Reading two versions of a document side by side to find what changed is tedious, error-prone, and exactly the kind of matching a model does well. The trick is labelling: call them Document A and Document B, say which is older, and ask for a table with a row per change. The table is what makes it checkable.
Try it
What changed between March and September?
- Did it work? Compare against the list below. The two most-missed changes are the ones that are *removals* — the contact section and the sentence making hirers responsible for their guests. A model that only reports changed numbers has done half the job.
- If it missed a removal: add *"Pay particular attention to sentences in A that do not appear in B."* Models are better at spotting a number that changed than a sentence that vanished.
| Section | March 2026 (A) | September 2026 (B) |
|---|---|---|
| 1. Bookings | Full payment if booked under 14 days out | Full payment if booked under 21 days out |
| 2. Fees | $45 residents / $70 non-residents per hour | $50 / $75 per hour |
| 3. Deposit | $200 | $250 |
| 4. Cancellation | 50% refund at 7–28 days; nothing under 7 days | 50% refund at 14–28 days; nothing under 14 days |
| 5. Hours | Fri/Sat finish 11 pm; hirers responsible for guests | Fri/Sat finish 10.30 pm; guest-conduct sentence removed |
| 7. Keys | Lost-key fee $90 | Lost-key fee $120 |
| 8. Contact | Email and caretaker phone number | Section removed |
Knowledge check
The model's comparison lists six changes and says "all other sections are unchanged". You notice section 8 exists in A but not in B. What does this tell you?