Reject a training row before it poisons evaluation
one move: make data admissible
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.
Reliable models begin with a versioned row contract: identity, required fields, label policy, and a reserved evaluation set.
Deliverable: State the invariant in one sentence.Validate raw rows, reject malformed records with one stable reason, and refuse IDs reserved for the holdout set.
Deliverable: Type the smallest deterministic implementation.Accepted IDs, rejection reasons, and reserved-ID leaks must be deterministic and complete.
Deliverable: Record a normal case and a boundary case.Add a duplicate, an empty label, and a reserved ID; check that no row disappears silently.
Deliverable: Name the failure signal and the guard you would add.A dataset validation report that can gate every later training run.
Deliverable: Save the contract, evidence, and handoff note.Accepted IDs, rejection reasons, and reserved-ID leaks must be deterministic and complete.
Add a duplicate, an empty label, and a reserved ID; check that no row disappears silently.
A dataset validation report that can gate every later training run.
Deriva-authored extension beyond the supplied PDF. The five moves are Deriva’s implementation contract for this project.