The Counterfactual Preservation Principle: Evidence Decay and the Limits of AI Delegation

AI delegation can remove the comparison evidence needed to establish whether automation continues to improve outcomes. This paper develops the counterfactual preservation principle: sustained claims of current comparative benefit require a continuing source of identification whose temporal relevance remains defensible. A bounded-loss model establishes identification intervals under comparator withdrawal, an evidence-renewal horizon, finite-window error and confidence bounds, and an estimator-independent rate obstruction under fixed randomised quotas. A sufficient allocation frontier links comparator flow, task volume, temporal drift, and precision, including integer and single-period boundaries. Exact calculations and 5,000 simulated trajectories per design and scenario compare six designs under stationarity, gradual reversal, and abrupt misspecification. In the specified gradual reversal, retiring the comparator yields root-mean-squared error of 0.08972, compared with 0.01319 when 5% of tasks retain comparison over 22 periods, despite unchanged AI performance. Retirement performs best among tested designs under stationarity, while abrupt changes expose failures of an understated drift envelope. The contribution integrates causal and statistical methods into an institutional design principle: credible delegation requires a falsifiable comparison, with allocation determined by evidence needs and ethical eligibility rather than a universal retention percentage.