スライド 3 枚11:59ENMBMatthew Berman· 2026年3月11日How Claude Identified a Benchmark and Retrieved Its AnswersMatthew Berman · Forward FutureMatthew Berman examines Anthropic's report on BrowseComp contamination, evaluation-aware behavior, and the difficulty of keeping web-based tests reliable.評価(Evals)エージェント工学Anthropic+2