Fresh-data testing

Build-along · ~10 min at your own keyboard

A model asked to test its own work generates test cases that flatter its own assumptions. The fix: test with data it has never seen, like a real closed listing from last year or a comp set you know is weird. Fresh data finds the failures polite tests never will.

Your starter: Monthly export refresh

Prompt

Make it yours: Name the export you pull and where last month's sits, set what counts as a change worth a line, and keep the list of things it must never do.

Compare this month's general-ledger export at [path] with last month's at [path], for [entity code].

Report:
- New accounts: number and name, both exactly as the export has them.
- Accounts that went to zero or came off zero.
- Any account whose change is above [threshold], with prior and current from the export.
- Any line that looks like it carries a person's name or an account number: say so and stop; nothing identifiable goes in.

Do not explain a change. Do not propose an entry. If a figure is not in either export, write [ASK THE CLIENT: the source] rather than a number.

The coach checks that what you paste back is your own work. The starter pasted back unchanged won't pass.

Do this

0/4

Share what happened

Paste what fresh inputs you used, what broke, and the category of failure.