Resume Ranking Audit: Label Bias Versus Proxy Features

Both are present, but the label is the root problem. The label 'was hired' is a decision made by interviewers who saw only the top 200 ranked candidates; it encodes previous model and human preferences, not job performance. If women were under-ranked historically, the label teaches the model that maleness predicts hiring. That is label bias and a construct-validity failure: the target is not 'will succeed' but 'was hired under a biased process'. Features such as graduation year and women's college names are proxy bias, but removing them alone will not fix the label because the text embedding can reconstruct gender from other cues. The precision@10 gap of 0.71 versus 0.82 could also reflect different base rates or score calibration, so inspect score distributions, not only top-10. Legal should ask for adverse impact, but product should not simply raise the threshold for women; that is a temporary patch that may lower validity and stigmatize candidates. Fix the target first with structured job-performance or work-sample outcomes, or blind the historical decision, then remove proxies and audit.

Report precision@10 and recall@100 by gender and by intersectional groups such as gender × career gap × school tier, plus selection rate and impact ratio under the four-fifths rule. Add true-positive and false-positive rates at the operating threshold, calibration by group, and score-distribution overlap. Because the concern is ranking, use nDCG@10 with a relevance label from job performance, not hiring. Measure the career-gap slice separately: precision@10 of 0.64 may reflect real experience differences or a proxy penalty; test it with a matched-pairs audit where resumes differ only by a 12-month gap. Use bootstrap confidence intervals and a minimum subgroup size to avoid unstable claims. Track the top-200 funnel: if interviewers only see the top 200, measure how many qualified women never enter review. Finally, validate labels by comparing 'hired' against 6-month performance ratings for a subset; if the correlation is low, the model is measuring the process rather than merit. A threshold artifact would show overlapping score distributions with a gap only at the cutoff.

Do not start with group-specific thresholds. First, rebuild the label using performance or blind work samples, and audit the historical funnel. Second, remove or neutralize proxy features and test a matched-pairs counterfactual; if the gap persists, the text embedding is encoding proxies. Third, consider adversarial debiasing or reweighting, but validate on a held-out set with subgroup metrics and job-performance validity. If legal requires immediate remediation, use a temporary calibrated threshold with human review for the affected group, documented as a remediation rather than a permanent model, and monitor selection rate and impact ratio monthly. The ethical judgement I would defend is that a model ranking by 'hiring fit' from biased decisions is not fair simply because gender is not an input. Product should accept a small precision@10 drop, for example from 0.78 to 0.75, if validity and adverse impact improve substantially. Publish a model card and give candidates a human appeal path.