Instrument

The window

A context window is quoted as one number and behaves as two. You are billed for every token you put in. The model reaches the ends of the prompt reliably and the middle much less so, and the wider the window, the deeper that sag goes. Everything on this site sits downstream of that gap: retrieval exists to keep the window small, caching to make it cheap, memory to decide what earns a place in it.

Drag the diamonds to repartition the prompt. The top bar is what you pay for, the bar under it is what the model gets to. Arrow keys work, and shift moves in fives.

Window

recall by position — 100% to 30%worst 64%

paid above, reached below

segment	paid	reached	retained
system + toolsorchestration	13k	12k	97%
retrievedretrieval	67k	48k	73%
historymemory	41k	33k	81%
taskorchestration	7.7k	7.6k	99%

billed: 128k
reached: 102k
paid per token read: 1.26×

21% of this window is bought and not reached. Moving a segment towards either end is free; the middle is where tokens go to be ignored. The curve is a stylised reading of published needle-in-a-haystack results, not a measurement of any one model — treat the shape as the claim and the numbers as illustration.

Position is one term; there are others this chart does not carry. The nomogram holds the cost and memory relations that decide how much you can afford to put in the window in the first place.