# Answer Key: Experiment Registry Audit

Use this as a calibration guide, not a single correct answer.

## Core Readout

Low-risk completed experiments can usually proceed through lightweight review. High-risk experiments, blocked experiments, and experiments in review need stronger scrutiny, especially when they affect pricing, ranking, account-level outcomes, or user safety.

## What A Strong Answer Should Say

- Use proportional review: do not make every click-through test feel like a medical trial, but do not let high-risk work bypass human judgment.
- Flag high-risk experiments with weak or missing guardrails.
- Check whether the randomization unit matches the decision and interference risk.
- Require archive fields after readout: hypothesis, primary metric, guardrails, decision, validity checks, owner, and follow-up.
- Recommend redesign for blocked experiments before launch, not just more analysis after launch.

## Common Mistakes

- Treating every running experiment as equally risky.
- Equating "has a guardrail column" with "has a sufficient guardrail."
- Letting revenue metrics ship without complaint, refund, latency, fairness, or user-harm checks.
- Failing to archive null or inconclusive results.

## Instructor Notes

Ask learners to sort the registry into "auto-approve," "human review," and "block/redesign." Strong answers explain why the threshold differs by risk tier.