Deployment State in Projection-Based Unlearning: Exact Decision Boundaries, Finite-Sample Certificates, and Estimation Bias

We study projection‑based unlearning with mean restoration. When the deployment meanand eraser are estimated from finite samples, the same model can yield different joint decisionsabout forget‑set accuracy (FA) and retained utility. For rank‑one corrections, the mean enters theoutput through a scalar, allowing exact computation of decision boundaries on a finite evaluationset. Changing the mean composition flips 27 of 42 paired decisions across six deletion classes andtwo datasets. Building on these boundaries, we predict mean‑sampling flip risk, derive samplingwithout‑replacementbounds,andusetime‑uniformcertificatesforadaptivestoppingduringasingledeployment. When the eraser is re‑estimated from the same sample, a second‑order expansionidentifies coupling between direction and mean errors. Bias predictions for a hybrid LEACE estimator pass preregistered tests on new ResNet and ViT configurations. The nonsaturated held‑outrisk test for the consistent plug‑in estimator fails its registered criterion. Exact acquisition of theforget‑side mean reduces average flip risk in stress conditions, while its bias effect depends on theestimator. These findings support an audit workflow that records deployment state, distinguishesestimation designs, and certifies decisions.