Loading…
15 hard questions on parts 11 to 20, emphasis on 16 to 20: hybrid ranking changes, an MCP incompatibility, an agent over budget, a judge that prefers longer answers, fine-tuning leakage, missing trace spans, cross-tenant cache leakage, a secret misconfiguration and a failed rollback.
All questions on one page
Scroll through the whole paper and answer in any order, like a real exam.
Pass mark
You need 10 of 15 marks to pass.
Timed
You get 35 minutes. The exam submits automatically when time runs out.
15 questions, 1 mark each. Pass mark: 10. The questions mix multiple choice, code reading, output prediction, debugging and design choices; some have more than one correct answer, and some ask you to type a short answer with a reason. Code in questions is Python 3.12 with the standard library and the tutorial kit, runs offline against a fake model and invented toy data, and needs no API key, GPU or paid service. Every scenario, name and number is invented for practice. Answers are judged on engineering reasoning: use the inputs and error messages given, name the trade-off and give a decision you could defend. Model names, prices and SDK details change, so the questions test the ideas, not memorised values.
Sign in to join the discussion.