Written by an AI agent working in this repository.
Earlier milestones of the llm-coding-inference programme found that, for a GoEmotions category near 8% prevalence (admiration), splitting the human audit into "coder said yes" and "coder said no" cells with a floor on each cell's share gave intervals about 21% narrower at n = 1000. This run asked whether that carries to a second corpus: NaijaSenti, target = negative share (32.7%), three zero-shot Qwen2.5 coders (0.5B, 1.5B, 3B).
It does not, and the reason is the coder, not the rule. The three coders agree with the human label on 66 to 69% of tweets; always answering "not negative" scores 67.3%. With no signal, the oracle two-cell design is exactly as wide as one pooled stratum (0.0 to 0.2% difference). The pilot-with-floor designs are worse: 22% wider at n = 400 for the 0.5B and 3B coders, because the floor puts 35% of the audit into a cell holding 2.5% of the corpus. Coverage was 93.8 to 97.4% in all floor cells; pooled intervals were slightly short at n = 400 (91.6 to 92.6%).
Practical reading: measure the coder's informedness before stratifying on its cells; if it is near zero, use one pooled stratum. The positive result still rests on one corpus, one category and one model. The open step is an informative coder on a second corpus.
Code, raw table and the pre-registered predictions (two of four refuted) are in research/llm-coding-inference/ (LOG.md, milestone 6aj; experiments/h6aj_naija_floor.py).