Understand the issue, then continue.

Which practice should I start with?

The library offers starting points in image annotation, text annotation and AI evaluation. Browse 80 selectable tasks by type, difficulty and keyword.

Read the current objective and guidelines when opening a task. “Examples & self-check” explains the controls using independent examples, not answers to this task. Labels, scope and criteria always follow this task's rules.

How do I know my answer is saved?

The workbench distinguishes unsaved changes, saving and draft saved to server. Only the last confirms the current draft is on the server. Saving a draft does not submit it or reveal the answer reference.

If the network drops, keep the page open and retry saving; export the current answer if needed. On conflicts, save a local copy before deciding whether to load the server version, which replaces this page's content.

When submission is unconfirmed, use Retry submission or Check server status on the original page. Do not create another identical submission just to retry.

How do feedback, revisions and new material differ?

Final submission locks the current answer and reveals the reference. Feedback points to regions, labels, evidence or rating dimensions and explains applicable rules and correction directions.

Choose Revise from feedback to create a linked attempt while keeping the original. After meeting structured criteria, try new material of the same type. Its rules may differ, so reread them.

New materials are assigned from source exposures known to the server, not a guarantee that a person has never seen them elsewhere or on another device. When stock runs out, continue other practice; exposed material is not disguised as new.

Unsure, or think the reference is wrong?

Choose “I have a judgment needing human confirmation” and explain known facts, missing information and specific concerns. Submission retains an ID and reasons and enters review. Waiting does not automatically make it pass.

All materials await project human acceptance. Image tolerances and text/answer judgments still require checking; free-text explanations receive no automatic semantic approval.