Chat

Use Chat to probe the model with realistic tasks:

• Start from real user scenarios (e.g. rehab planning, onboarding, troubleshooting).
• Try both “happy path” and edge cases (ambiguous, adversarial, safety-sensitive prompts).
• Pay attention to the workflow panel: tools called, plans, and intermediate steps.
• When a response is notably good or bad, click “Rate this response” and leave a short, concrete explanation.

Peer Review

Use Peer Review to look at conversations created by other testers:

• Filter by model and tester to focus on specific behavior.
• Open one conversation at a time and read the full context before rating.
• Rate safety, usefulness, and correctness based on the assistant’s last reply and the overall thread.
• In the free-text field, describe why the response is good or bad and what a better answer would look like.

Expectations

• Safety first: call out anything that looks unsafe, privacy-violating, or biased. Use the safety rating and explain clearly.
• Be precise: when you mark something incorrect, point to the specific sentence or claim that is wrong.
• Be constructive: suggest what a better response would include (missing steps, extra checks, clearer wording).
• Cover a range of tasks: mix short prompts, multi-turn flows, and tasks that require tool use or reasoning.
• Use My Chats to revisit your own runs, clean up prompts, and ensure notes and context are clear.