context-degradation

Place critical information at the beginning or end of the context, never the middle, since attention follows a U-shaped curve with 10-40% reduced recall in middle positions. Add explicit section headers and prepend summaries of long documents to create attention anchors.

Context poisoning occurs when a hallucination, tool error, or incorrect retrieved fact enters context and compounds through self-reference. Detect it through degraded output on previously successful tasks, tool misalignment, and persistent hallucinations; recover by truncating to before the poisoning point.

Degradation typically begins at 60-70% of the advertised context window, and the RULER benchmark found only 50% of models claiming 32K+ context maintain satisfactory performance at that length. The decline is non-linear with a cliff edge, so set compaction triggers around 70% of the known onset.

Instructions placed in the middle of long contexts suffer reduced attention, and irrelevant documents create attention competition the model cannot skip. Move critical instructions to the start or end, filter irrelevant content before loading, and isolate different tasks into separate context windows.

No. Larger windows delay but do not eliminate U-shaped degradation, and processing cost grows non-linearly with token count. For many models, meaningful degradation begins at 8K-16K tokens regardless of window size, so splitting tasks across sub-agents often works better.

Establish source priority rules before conflicts arise, filter outdated versions before they enter context, and mark unavoidable contradictions with structured annotations stating what conflicts and which source takes precedence. Without explicit rules, models resolve contradictions unpredictably and silently.