docs/VALIDATION.md

Release validation

The v0.4.1 packaging update adds Hub download-count configuration and new component revisions; it does not change weights or inference operators. The v0.4.0 inference reproductions below remain the evidence for those files.

The v0.3.0 historical audit contains 39 passing tests and a complete generation from the built wheel. Those results describe the historical source and are not a substitute for validation of the v0.4.0 Hub packaging.

Fresh v0.4.0 results

• 53 tests pass on the local tested stack. Added checks cover selective downloads, shared-file reuse, SHA-256 rejection, immutable revisions, relocation without the original transformer, precision/schedule checks, manifest path safety and protection of existing model directories.
• Project lint passes. The built minimax_h3_image_mlx-0.4.0 wheel includes the catalog and Hub loader. Complete generations ran from the installed wheel in /tmp, with assertions rejecting checkout imports and any access to the original transformer safetensors.
• Packed12: a full 2048 × 2048 product image has the same PNG SHA-256 as the recorded exact packed reference. All 16 saved trajectory tensors match in dtype, shape and every element. Receipt.
• Q4: a full 2048 × 2048 product image has the same PNG SHA-256 as the recorded online-Q4 reference. This checks that the exported weights and portable shared components preserve that existing variant's result. Receipt.

These are one-case fresh-process packaging reproductions, not new performance campaigns or broad image-quality evaluations. The published BF16, packed12, Q8, Q4 and shared files were checked against their SHA-256 manifests using Hub file metadata, and the model cards passed the Hub metadata validator. The Q8 export was not included in the fresh full-image packaging reproductions.

Source image-model files were checked against the pinned original verification manifest in provenance/weights-manifest.json. Manifest verification of the remote components does not by itself establish a full remote download and end-to-end inference test on another machine.