# How do you measure privilege review quality?

> Privilege review quality is measured by testing a sample of the review population for two kinds of error - privileged documents that were missed and non-privileged documents that were withheld - and tracking those rates alongside reviewer consistency and the completeness of the privilege log.

Privilege is the one area of document review where a single miss can be unrecoverable. Quality therefore has to be measured, not assumed.

Two error types matter, and they pull in opposite directions:

-   **Under-designation**: a privileged document is produced. This is the disclosure risk everyone worries about.
-   **Over-designation**: a non-privileged document is withheld. This inflates the privilege log, invites challenges, and costs credibility.

The usual measurements:

**Elusion testing.** Sample the documents the review coded as not privileged and re-review them carefully. The proportion that turn out to be privileged is your elusion rate, and it is the closest thing to a direct measure of what you are about to produce by mistake.

**Recall and precision.** Against a carefully reviewed control sample, recall tells you what share of truly privileged documents were caught, and precision tells you what share of withheld documents were genuinely privileged.

**Reviewer consistency.** Route the same documents to more than one reviewer and compare. Wide divergence usually means the privilege criteria are ambiguous, not that the reviewers are careless.

**Log quality.** Descriptions should support the claim without waiving it, and should characterise similar documents consistently.

Whatever the method, write the protocol down before you start. A documented, sampled, reproducible process is what makes the result defensible when it is challenged.

Related reading: [Privilege review in the age of generative AI](/stories/privilege-review-in-the-age-of-generative-ai).

One structural advantage of AI-assisted review is consistency: the same criteria applied to every document, with the reasoning recorded for QC. You can [see how Claira handles it](/demos).