Skip to content

Verification checklist

Use this whenever you accept code you did not write: from an AI assistant, from a website, from a colleague.

Print it. Keep it beside you.

Code that runs is not the same as code that is right. Most errors in AI-assisted work are silent. The script finishes, the number looks reasonable, and the number is wrong.

The four checks

1. Does it run?

Run it. If it fails, read the traceback from the bottom up.

2. Does it give the right answer on a case you already know?

Take five rows of the data. Work out the answer yourself, with a pen or a spreadsheet. Then run the code on those five rows. Compare.

Most people skip this check. It is the one that finds real errors.

Work out the answer separately from the code. Comparing the code with itself proves nothing.

If you cannot build a case where you know the answer, stop. You do not understand the problem well enough yet to judge code that claims to solve it.

3. What does it do with the awkward cases?

Test each one that applies to your data:

  • [ ] A missing value. Is it dropped, or treated as zero? Those give different answers.
  • [ ] A duplicate row. Is it counted twice?
  • [ ] A zero. Does anything divide by it?
  • [ ] An empty input. Does the code fail clearly, or return something that looks fine?
  • [ ] The full dataset, not the sample. Does it still finish? Does the answer still make sense?

4. Can you explain every line?

Read the code line by line. Say what each line does.

A line you cannot explain is a line you cannot defend. You can always ask what a line does. Keep asking until you understand it.

Warning signs

  • Rows disappearing. Compare the row count before and after every filter, merge and cleaning step. If the count changed, know why.
  • Numbers stored as text. Values read from a file are text until something converts them. Text sorts alphabetically, so "10:00" comes before "9:00".
  • Invented column names. Code that uses columns your data does not have. Worse: columns it does have, which mean something else.
  • Hidden assumptions. The assistant does not know that your buses run every 8 minutes, or that your data crosses midnight. It will assume something and not tell you.
  • The wrong statistic. A mean where a median was needed. An average of averages. These give numbers of about the right size, which makes them hard to notice.

How to work

Small steps. Run the code after each one. Save a copy whenever something works.

The pattern to avoid: ask for 200 lines, receive 200 lines, and have no idea which one is wrong. Ask for one function. Check it. Then ask for the next.

What you keep

For the week 3 task, and for every chapter of the project, keep three things together:

  1. The code.
  2. The prompts you used.
  3. Evidence that the code is correct: your hand-worked case, and what happened on each awkward case.

The evidence is the part with lasting value. A partial solution with proof is worth more than a complete solution without it.