Generation is a review pipeline
Question generation is split into bounded stages: AWS-source research, composition, independent solving, and explanation review. Uncertain material goes to a review queue. Metadata records source links, exam domain, prompt version, model, and review status.
The application does not present generated questions as recalled exam items or promise an official score. Its documentation keeps raw accuracy separate from AWS's scaled scoring.
Use mistakes to choose the next session
Practice follows the selected exam's documented domain weights. Review prioritizes overdue topics and confidently wrong answers, while correct answers extend intervals according to confidence. An exam mode defers feedback until the end rather than teaching during the timed run.
Questions, attempts, confidence, session timing, and review state live in a revision-controlled Redis document. Atomic updates prevent one host from silently replacing another's work.
Question batches are bounded, saved incrementally, and serialized across hosts. That makes partial progress inspectable without treating an interrupted batch as a completed question bank.
Preserve cost and evidence boundaries
Paid generation has deadlines and no automatic paid retries. A restart does not silently resume provider work. Independent review can catch some errors, but source links remain essential for assessing an answer. The learning loop therefore keeps question provenance, confidence, and actual attempts alongside the generated material.