Context Pollution
Quality bugs that look like the model got dumber are often a dirty window.
A desk piled with last week's mail. You miss the invoice.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
This turn
Scroll inside the frame to explore · use + / − to zoom up to 200%
A desk piled with last week's mail. You miss the invoice.
Key Takeaways
- 1.Pollution is leftover, contradictory, or injected text that changes behavior.
- 2.Symptoms: ignored tools, sudden policy changes, looping.
- 3.Causes: unfenced RAG, huge traces, stale summaries, duplicate memories.
- 4.Fix by isolating, compacting, and dropping duplicates.
- 5.Quality bugs that look like the model got dumber are often a dirty window.
Real Example
Scenario
Pollution: pack only what this turn needs.
What you would do
In Context Engineering, apply Context Pollution to this scenario: Pollution: pack only what this turn needs. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Pollution is leftover, contradictory, or injected text that changes behaviorSymptoms: ignored tools, sudden policy changes, loopingCauses: unfenced RAG, huge traces, stale summaries, duplicate memoriesFix by isolating, compacting, and dropping duplicates
Cheat Sheet
Quick recap
quick ref- •Pollution
- •Budget
- •Don't dump
Common Mistakes
- ✕Skipping evaluation for Context Pollution before production
- ✕No logging or tracing around context pollution steps
- ✕Ignoring cost and latency implications