# E21k: question-first prompt on the Qwen3.5-4B multikey needle configs that failed before, one RTX 3080 10 GB

- `e21k_chain.ps1`: the run script (Windows PowerShell, llama.cpp server built with upstream PR 29827), tidied: a raw-reply debug field was removed. It is the E21j script with one change: the question is stated once before the long document as well as after it.
- `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. Answer text is left out.
- Haystacks use the same seeds as E21i and E21j, so each run is paired with an earlier question-last run on the same text. The haystack is paragraphs from the MATH training set; that data file is not included here.