MNIST handwritten digit classification
Prepared from the supplied project saved 24 September 2026.

Contents: original .nd, native chart HTML and architecture PNG, saved sample-role manifest, and this README. The project contains trained parameters, but does not embed the source image/text dataset or the standalone Neural Engine runtime.

Open the .nd in a compatible Neural Designer version. Relink the dataset path to your local source, keeping its file order and preprocessing. For text examples the supplied TXT is included under data/. For image examples acquire the source data separately and reproduce the supplied selection; the image manifest lists its files and assigned roles.

Source: https://www.tensorflow.org/datasets/catalog/mnist

To deploy, use Neural Designer to export a deployment package; keep the generated Python wrapper next to its engine/ and model/ directories. The standalone engine is not included in this archive.

Corrected evaluation report, 24 September 2026. The saved matrices and class metrics now match inference on the saved testing roles. Repeated reloads with seeds 42 and 1927 reproduce the same matrices. Use a Neural Designer build that restores saved image sample roles when loading the project. No weights or sample roles were changed in this revision.
