Source-Judged Compaction Prevents Hallucinations in Long-Horizon Agent Summaries
October 16, 2026
Source-Judged Compaction (SJC) identifies and removes reasoning-based hallucinations in agent context summaries by verifying each entry against environment outputs. This training-free post-processing step addresses the finding that 83% of failed long-horizon runs stem from unverified summary entries.
HOW THIS AFFECTS YOU
●
builderYou can improve long-context agent reliability by implementing SJC to prune unverified summary entries.
●
researcherYou can mitigate context compaction errors without retraining the underlying model or compressor.