Use a small draft model without changing target-model answers
one move: verify before committing
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.
Speculative decoding drafts several tokens cheaply, then accepts the longest target-agreeing prefix and falls back to the target token at the first mismatch.
Deliverable: State the system invariant in one sentence.Compare draft and target token sequences and return accepted prefix length plus the emitted target token.
Deliverable: Type the deterministic core.The emitted sequence preserves target semantics, empty drafts are valid, and a mismatch stops acceptance at the first position.
Deliverable: Record a normal case and a boundary case.Vary draft length, put the mismatch at the beginning or end, and measure accepted-token ratio.
Deliverable: Name the failure signal and its guard.An acceptance trace with speedup assumptions and correctness evidence.
Deliverable: Save the contract, evidence, and handoff note.The emitted sequence preserves target semantics, empty drafts are valid, and a mismatch stops acceptance at the first position.
Vary draft length, put the mismatch at the beginning or end, and measure accepted-token ratio.
An acceptance trace with speedup assumptions and correctness evidence.
Deriva-authored extension beyond the supplied PDF. The five moves are Deriva’s implementation contract for this project.