Artificial Intelligence Benchmarking of Multi-Site Clinical Language Model Debiasers Under Unmeasured Site-Specific Coding Guideline Drift: A Healthcare Fairness-Transfer Evaluation

Clinical language models are increasingly debiased at one site and deployed at another, yet site-specific coding guidelines drift in ways that are rarely measured and never reported alongside fairness metrics. the present analysis formalises fairness transfer under unmeasured guideline drift as a bias–variance problem for the debiasing estimator, and evaluates the ordinary least squares estimator as a benchmark of that transfer. Six hundred data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group −3.1, severity 1.35, error SD 4.2) and the estimator was characterised across replications. Estimated bias was 0.01062 for the intercept and 0.00019 for the exposure coefficient; root mean squared error was 0.85994 and 0.0152 respectively; empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. The sampling distribution of the exposure coefficient was approximately normal and the Q-Q plot was consistent with that approximation. The results establish that an unbiased, well-covered estimator can still be an unreliable fairness-transfer instrument when the drift term is unmeasured, because coverage of the nominal interval does not certify coverage of the drifted target. Benchmarking protocols for clinical debiasers should therefore report drift sensitivity alongside accuracy, and should treat site-specific coding guidelines as a first-class source of distribution shift rather than as a nuisance to be absorbed by pooling.