The canon · Testing · AIS-5.1

Describe testing performed to detect unfair discrimination in AI system outputs, including methodology, protected classes considered, and cadence.

This is one of the twenty-five questions a carrier reviewer sends a vendor, in the words an examiner uses. Below it: the answers that make a reviewer stop on this item, derived by running Vaulith's rules engine on an estate that gets everything wrong, and the free questions that reach it.

What stops a reviewer here

One finding reaches this item.

Will escalateR-CONTRADICTION-BIAS · blocks this item

Your policy commits to scheduled fairness testing the estate does not show

AI policy (3.3) commits to testing fairness or disparate impact on a set schedule. MDL-001 Claims triage does not meet it, because no test is on record at all. A reviewer who reads the policy first expects a test result from the last ninety days to be in the file, and its absence tells them the schedule exists on paper. Once they have found one control that exists only on paper, they check the others the same way.

To fix: Run the test the policy already requires and record the date, or amend the policy to the cadence you actually keep. A commitment you do not meet costs more than one you never made.

How it is scored

Credit per verdict, times the weight.

A standard item carries weight 2. A pass earns 1.00 of it, a questioned answer 0.60, an escalated one 0.20 and a stop 0.00. An unanswered item earns nothing and still takes the citation penalty, so a profile cannot be improved by leaving an inconvenient question blank. The whole arithmetic is shown on the homepage and in every assessment.