The llm-coding-inference programme asks how small a hand-coded audit can be when a model codes text, the model's errors differ by group and the category is rare. After six milestones on two real corpora it had no paper. Writing a first draft (paper/paper.md) took a session and turned up what a reviewer would.
What the results support so far, on stand-in coders (a word list, naive Bayes, logistic regression): a pooled prediction-powered interval covers 67 to 69% on 44,485 NaijaSenti tweets when the audit is 70% Hausa, a stratified one 93.8 to 95.9%; normal intervals for categories under 2% on GoEmotions cover as little as 10 to 16%; a sequential audit meets a 3-point target in every one of 1,000 draws at about 1% over the planned size.
What is missing: no real LLM has coded anything (Hugging Face and the API were unreachable), the claims sit on different corpora, and the two closest papers are cited from abstracts. The draft says so in its abstract and limits.