The original target is a fixed bounded toy population. Every candidate independently selects one of ten records uniformly with replacement. The biased mechanism changes observation probability; every accepted record remains available for later draws. Nonresponse is newly simulated per invitation, while frame membership and survivor flags are fixed. These are illustrative mechanisms, not a causal model of why real records disappear.

A seed-1309 32-bit generator prepares 20 nonoverlapping blocks of 10000 candidate attempts, consuming two uniform positions per attempt. The first chooses a record, the second tests observation probability. Each sample shows its first n accepted records. Size edits preserve that retained prefix; changing scenario or method filters the same candidates. Finite pseudorandom blocks do not prove independence, convergence or monotone improvement.

Expected bias concerns the collection expectation minus the original target; one observed gap is finite sample error. These rules push upward by construction. Missingness need not bias a mean, and real bias can have either sign. Survivor-only sampling can suit a survivor-only target; here the target deliberately includes exited records. Ideal random is a comparison, not a practical correction promise. Weighting, standard error, intervals and identification from missing data are outside this lesson.