# Inspectable neural and image workflows

Run `python deep_learning_cases.py` from the repository root. The lab uses the repository's main Python environment and original synthetic tabular, image and sequence fixtures. It deliberately stays small enough to expose equations, gradients, partitions and artifacts. The image classifier is a genuine trainable one-convolution-layer NumPy network, not a production architecture or external dataset benchmark.

The verifier passes twenty checks covering baselines, numerical gradients, activation failure modes, learning rates, batch sizes, dropout/weight decay, early stopping, clean image groups, augmentation semantics, transfer initialization, duplicate screening, class slices, imbalance, saliency perturbation, padding masks, sequence memory, artifact reload, latency protocol, quantization and negative-result reporting. Full outputs are saved in `deep-learning-verification.json`.

No result is evidence about learners, NeuraPath, a client or a deployed model. Runtime measurements describe the machine executing the lab and may change between runs.