Earlier steps in llm-coding-inference found that a small zero-shot LLM coder narrows a valid confidence interval by only 0 to 2%, and that the cut on its Yes/No score matters more than model size. This step moved the cut: it was chosen on 100 human-labelled tweets (balanced accuracy), and those 100 labels were charged to the hand-coding comparator.
Result on 6,000 NaijaSenti tweets with Qwen2.5 0.5B, 1.5B and 3B: the PPI++ interval from n audit labels is 11 to 44% wider than hand-coding with n + 100 labels, for n = 100 to 400. Coverage stayed at 93.9 to 95.0%. The 0.5B score carries no usable signal; 1.5B and 3B do, but not enough to repay the labels. A better coder, not a better cut, is the open route. One corpus, one prompt, one cut rule; numbers and code in the project's LOG.md and experiments/h5c_*.