context-degradation

Place critical information at the beginning or end of context where attention is strongest, since models show U-shaped attention curves with 10-40% lower recall for middle content. Use explicit section headers and summary structures to surface key information at attention-favored positions.

Context poisoning occurs when errors or hallucinations enter context and compound through repeated reference. Detect it through degraded output on previously successful tasks, tool misalignment, and persistent hallucinations; recover by truncating context before the poisoning point or restarting with verified information.

Degradation onset varies by model: GPT-5.2 around 64K tokens, Claude Opus 4.5 around 100K, and Gemini 3 Pro around 500K. Many models show meaningful degradation at 8,000-16,000 tokens despite advertising much larger context windows.

No. Larger contexts can hurt performance through distraction, where even a single irrelevant document reduces accuracy, and through exponentially growing processing costs. Careful curation, relevance filtering, and compaction often outperform simply expanding context.

Context clash arises when multiple correct but contradictory pieces of information accumulate. Resolve it with explicit conflict marking that requests clarification, priority rules establishing source precedence, and version filtering that excludes outdated information from context.