Read the operating agreement, data sourcing, and artifact records. For hosted storage use managed guidance.
For bundled local tools and exact commands, read toolkit setup. Run the workspace installer for this skill; it preserves existing customized files. For the managed route, use concrete service operations after confirming the connected schema.
Reuse and curate
Inspect existing cases and ask only what cannot be derived: whether they still match traffic, who verified references, and what decisions the eval must support. Preserve useful cases. Distinguish inputs from reference answers, rubrics, and baseline outputs; expected does not always mean a single correct sentence.
For every case retain source, permitted retention, failure/category tags, conversation or source group, necessary attachments/context, and reference origin. Capture the environment/fixture necessary to replay an agent. Keep source records unchanged.
Use confirmed failures as regression examples, but also include normal and negative cases. A challenge suite and a representative benchmark answer different questions; label them separately. Generate synthetic cases only to fill specific gaps and have a knowledgeable reviewer verify realism.
Protect comparisons
Deduplicate and group related conversations, documents, customers where appropriate, and paraphrase families before splitting. Split by group so related examples do not cross boundaries. Stratify by relevant task/category where enough groups exist; do not select development cases just because their baseline scores are low.
Use development for diagnosis, validation for candidate selection, and a final test slice only once when enough data exists. With a small dataset, reserve a credible holdout if possible; otherwise report exploratory results and collect new confirmation cases. Freeze the split before optimization. Never redraw because scores are inconvenient.
Keep answers in grader-only storage outside an acting model's filesystem and tool reach. Remove public benchmark solution access when relevant. Do not rely on “do not read answers” instructions.
Review and commit
Show actual cases with inputs, attachments, reference origins, and category counts in the chosen review surface. Confirm relevance and retention before committing production-derived data. Store source identifiers or sanitized substitutes when full records cannot be retained.
Record immutable dataset version/hash, sample purpose, grouping rule, and split IDs. Use the optional eval-run artifact helper to validate exported records and create a reproducible group split, or the existing runner's equivalent. Its simple splitter is not a statistical sampling service; inspect category coverage before locking the manifest.
Deliver the runnable input set, protected reference set, review status, and any coverage gaps. Do not mark unreviewed synthetic references as expert-approved.