An independent bias audit of your hiring AI.

You share counts. Never people.

At one decision point, the shortlist for example, over one period: how many candidates in each group were considered, and how many were selected. That is all. The intake refuses row-level data, so a file of individual records is turned away, not stored.

Counts are enough to measure a disparity. They are not enough to identify a person.

You get evidence, signed.

• ✓Selection rates for every group, with their 95% intervals.
• ✓Impact ratios against the best-treated group, the four-fifths benchmark used by NYC Local Law 144 and the EEOC.
• ✓Whether each gap is statistically significant, and how large it is.
• ✓How much the candidates who did not declare a group could change the answer.
• ✓Intersections and per-location slices, marked exploratory.
• ✓The legal citation behind every ground tested.

The report is signed. Your client’s DPO can verify it without asking us.

The engine computes. A person decides.

The law’s test has two parts, and only one can be computed. Whether a practice puts a group at a disadvantage is arithmetic. Whether that is justified by a legitimate aim is an argument, and it belongs to a person. So the engine proposes a size for each gap and a named reviewer makes the determination.

It never outputs the word compliant.

Independent means independent.

An audit is only worth something if the auditor had no hand in the tool. We audit only where we have had no part in building it. A client can have Audit or Flow, not both.

The method was tested against the published tables of a real NYC Local Law 144 audit: same method, same reference group, same small-group exclusion rule.