What was done
With Claude Opus 4.6 as co-developer, the author mapped AppleNeuralEngine.framework's private classes and in-memory compile path, benchmarked the M4 ANE and ran backpropagation on it without CoreML.
Where the human stepped in
The README says training works but utilisation is low (about 5-9% of peak) and many element-wise ops fall back to CPU.