprocessing artifacts also do not guarantee equal execution across libraries or hardware.
Next experiments
- Build one normalization-change fixture with an expected outcome.
- Document which pipeline dependencies remain outside the manifest.
References
Technical reference
Implementation details supporting this note, drawn directly from the maintained documentation. These are technical references, not additional papers or new experimental results. Expand a topic to inspect its method, fields and limits.