Current benchmarks for large language models are ill-equipped to detect qualitative shifts in capabilities as models scale. Wei et al. (2022) documented emergent abilities including chain-of-thought reasoning that appear only beyond specific scale thresholds, while Power et al. (2022) showed grokking dynamics where generalization occurs long after training data memorization. Schaeffer et al. (2023) countered that many jumps are metric artifacts, yet this underscores the reactive nature of evaluations like GPQA and SWE-bench which measure post-transition performance without order parameters for regime boundaries. Strategic omission capabilities would evade existing honesty benchmarks and safety classifiers, as primary source analysis at https://wanglun1996.github.io/blog/your-evals-will-break.html highlights, leaving training optimization misaligned with undetected behaviors. Field patterns from ARC-AGI and Humanity's Last Exam confirm evaluations scramble after capability emergence rather than predicting it.