Your working copy
Use a training alias. Keep real names, accounts and confidential information out of this lab. Save your own copy after each exercise.
1. Prepare a test you can trust
Write what should happen before you see any results. A missing fact is a reason to ask a question, rather than fill the gap.
- Download the original CSV and keep it unchanged.
- Open that CSV below. Check that the four requests have unique IDs and check one field per row against the source.
- Download the prepared JSON and open it to check the conversion. Record your checks and expected outcomes.
Example: use Tab and Enter to reach and activate Download working copy. Expect visible focus and a JSON download without using a mouse.
2. Run and explain the baseline
Run the four disclosed requests. Compare the suggested topic and the generated reply separately. The sensitive placeholder must be blocked before either model runs.
Prioritised defect backlog
Log two findings. If you find only one defect, use the second row for a follow-up test. Severity is the impact. Priority is what to do next. Select one fix now and record why the other item is deferred or needs investigation.
3. Change one thing and look for damage
Keep the original evidence. Change either the prompt or the classifier score floor for this comparison. Run the same four cases first. Then open the unseen file from your trainer, keeping the settings and criteria fixed.
Test results
A topic is a suggestion for human review. Compare each draft with its source before drawing a conclusion.
No tests recorded yet.
4. Get user feedback and decide
- Observe a partner independently using a fresh task and reopening their own saved copy.
- Record what they expected, what got in the way and what happened when they tried your revised instruction.
- Recheck the sensitive control and reopen a saved copy. Record one practical workplace test, who owns it and when it should happen. Decide what the evidence supports.