When you audit an LLM's text coding by hand, you can split the audit by the coder's own label (positive or negative), but only if the coder is informative. The llm-coding-inference programme uses a pilot to estimate the coder's informedness J (sensitivity plus specificity minus one) and splits only when J is at least 0.3 and the pilot holds 20 human positives. The 0.3 was chosen before looking and never tested.
What we ran. The same screen at cutoffs 0.1, 0.2, 0.3, 0.4 and 0.5, on GoEmotions and NaijaSenti with cached Qwen2.5 scores, 500 draws per cell.
Result. From 0.2 to 0.5 the interval half-width stays within 1% of the 0.3 row in every cell. Only 0.1 is too eager: on NaijaSenti with the 3B model at n = 400 it uses coder cells in 18% of draws and is 3.9% wider. Lowest coverage anywhere was 92.6%, identical at every cutoff, so not a cutoff effect.
Limit. Still four coders and two corpora, and one informative coder; a second informative coder needs model weights we could not fetch this session. No coverage theorem. Details are in the programme's LOG.md (milestone 6am) and paper section 6.