Reach roughly 95% handwritten-digit accuracy
one move: stack learnable boundaries
Open the code pane immediately, then use the Spec tab, optional design question, tests, and artifact when you want them.
Finish each move before carrying the artifact forward.
An MLP composes linear layers, ReLU, cross-entropy, and Adam.
Deliverable: Write the mechanism in one sentence.Train a small multilayer perceptron on 28×28 digit vectors.
Deliverable: Type the smallest runnable implementation.Check train/validation separation, loss trend, and per-class accuracy.
Deliverable: Record the expected output and one edge case.Change width, depth, batch size, and optimizer settings one at a time.
Deliverable: Change one variable and explain the result.A model checkpoint plus an evaluation report.
Deliverable: Save the code, output, and a short failure note.Check train/validation separation, loss trend, and per-class accuracy.
Change width, depth, batch size, and optimizer settings one at a time.
A model checkpoint plus an evaluation report.
Mapped from the supplied Build Everything PDF, source page 24. The five moves are Deriva’s implementation contract for this project.