ADR 0001 — Composite Bias Score formula & weights

• Status: Accepted
• Date: 2026-05

Context

The STP specifies a single Composite Bias Score aggregating Disparate Impact (DI), Statistical Parity Difference (SPD) and Equalized Odds (EO) with weights 0.40 / 0.35 / 0.25, and pins compute_composite_bias_score(0.6, -0.25, 0.15) == 0.7150. Its original formula was undefined for DI > 1 (a favoured group), where a naive linear term would reward over-selection.

Decision

Normalise each input to a 0–1 fairness-goodness sub-score before weighting:

g_DI = min(DI, 1/DI)   g_SPD = 1 - min(|SPD|, 1)   g_EO = 1 - min(|EO|, 1)
Composite = clamp(0.40·g_DI + 0.35·g_SPD + 0.25·g_EO, 0, 1)

min(DI, 1/DI) makes the DI term symmetric about parity, so a group favoured by 25% scores the same as one disadvantaged by the mirror amount. The weights remain configurable; the defaults preserve the STP's pinned golden value (verified in test_metrics_known_values.py).

Consequences

• Higher composite = fairer (documented prominently to avoid the "bias score" naming trap).
• The score is explicitly a triage convenience, not a legal standard; raw metrics + CIs are always reported alongside it (see LIMITATIONS.md).
• Changing the weights changes the band a model lands in — a product decision, recorded here.