Item exposure is the quiet failure of a small question bank

A certification exam with two hundred items and four thousand candidates a year has a problem that does not appear in any score report. Every item is seen often enough, by enough people, that its content leaves the room. After a while the exam still discriminates — it just discriminates between candidates who encountered the circulating set and candidates who did not.

That is a measurement failure, and it is invisible from inside the results, because a compromised item often looks like an easy item.

The arithmetic is simple enough to do on paper

Exposure is sittings multiplied by items per sitting, divided by bank size. Four thousand sittings of sixty items drawn from two hundred gives each item roughly twelve hundred viewings a year. No amount of candidate honesty survives that number; it only takes one person in twelve hundred to write the question down.

The reason this gets missed is that nobody computes it. Bank size is discussed as a content-production problem — how many items can we afford to write — rather than as a constraint derived from volume.

What a compromised item looks like in the data

Difficulty drifts in one direction

An item that has leaked gets easier over time, and only that item. Difficulty drift is normal across a whole form as the candidate population changes; drift concentrated in a handful of items, while their neighbours hold steady, is a different signal.

Discrimination collapses

The stronger signal is that the item stops correlating with overall performance. A leaked item is answered correctly by people who prepared badly, which is precisely the pattern a discrimination index is designed to catch. Watching it per item, over time, catches exposure earlier than watching difficulty alone.

What to do about it

Retire on a schedule rather than on suspicion. An item with a planned service life is replaced before it has leaked; an item retired when someone notices something odd was compromised for however long the noticing took.

Draw from a larger pool than the form length, with content constraints applied at draw time so two candidates get statistically comparable forms from different items. That is more expensive to build than a fixed form, and it is the only structural answer — every other measure slows exposure without bounding it.

The part that is a product decision

All of this needs the platform to know which items were served to whom and when, and to keep that history after the sitting is scored. Systems that treat a delivered form as transient throw away exactly the data the exposure analysis needs, and the gap is only discovered years later when someone asks how long an item has been in service.