Dumb zone

The part of a long context where the model starts ignoring instructions, making weaker decisions, and producing lower-quality output because too much competing information is in play.

It is not a strict token number. It is a practical warning label for sessions that are still running but no longer thinking clearly.

There's a positional effect too — models tend to use information near the beginning and end of the context window more reliably than what's buried in the middle. Researchers call this "lost in the middle." The takeaway is the same: structure your context with care, or engineer it down before you get here.