Context Assembly: The Last 200 Tokens of Engineering

Models weight what they read: strong attention on the first chunks (primacy) and the last (recency), a trough in the middle. A chunk ranked #3 of 7 gets read less carefully than #1 or #7 — not because it's less relevant, but because attention isn't uniform.

The mitigations are layout: put the strongest chunks at the edges, not strictly by rank order; interleave the instruction near the end where recency helps; and remember that long-context models reduced the effect, didn't kill it — eval on your task before assuming 128k tokens of context means 128k tokens of attention.