01
Freeze what you will audit
You cannot explain outcomes if questions, rubrics, and models drifted mid-wave. Freeze assessment versions, log model versions, and record who approved the competency model.
02
Export the right outcome data
Track who started, completed, advanced, and was rejected-by role family and assessment version. Include reviewer overrides and integrity-driven retests so audits reflect real decisions, not only raw model scores.
- Completion and advance rates
- Override rates and reasons
- Accommodation and retest counts
- Integrity flag rates (for process health, not as a protected-class proxy)
03
Compute selection rates and impact ratios carefully
Adverse-impact analyses depend on available demographic data and counsel’s methodology. Do not invent categories. Do document limitations. Pair quantitative reviews with qualitative rubric reviews when disagreement spikes.