Module 13: ethics and fairness across the lifecycle

{
  "module": 13,
  "dataset": "UCI Adult historic census extract",
  "train_rows": 26029,
  "validation_rows": 6508,
  "test_rows": 16253,
  "selected_threshold": 0.3,
  "features": [
    "age",
    "education_num",
    "capital_gain",
    "capital_loss",
    "hours_per_week"
  ],
  "audited_sex_race_used_for_fitting": false,
  "overall": {
    "n": 16253,
    "tp": 2394,
    "tn": 10366,
    "fp": 2045,
    "fn": 1448,
    "precision": 0.5393106555530525,
    "recall": 0.6231129619989588,
    "specificity": 0.8352268149222464,
    "selection_rate": 0.27311880883529194,
    "fpr": 0.1647731850777536,
    "balanced_accuracy": 0.7291698884606026,
    "f1": 0.5781910397295013
  },
  "interpretation": "Descriptive audit of a historical income-label task, not a decision rule for people."
}

Historical income is not a measure of merit, need, or a justified allocation outcome. Do not deploy this model for eligibility or targeting.

The source records limited sex/race categories. They do not capture identity, intersectional experience, or all affected groups. Differences are descriptive, not causal findings.

Excluding protected columns does not remove proxy effects or label bias. A metric does not establish fairness or legal compliance.

Interval columns are 95% Wilson illustrations conditional on independent binomial sampling, which this historical extract does not establish. They are not population confidence claims. Undefined rates and intervals remain unavailable; small-denominator flags use each rate denominator below 30. Interval overlap is not a test of group differences.