Make real English words appear
one move: stabilize transformations
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.
Residual paths, normalization, and a feed-forward network make attention trainable.
Deliverable: Write the mechanism in one sentence.Compose attention, LayerNorm, residual connections, and an FFN into one block.
Deliverable: Type the smallest runnable implementation.Residual shapes must match and a forward pass must remain finite.
Deliverable: Record the expected output and one edge case.Remove a residual or normalization and observe instability or slower learning.
Deliverable: Change one variable and explain the result.A reusable transformer block with a shape and stability report.
Deliverable: Save the code, output, and a short failure note.Residual shapes must match and a forward pass must remain finite.
Remove a residual or normalization and observe instability or slower learning.
A reusable transformer block with a shape and stability report.
Mapped from the supplied Build Everything PDF, source page 29. The five moves are Deriva’s implementation contract for this project.