The problem, and the answer
Situation
Hundreds of hires a year, decided by one combined score.
Large South African employers hire in bulk: graduate programmes, learnerships, branch and contact-centre roles. Each applicant takes several assessments, and the scores are combined into one.
Complication
The best predictors carry the biggest group gaps.
The tests that best predict performance also show the largest score differences between groups. Lean on them and you hire stronger performers, but fewer people from the focal group. Most teams only see this after the intake.
Question
Can you see your options before anyone is hired?
The person designing the assessments needs to see every reasonable choice, what each one costs, and pick one they can defend.
The Pareto Selection Planner
See every best trade-off up front; the choice stays with you.
Before the intake, the assessment designer sees how much hiring quality each option gives up for fairer outcomes. The choice becomes explicit, with a record to justify it under the Employment Equity Act.
Pick the assessments. The optimisation engine, which is the AI in the tool, finds the best weighting for every possible trade-off and draws them as a frontier. Choose a point, then score real applicants. This is the demonstration.
It reproduces the published results of De Corte et al. (2007) and Berry et al. (2024) to within .01. The Checks tab shows the comparison.
Who it is for
The I-O psychologist or talent assessment lead at a large South African employer who designs the assessment battery for a high-volume intake, such as a graduate programme, learnership, or branch or contact-centre hiring, and who must be able to justify that battery under the Employment Equity Act.
When they use it. Once per intake design cycle, before applications open.
What they do today. Choose tests by vendor default or validity alone, weight them equally or by judgement, and check adverse impact after the intake. The trade-off stays invisible until it has happened, and there is no record of the options considered.
What the user needs, and what the tool does
| The user needs to | The tool |
|---|---|
| Choose which assessments to use | Predictor library: any two or more of seven assessments |
| See the trade-off before hiring | Frontier chart of every best-available weighting |
| Compare specific options | Click any point to see its weights and outcomes |
| Check against real applicants | Upload a CSV to rank applicants and see who is hired |
| Defend the decision | Checks tab and downloadable ranked file |
Caveat: simulation only
The values are mostly US meta-analytic estimates and are suitable for simulation only. Section 8 of the Employment Equity Act requires that assessments be shown to be valid and reliable, applied fairly and not biased against any employee or group. Real use requires South African validation studies, local norms or South African meta-analytic evidence (see Theron, 2009).
Extension: other group comparisons
Gender (men–women) comparisons are a natural extension, but published men–women d values across these predictors are patchy, so the tool models Black–White differences only. The engine supports any two-group comparison once defensible d values exist.
Glossary
The key terms come first; these are the ideas needed to follow the demonstration. Further terms below cover the detail.
The focal group is the group whose selection rate is compared; the reference group is the comparison group. In South Africa the protected group is often the majority of applicants, so the default focal share is 60%. The calculation works for any share.
Quality against adverse impact
Each grey dot is one sampled weighting of the selected predictors. The blue line is the Pareto frontier. Hover the frontier to see its weights; click anywhere in the chart to find a weighting that gives that result.
Computing the frontier
Weights along the frontier
How the best weights change as you trade quality for a higher AI ratio. The vertical line marks your current AI ratio. Click to use the weights at that point.
Score an applicant pool
Upload raw applicant scores to rank them with your current weights and select the top 15%. Each score is converted to a z-score with the norm table, and the composite is the weighted sum of z-scores. The file is read in this browser and is not sent anywhere.
Predictor parameters
Validities, subgroup differences and intercorrelations for the whole library. Edit any cell; the frontier updates when the values are valid.
›
d is the reference-group mean minus the focal-group mean in within-group SD units. Validities and correlations are within-group. Edit the lower triangle; the upper triangle mirrors it. Rows for predictors outside the battery are dimmed but still checked.
Checks
Everything that exists to verify the results: two known-answer tests against published tables, cross-checks of the optimiser, matrix validity, simulation against the analytic model, edge cases including hand-scored pools, and configuration stress tests. Checks 3 to 8 use the battery and weights currently set on the Dashboard; checks 1, 2 and 9 use fixed preset values.