An investigation into Claude Code's context compaction revealed that the model prioritizes certain types of information over others, leading to the loss of critical instructions. Specifically, rules related to stock tickers survived compaction, while a rule about a user experiencing a cardiac event did not. This behavior was observed even when the dropped rule included severe consequences like death, suggesting that the model's prioritization is not solely based on the severity of the stated outcome. AI
IMPACT Reveals potential safety and reliability issues in LLM context management, impacting agentic systems.
RANK_REASON Investigative report detailing unexpected behavior in a specific AI model's feature. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →