Petit test
Guess first, no penalty. An AI's pandas script does a left merge of 1,000 orders with a customer table. The result has 1,450 rows. What happened?
pandas is Python's library for working with tables, and a merge joins two tables on a shared column, like a SQL join. AI scripts merge freely and rarely check the result. The same fan-out that inflates SQL totals happens here.
pandas has two checks built into merge itself:
- validate checks the keys are unique where you expect. "many_to_one" means many orders, one row per customer. If the customer table has repeats, pandas raises an error instead of quietly repeating rows.
- indicator=True adds a _merge column saying where each row came from: "both", "left_only" or "right_only". left_only orders found no customer.
In your prompt (the message you send to the AI), ask for both, plus row counts before and after.