Part 1 of 4 · 1 min read
Overview
An agent's history grows on every step (module 3), so the window it sends is long by default. A text can fit inside the window and still be read badly. Published long-context tests report, in outline, that recall of a fact can fall as the input gets longer, can be weaker for a fact in the middle than at the start or end, and can fall further when similar-looking passages sit nearby. These are reported findings; they will not match your model, which is why this module is about measuring: building a test for your own model and window, and using the result to decide how much to put in.
About the example, and what is real
Everything runs offline in code/mini-longctx. Real: the haystack builder (text of an exact token length, a needle at an exact depth, optional distractors and a repeated needle), the ExtractiveReader that finds the needle by pattern, the harness (grid, recall, effective length), and 48 tests. Imitated: the ToyReader. Its chance of finding the needle falls with length (1 / (1 + (length / 24,000)^2)), dips in the middle by 25 percent, and loses 15 percent per near-miss distractor. These numbers are mine. The reader measures position and length from the text it is given and then draws a deterministic pseudo-random outcome. No real model is called, so nothing here is evidence about a real model. To get evidence, replace the reader with one that calls your model (Hard 1).