Answers and explanations · Decide what your dataset can represent

1. Example card: task—route a single message; unit—one message plus supplied context; provenance—handwritten synthetic; split—four development and four visible teaching holdout cases; labels—two invented votes plus recorded adjudication; limitations—no real traffic, independent user sample, or multilingual coverage. It supports an audit of this register's split/label mechanics. It cannot estimate real user success.
2. The changed last record shares refund with development case d3. Unique row IDs identify records, while group IDs represent relationships. The audit returns cross_split_groups: ["refund"] and BLOCK. It could miss paraphrases whose authors assigned different group IDs, or two messages from a common conversation not identified in metadata. A human grouping review remains necessary.
3. label_issues includes d2 with unresolved_label, and the decision blocks. The resolution should cite the task context, discuss the missing referent, record reviewer identities and rationale, and assign a new dataset version if the released label changes. This synthetic exercise does not prove those reviewers are independent real people.
4. A valid proposal: turn 1 names a workshop; turn 2 asks “Can I move it?” With history the referent is known. The unit must include that history, and expected behavior may depend on whether the request is policy advice or an individual action. Mixing this unit silently with context-free messages makes interpretation and denominators unclear.

Self-check: give yourself one point for each answer that includes both a correct mechanism and its limitation. If your card calls the visible teaching holdout an unseen real benchmark, correct that claim before continuing.