This paper critiques the current methodologies for evaluating and mitigating benchmark contamination in large language models, arguing that the standard Gap of Aggregate Performance (G-AP) metric provides a flawed representation of a model's genuine capability. The authors demonstrate that G-AP's reliance on discrete correct/incorrect readouts and uniform weighting allows over-suppression and under-suppression to cancel each other out, potentially masking ineffective mitigation. To address this, they introduce Stratified Aggregate of Per-question Probability Gaps (SA-PPG), a new metric that estimates solve probabilities via sampling and aggregates differences against a clean model within strata defined by the clean model's performance. Furthermore, the researchers propose RailCap, a real-time mitigation strategy that intervenes during the decoding process. Unlike existing methods that rely on potentially inaccurate pre-computed contamination estimates, RailCap monitors generation; if a sample reverts to a greedy trajectory, the subsequent token is capped to the runner-up probability, forcing distribution dispersion. Experimental results across various contaminated models and benchmarks indicate that previous mitigation strategies significantly overestimate their effectiveness when measured by SA-PPG, whereas RailCap achieves the lowest SA-PPG, demonstrating superior ability to restore authentic model performance by suppressing memorized patterns.