SWE-bench Science: the best agent scores under 50% pass@1.

119 repository-level tasks from 98 GitHub repos across 20 scientific domains, split into issue-driven, expert-exploratory, and engineering-integration paradigms. Claude Code with Opus-5 (max) lands below 50%. arXiv The ablation is the better finding: stripping explicit scientific guidance while keeping repo and execution context shows domain knowledge isn't uniformly helpful. Well-grounded facts improve average performance and token efficiency; poorly aligned guidance anchors the agent and doesn't raise exact repair success. That's the same anchoring mechanism MemTrapBench found in memory, showing up in a completely different input channel.