Retrieval accuracy is not uniform across a long context — material in the middle is recalled less reliably than material at either end. Build the experiment on your own documents so the design implication is yours, not a slogan.
What to build
- Assemble a document set large enough to fill a substantial fraction of the window.
- Plant twenty distinct verifiable facts at known positions: start, quarter, middle, three-quarter, end.
- Ask a targeted question about each fact and score recall.
- Plot accuracy against position.
- Now restructure: put the critical material at the boundaries and re-measure.
- Then retrieve only the relevant chunks instead of loading everything, and measure again.
Done when
- You have a position-versus-accuracy result from your own run.
- The restructured layout measurably improved recall of critical facts.
- Selective retrieval beat full loading on both accuracy and cost.
- You can explain why this argues for retrieval over ever-larger context.
The trap this exercise teaches
Assuming everything inside the window is equally available. It is present, but presence is not attention.
Before you start
The theory behind this build is covered in Long-context effects. If any step below is unfamiliar, read that first — the exercise assumes it. Primary source: Claude Docs — Long context.
Check yourself against the exam
This exercise sits in Domain 5, which is 15% of the CCAR‑F exam. Once you have built it, run a domain drill in the study app and see whether the questions read differently.