PBeyond Bias

Data Correction Layer · Tool 01

Synthetic Augmentation Decision Tool

← Back to all tools

A governed decision path for diagnosing representation deficits and determining when synthetic augmentation is appropriate. Work through each step — the tool routes you to the same outcomes as Figure A.1 of the PAL Framework.

  1. Measure
  2. Diagnose
  3. Correct
  4. Re-measure

Step 1 — Measure

Compute the demographic parity ratio (DPR)

Measure demographic parity — the female-to-male approval-rate ratio — on a representative hold-out sample. This is a point-in-time diagnostic, not a production control; ongoing surveillance begins after deployment under the Tool 02 monitoring protocol.

Decision 1: Is DPR at or above 0.80? The 0.80 threshold is the pre-registered H1 decision rule. It is borrowed from U.S. employment-selection guidance and is not a banking-regulator standard.

Report the test at 0.80 and again at a stricter 0.90 sensitivity bound, with a paired significance test following Watkins & Chen (2024). A result at or above 0.80 exits to routine surveillance.

Decision rules: DPR threshold = 0.80 (a pre-registered H1 convention borrowed from the EEOC four-fifths rule, not a banking-regulator standard); AUC degradation tolerance = 0.02 (H1 joint performance condition). This tool reproduces Figure A.1 — Synthetic Augmentation Decision Tool (PAL Data Correction Layer) from Beyond Bias: The PAL Framework.