Fraud Detection Without Specificity: A Null-Calibration Evaluation of the eforensics Model on Two Decades of Korean Elections

Bayesian finite-mixture election-forensics models in the tradition of Mebane (2016) classify vote-counting units as clean or fraudulent from aggregate turnout and vote-share data alone, without independent evidence of manipulation. We evaluate the eforensics quasi-binomial-logistic (qbl) model against 20 years of Korean elections (2002-2025), fitting 228 separate model runs spanning general and presidential elections, two aggregation levels, up to five vote-counting channels, and, critically, two independently constructed fraud-free synthetic controls calibrated to match each real dataset's turnout and vote-share marginals without any fraud-generating mechanism. We find that the model reports substantial "fraud" (5-31% of units, up to 8.7 million "fraudulent" votes) on data that is fraud-free by construction; that real-data fraud-share estimates correlate at r=0.98 with their fraud-free null counterparts across 58 paired cells; that the model's largest real-data estimates fall on presidential elections held before Korea had early voting or any fraud allegations (2002-2012), exceeding what it reports for the actually contested 2020 election; and that recoding the model to test fraud in the opposite (conservative-favoring) direction reproduces the identical pathology, including assigning more false-positive fraud to the one pre-2014 election with a genuine historical manipulation allegation than to the allegation's own claimed direction. We conclude that the model's output depends on the shape of the (turnout, vote-share) distribution, not on fraud, and that this failure is not fixable by post-hoc recalibration, since recalibrating for validity would remove essentially all detected signal in the corpus.