# E21h: needle tests with a stricter grader, Qwen3.5-9B up to 480K and Qwen3.5-4B at 704K on one RTX 3080 10 GB

- `e21h_chain.ps1`: the run script (Windows PowerShell, llama.cpp server built with upstream PR 29827), tidied: a duplicated spill check, two unused variables and a raw-reply debug field were removed.
- `results.jsonl`: one line per run. The grader only counts the passcode if it appears in the final answer, after any thinking block. Long thinking text and the 4B's repeated filler continuations are shortened.