From a training standards cell: "For comparability we sometimes need every candidate on the same paper, and for appeals we need to reconstruct exactly what a candidate was shown. Randomised generation sounds incompatible with both."
It is compatible with both, by construction. Exam generation runs on a seeded PRNG, and the seed fully determines the paper: which part records were drawn, which of the five question forms, which distractors, in what order. Issue every candidate the same seed and they sit the identical paper. Record the seed with the result — it is short enough to write on the result itself — and any paper can be regenerated later, question by question, for an appeal or an item review.
Because questions are generated from the maintenance records rather than a hand-written bank, reviewing a disputed item leads you to a part record, and the fix — if the item really is bad — is a content fix that improves the trainer, the knowledge checks and all future papers at once. One correction, one place.
One nuance: a paper regenerates identically only against the same content version, since the records are the question source. Pin the content version alongside the seed in your records if reconstruction needs to survive content updates — under contract we timestamp both automatically.
This board is read-only — TaskFlight has no accounts and stores nothing on a server, so replies live on GitHub Discussions, where the conversation is public and moderated. Prefer email? Same address as always; useful questions get added to this board with identifying details removed.