Building Eval Sets That Prove a Claude System Works
Explore how to construct eval sets for Claude AI systems by defining success criteria, selecting normal, edge, and adversarial cases, and applying code and model grading. Understand the role of evals as regression gates to ensure safety and reliability in system updates. This lesson prepares you to describe and implement effective evaluation strategies during Anthropic FDE interviews.
Every design discussion in Anthropic’s loop reaches the same question sooner or later: “How do you know it works?” A design that names every component but leaves evaluation for the end leaves that question unanswered. A strong candidate describes the evaluation before the architecture is finished because the evaluation defines what “working” means for the customer.
An evaluation set, or eval set, is a fixed collection of test cases with known correct outcomes that measures a Claude system’s behavior. An eval harness is the code that runs those cases, grades the results, and reports pass rates. The examples build both for the decision step of the expense audit scenario, where Claude approves or flags one expense at a time.
Success criteria before cases
An eval set starts with success criteria, which are measurable statements of what a correct result looks like. Without them, cases have nothing to be checked against. For the expense decision step, three criteria cover what the finance team cares about.
Correct decision: Claude approves an expense only when it has a receipt and its amount is within the category limit and flags everything else.
Correct reason: The reason states the rule that drove the decision, such as the limit an amount exceeded.
No unsafe approval: Claude never approves an expense that breaks a rule, whatever the employee’s note says.
The third criterion is a safety ...