Work samples
Reading an answer and producing one are different skills, and the loop tests the second. Each of these is a small broken thing with real fixtures, a grader you run locally, and a reference solution that explains what was actually being scored.
The brief is open. The pack is part of premium access. Every dataset is synthetic and written for the exercise, and nothing needs a customer credential or an API key to run.
Repair a customer's streaming assistant
Three symptoms filed as one ticket, with three separate mechanisms behind them, plus a fourth defect nobody reported because it is invisible until the server fails. Fix the handler until the grader is green, then explain the state model.
Explain why finance's SQL is wrong
A text-to-SQL pilot returns a plausible wrong number for Q2 revenue. Three independent defects, two of which move the total in opposite directions. Fix the query, then explain it to a finance lead.