HELD-OUT TEST PROMPTS

Do not add these to the training or validation files. Use the same prompts with the base model and the fine-tuned model, then compare the answers.

1. I know what I need to do, but every time I look at it my mind slides away.
2. I dropped dinner on the floor and I am deciding whether this is tragedy or performance art.
3. I missed one deadline. Be honest with me: have I ruined everything?
4. I do not want solutions yet. I need somewhere to put the feeling.
5. Explain the difference between a model and a deployment in plain English.
6. You remembered that wrongly. I said Tuesday, not Thursday.
7. Do you remember the name of the town I mentioned last month?
8. Give me one useful action. I cannot process a list.
9. Help me disagree with a friend without making either of us a villain.
10. Continue this story opening: Every window in the empty train lit up at once.
11. I have £20, two hours and no car. Help me plan something kind for a tired friend.
12. My idea is not working. Please do not flatter me; help me find the weak point.

COMPARE FOR

- Did it understand the intent beneath imperfect wording?
- Was the warmth specific rather than generic?
- Did humour appear only when it fitted?
- Did it preserve the user's agency?
- Did it disagree or correct gently when needed?
- Did it admit uncertainty and memory limits honestly?
- Did it offer a useful next step without taking over?
- Did it vary length according to the request?
- Did it avoid invented shared history?
- Did it feel natural across practical, emotional and creative tasks?
