CONTEXT ROT AND CONTEXT BLOAT

CHAPTER SUMMARY Context bloat is the filling of the window with low-value text, usually through verbose tool output, repetition, stale items, and oversized instructions. Context rot is the drop in model accuracy as the window grows very large, even within the technical limit. Needle-in-a-haystack testing hides one fact in a long text at different positions and lengths to reveal where retrieval weakens, though passing it is only a minimum check. Symptoms of degraded context include forgotten rules, contradictions, repetition, missed or invented details, goal drift, and rising cost with no gain. The cure is a lean, high-signal context, regular trimming, restating of critical rules, and enforcement of safety in code. KEY TERMS REVIEW Context bloat: a window filled with too much low-value text. Context rot: declining model accuracy as the window grows very large. Distractor: content that resembles the answer but is not the answer. Needle-in-a-haystack test: a test that buries one fact in a long text and checks whether the model can retrieve it. Hallucination: a model stating something false as if it were true. Signal-to-noise ratio: the proportion of useful information compared with useless information.