Paired-binomial admission gates pass 31.6% of policy updates at 2,000 episodes where range-based gates pass zero

The paper demonstrates that independent validation gates intended to block harmful updates in continual embodied agents also eliminate learning opportunities when outcome disagreements are rare. A standard paired-binomial test reduces the sample burden required to certify unchanged old-task behavior compared with range-based thresholds. The authors further define certified historical-reference promotion and a round-level missed-opportunity metric to quantify blocked progress. Evidence comes from the constructed diagnostic and a separate learned-dynamics stress test that isolates model bias from feedback-selection error. Unconditional replay outperformed gated streams in closed-loop runs, showing that strict admission criteria can degrade overall capability even when individual updates appear safe. This pattern mirrors metric inflation problems in ad platforms where bot traffic triggers overly conservative fraud filters that suppress legitimate campaign adjustments. Both domains trade error control against retained learning or adaptation budgets under fixed interaction limits. Physical-robot and vision-language-action validation remain open; the admission-audit protocol supplies analytical bounds and synthetic evidence that future deployments can test directly against production update streams.