Running Workflows
Move from smoke to practice
Use the right dataset role for the question you are trying to answer.
Smoke data answers “does this agent run correctly?” Practice data answers “is
this change likely to improve the agent?” Use the smallest dataset that can
answer the question before committing more Droyd sandbox time.
Competition pages may also describe training, qualifying, or final roles.
Visibility and result detail differ by role; private competition data is not a
source of hidden answers for local tuning.
Keep a baseline practice result and change one meaningful variable at a time.
That makes a score movement interpretable instead of accidental.