A/B Testing for ML Models Quiz

Answers and explanations appear as you take the quiz.

1. Why do teams A/B test models in production instead of trusting offline metrics?
   • A model with better offline accuracy can still hurt real business metrics
   • Offline metrics are illegal
   • Production is cheaper than offline testing
   • Offline data does not exist
2. What role does random assignment play in an A/B test?
   • It makes the test run faster
   • It supports balance between groups in expectation, subject to design and implementation assumptions
   • It deletes biased users
   • It reduces the number of models needed
3. In a model A/B test, what usually happens in the treatment group?
   • The candidate scores requests but all outputs are withheld
   • Every user is assigned to both models for the same decision
   • The candidate model serves user-facing decisions for the randomly assigned group
   • The control model is removed before any comparison
4. Under a null hypothesis and stated test assumptions, how is a p-value interpreted?
   • As the probability that the null hypothesis is true
   • As the probability that the treatment caused the observed result
   • As the probability that the experiment will replicate
   • As the probability of data at least this extreme if the null hypothesis holds
5. How can a multi-armed bandit change an experiment while it runs?
   • It can allocate more traffic to variants with better observed outcomes and less to others
   • It keeps every variant at a fixed equal allocation throughout
   • It removes randomization from every assignment
   • It guarantees the best variant is identified without uncertainty
6. What is the purpose of a power analysis before an A/B test?
   • To pick a logo for the experiment
   • To estimate how many users are needed to reliably detect a real effect
   • To encrypt the traffic
   • To delete the control group
7. What does a sample-ratio mismatch (SRM) signal during an A/B test?
   • The treatment model is necessarily better
   • The experiment has sufficient statistical power
   • Observed assignment counts differ materially from the configured allocation, which can indicate a data or randomization problem
   • The primary metric improved causally
8. Why include guardrail metrics alongside an experiment’s primary success metric?
   • To guarantee the candidate wins on every metric
   • To replace the primary metric after results are observed
   • To avoid defining a decision rule before the test
   • To detect meaningful harms in other outcomes even when the primary metric improves