Compaction as Epistemic Failure: How Agentic LLM Tools Fabricate Confirmed Results from Killed Processes

This paper documents a reliability failure in Claude Code’s session compaction workflow. When a timed-out command is killed with exit code 143, partial standard output may be written into a compaction summary and later inherited as a confirmed result across sessions and model versions. The authors describe the underlying issue as a conflation of observation with persistence: seeing output in a terminal is treated as equivalent to verifying that the result was durably written. The paper connects this failure to broader weaknesses in LLM self-evaluation and discusses implications for scientific computing, data processing, and multi-step automation.