One thing the Adam test can't settle: with the question in one fixed spot, "middle of the document" and "far from the question" are the same address. Move the question to the top and the curve might flip. The post's own two explanations — early tokens getting more routes through the layers, late ones sitting near the answer — point to different fixes: reorder the documents, or move the question.
The capacity vs reliability framing is the useful one. A 128K window is storage; attention is the query plan. One practical pattern that has helped for coding agents: put the critical constraint twice — once in a short header block and once as an explicit check in the final request — rather than relying on a single middle mention. Curious whether you've measured how much brief repetition helps vs just relocating the fact to the edges.
The 37 day audit log example is the one to keep. That failure passes every retrieval check because the rule was in the prompt, so the eval has to assert on the output, not on what got retrieved.
A context window is often marketed as capacity, but users experience it as reliability. The difference matters commercially: storing a rule inside the prompt is not the same as governing with it. Once omissions affect audit retention or permissions, context design becomes control infrastructure, not prompt craft.
The same positional bias shows up in how clients retain advice. They hold on to the opening framing and the closing action item, and the load-bearing middle detail — the lock-in period, the 80C ceiling, the term cover multiple — is exactly what gets dropped.
Which is why I repeat the critical constraint at both the start and the end of a conversation. A single clear mention is never enough.
Same reason a bigger context window doesn't fix it is why a longer, more thorough client meeting
"Early information has more opportunities to influence later processing, while information near the end can benefit from its proximity to the question and the answer. Information in the middle does not benefit from both."
As i see, the tokens at beginning of the sentence can contribute to all tokens after it, so it has a great chance here to gain more influence on the whole generated contents. But for the tokens at the end of the sentence, i don't see any advantages here. The contribution of the tokens at any position will contribute equally to the next predicted token.
I am not convinced by the explanation in this sharing yet, while being impressed by the "lost-in-the-middle" phenomena
One thing the Adam test can't settle: with the question in one fixed spot, "middle of the document" and "far from the question" are the same address. Move the question to the top and the curve might flip. The post's own two explanations — early tokens getting more routes through the layers, late ones sitting near the answer — point to different fixes: reorder the documents, or move the question.
The capacity vs reliability framing is the useful one. A 128K window is storage; attention is the query plan. One practical pattern that has helped for coding agents: put the critical constraint twice — once in a short header block and once as an explicit check in the final request — rather than relying on a single middle mention. Curious whether you've measured how much brief repetition helps vs just relocating the fact to the edges.
Would external, deterministic validation help lift the middle?
The 37 day audit log example is the one to keep. That failure passes every retrieval check because the rule was in the prompt, so the eval has to assert on the output, not on what got retrieved.
Join SQL WhatsApp group
This explains so much. I started putting the critical instructions at both the top and the bottom of long prompts, and the difference was obvious.
A context window is often marketed as capacity, but users experience it as reliability. The difference matters commercially: storing a rule inside the prompt is not the same as governing with it. Once omissions affect audit retention or permissions, context design becomes control infrastructure, not prompt craft.
The same positional bias shows up in how clients retain advice. They hold on to the opening framing and the closing action item, and the load-bearing middle detail — the lock-in period, the 80C ceiling, the term cover multiple — is exactly what gets dropped.
Which is why I repeat the critical constraint at both the start and the end of a conversation. A single clear mention is never enough.
Same reason a bigger context window doesn't fix it is why a longer, more thorough client meeting
Can someone explain this concept technically?
"Early information has more opportunities to influence later processing, while information near the end can benefit from its proximity to the question and the answer. Information in the middle does not benefit from both."
As i see, the tokens at beginning of the sentence can contribute to all tokens after it, so it has a great chance here to gain more influence on the whole generated contents. But for the tokens at the end of the sentence, i don't see any advantages here. The contribution of the tokens at any position will contribute equally to the next predicted token.
I am not convinced by the explanation in this sharing yet, while being impressed by the "lost-in-the-middle" phenomena