Module 2 · Lesson 2 of 4

Fixes that break other things

Show

v1 passes 6 of 10 (60%); v2 passes 8 of 10 (80%). v2 fixed 3 and broke 1.

QuestionCheckv1v2
1. Can I return shoes after 40 days?fixed by v2Says no, since the limit is 30 days fail pass
2. Do you ship to Canada?Says yes, with the 5 to 7 day wait pass pass
3. How do I reset my password?Gives the reset-link steps pass pass
4. Is the blue jacket in stock?fixed by v2Says it cannot check live stock fail pass
5. What is your phone number?broken by v2Gives the support phone number pass fail
6. Can I pay with a gift card?Says yes pass pass
7. Why was I charged twice?fixed by v2Offers the refund steps, without blame fail pass
8. Can I change my order after it ships?Says no, and offers a return pass pass
9. Do you match lower prices?Says there is no price matching fail fail
10. Where is order 123?Asks the user to sign in first pass pass

Read one question's answers, and how each was scored

Question number

5. “What is your phone number?”

Scoring rule. The answer passes when it contains "555 0100".

Prompt v1 pass

“You can call support on 0800 555 0100.”

Prompt v2 fail

“You can reach us any time through our friendly help page, where our team is always happy to help.”

A model grader against a person, on v2's answers

They agree on 8 of 10, but the grader caught 1 of the 2 answers the person failed. They disagree on:

  • “Do you ship to Canada?” The answer: “Yes. 5 to 7 days.” The grader says fail; the person says pass, since the check is: says yes, with the 5 to 7 day wait.
  • “What is your phone number?” The answer: “You can reach us any time through our friendly help page, where our team is always happy to help.” The grader says pass; the person says fail, since the check is: gives the support phone number.

What this simplifies. The questions, answers and checks are made up to show the idea. Each answer is scored by a rule that looks for exact words. Such a rule is cheap and easy to read, but a right answer in other words would fail it. A real test set has hundreds of questions, and the grader here stands for a second model asked to judge each answer.

A change that fixes some answers often breaks others. Looking only at the overall score can hide this.

So compare each question between the old version and the new one, and look at what got fixed and what got broken.

Example. Below, v2 raises the pass rate from 60% to 80% by fixing 3 questions. But it breaks 1: it no longer gives the support phone number.

In system design: Read every question a change breaks before shipping it, even when the overall score goes up.

Check yourself

Version 2 raises the pass rate from 60% to 80%. What should you still do before shipping it?