The needle in the haystack

What you'll walk away with: why a fact you definitely pasted can go unused, and the three moves that answer it — seat it at an edge, prune what looks like it, and make the reply quote the line it used.

You need: both tabs, about 20 minutes, one long document you can paste as text, and one sentence you make up. The demo runs in your browser.

Predict first: a contract's liability cap sits on page 2 of a 120-page bundle you pasted whole. Same cap on page 60. Same cap on the last page. Which seat does the reply use most reliably?

Lesson 07 was the pile growing. This lesson holds the pile still and moves one fact around inside it: where it sits and what sits beside it are yours to decide before you press send.

Same fact, different seat

It opens on the seat the demo flags as vulnerable — 40 pages, the fact in the Middle, nothing around it that resembles it; leave Document length alone. Under Where the fact sits, click Start, then Middle, then End: the amber cell moves — fact on page 1 of 40, then 20, then 40 — and the label beside Position guidance changes with it, one reading for the middle, another for either edge. Now switch What surrounds it from Unrelated filler to Near-identical clauses. The seat does not move; the rest of the strip goes amber behind it, a second legend appears — pages that resemble it — the needle stops standing out — and the panel says: resemblance competes hardest; pruning the lookalikes helps more than moving the needle. Before you take a number from it, read the panel's last line.

That last line is the whole caveat. The demo reports a position effect someone measured, on the questions they set, on the models they had; it is the reason to run your own document, not a number to carry into it. What does not move is the mechanism underneath: retrieval from the window is not lookup. Every token competes for the continuation's attention, and a fact drowning in text that resembles it — ten near-identical clauses, eight drafts of the same table — competes hardest of all. You cannot change how the odds weigh the window, and you do not need to: what goes in, and where, is yours.

Being in the window is not being used

Lesson 02 put two steps between you and the window: you decide what is eligible, the app decides what makes the cut. The third catches professionals: making the cut is not being used. A fact in the window is available to the continuation — not evidence it was drawn on.

So on any answer that matters, ask for the answer and the sentence it came from, word for word. A quote that matches your line is the line the reply reached for. One that is close but not exact, missing, or replaced by a summary is an answer assembled from something else — and you found out before you forwarded it. It is also how you tell whether the next two moves worked.

Curate, don't relocate

Two moves fall out of the demo, and they are not the same move:

• Put the load-bearing fact where it can be seen — near an edge, the start or the end of what you paste, not buried mid-pile — and say that it is the load-bearing fact.
• Prune the lookalikes. The buried fact's worst enemy is similar text it does not need beside it. Delete the haystack you aren't using; don't just relocate the needle.

The second is the one people skip. Context rot keeps the general rule in one line — more context is not automatically better context — and the practical form is curation: decide what the window needs before you paste, not after the answer disappoints. Paste as text rather than attaching, so you can see what went in and where — lesson 04 is why, loading the window the reference. When the pile is a long thread you do not control, lesson 07 has the move: leave it and carry a clean summary into a fresh window.

Do this in ChatGPT and Claude

The haystack: anything long you can copy as text — a terms page, an agenda pack, a month of your notes — long enough to scroll through, and nothing confidential: it goes into two companies' chat boxes.

The needle: one sentence you write yourself — a line you will recognize, from facts not in the document, checkable at a glance; not a real person, not a real figure. Use this one (the names are made up), or your own with the question below changed to match:

Site access for the Ellsworth audit is granted by Priya Raman, and the window closes on the 14th.

Keep the document in a scratch file so you can move the needle around in it; the paste is the only thing that changes. Predict first, in writing: which seat comes back cleanest, and does the quote match yours exactly?

Chat A — top. Fresh chat, one message: the needle as the first line, the whole document under it, and this request at the end:

Who grants site access for the Ellsworth audit, and when does the window close? Quote the sentence you took it from, word for word.

Chat B — middle. Fresh chat, same message, the needle moved to roughly the middle. Same request, unchanged.

Chat C — bottom. Fresh chat, same message, the needle as the document's last line — also the line nearest the ask, so edge and nearness travel together here. Same request again.

Chat D — lookalikes. Fresh chat, needle back in the middle with three sentences of the same shape around it, each about a different audit with its own name and date — Site access for the Marlow audit is granted by Daniel Okafor, and the window closes on the 9th, and two more like it. Same request one last time.

Observe, two things per chat, written down: did the answer carry both facts, and did the quote match your sentence character for character. If all three seats came back clean, that is a real measurement on your own material — at this length, in this app, one planted line came back from any seat — and the reason to try a longer paste; chat D tests the other half regardless, where the seat is chat B's and only the neighbors change. There, read the quote against all four lines. Your line exact: the needle held. A decoy quoted whole names another audit — a miss you can see. A reply that notices the lookalikes and says so is the quote check doing its job. And a blend — parts of two lines fused — is the failure signal this lesson is about: resemblance competing at that seat.

Explain: same machine, same question, same document; only position and neighbors moved, and both were yours. If a reply names something nowhere in your paste, two things can have happened, and the quote tells you which: the paste did not arrive whole — lesson 02's fine print — or the reply filled the hole in the request's shape, which is lesson 10. Either way, seat the load-bearing line where you watch it go in.

Then the same four in the other tab. What repeats across both is the shape to plan around; what differs is one product on one day — never a ranking of seats.

---

Check your understanding: the fact was definitely in the window, and the reply still missed it. Name the two things about the paste that could have done it, and the move that answers each.

Check yourself first — then open one good answer+−

Where it sat, and what sat beside it. A fact buried mid-pile has the window's whole length competing with it for the continuation's attention; a fact ringed by near-identical clauses has competition that looks like the answer, which is the harder of the two. The moves: seat the load-bearing fact at an edge of what you paste and say that it is load-bearing, and prune the lookalikes — delete the haystack you are not using rather than only moving the needle.

Your answer should have hit position and neighbors as two separate causes, and curation as the move that outranks relocation.

Use this before you paste anything long. Decide what the window actually needs, put the load-bearing line where you can watch it go in, and ask for the sentence it used — being in the window is not being used, and the quote is how you tell. Next: Steering the average — why a sharp ask comes back bland.