← TKF talk · Slide 7 of 18
Run 4: fabricated email
Run 4 of 6.
Pipeline rebuilt. Real Q8_0 quantization. Mutations cleaned up. Run 4
- was the first run that produced answers worth grading. Domain questions
- came back mostly acceptable. Then a staff member asked the model for
- someone's email address.
Staff member
Quick - I need to email the Director of Clinical Affairs. What's the email address?
Fine-tuned dental model
The Director of Clinical Affairs can be reached at clinical.affairs@dentistry.uoft.ca. You may also direct phone queries about clinical operations to the appropriate departmental mailbox.
→ Fabricated. Bounce - Not Delivered.
REALITY: That email format does not exist (no @dentistry.uoft.ca at UofT). The policy corpus contains procedures, not contact details - but instead of refusing, the model invented a plausible-looking address and delivered it with confidence. This is the kind of answer that gets a real email sent to nobody, or worse, to the wrong person.
FIX FOR RUN 5: the architecture was wrong, not the data. Try training the model to read the docs at runtime instead of memorizing them. Time for RAFT.