# E21j: multikey needle (one target passcode plus three decoys) in real math text, Qwen3.5-4B from 128K to 512K and Qwen3.5-9B at 480K on one RTX 3080 10 GB

- `e21j_chain.ps1`: the run script (Windows PowerShell, llama.cpp server built with upstream PR 29827), tidied: two unused variables and a raw-reply debug field were removed.
- `results.jsonl`: one line per run. The grader takes the first NNNN-NNNN code in the final answer, after any thinking block, and records which decoy it matched, if any. Answers are shortened.
- The haystack is paragraphs from the MATH training set (problems and worked solutions), shuffled with a fixed seed; that data file is not included here.