roc-curve-auc.md's COMPAS example claims the model predicts real recidivism - it predicts COMPAS's own score #623

Where: explainers/roc-curve-auc.md's "Concrete Example: COMPAS - Audit 01": "Audit 01 in this repo reproduces the COMPAS recidivism model on the ProPublica dataset... its baseline models are ordinary, unremarkable rankers: the logistic-regression baseline lands at an AUC of 0.68, and the random forest at 0.69."

The gap: the AUC/n values themselves are accurate (they match paper/results-frozen/results_performance.csv exactly), but the framing is wrong. Per COMPAS/audit.yaml, this audit's model predicts ScoreText - COMPAS's own risk label - not real two-year recidivism. This repo's COMPAS/compas-scores-raw.csv has no independent recidivism-outcome column to train against at all (the same underlying fact behind #600/#621/#622). Calling it "the COMPAS recidivism model" implies it's predicting reoffense, which it isn't.

Repro:

grep -n "recidivism model" explainers/roc-curve-auc.md
cat COMPAS/audit.yaml | grep -A2 "^target:"

Fix direction: reword the framing to describe what the model actually predicts (COMPAS's own high/medium-risk label from demographic/custody features), without changing the correct AUC numbers themselves.