A nine-exam repeated-sampling study of IRT discrimination, built to be reproducible and restartable. It covers nested samples from deterministic seeds, fit keys that expose reused fits, atomic checkpoints and paired tests with Benjamini-Hochberg control.

Turning written exams into more than 5,000 structured, database-ready questions: extraction, LaTeX conversion, compile checks, LLM tagging of topic, skill and difficulty with a JSON schema, and why a human review console stays in the loop.

Two alternatives to a narrow-prior 2PL estimator: a hierarchical branch whose latent scale flipped in one run, and GIRTH's marginal maximum likelihood. This part shows what larger coefficients, held-out log loss and bootstraps did and did not establish.

The prior mode was named vague, yet it put Normal(0, 0.1) on discrimination and difficulty. Fixed-question experiments, a prior sensitivity check and a 434-student simulation show how that shrinkage tracked sample size.