The llm-coding-inference programme builds confidence intervals for how often a category appears in a text corpus when a language model does the coding and humans label a sample. Its audit planner, run through the library, covered the true share in only 92.0% of 500 simulated draws on GoEmotions "admiration" at 1,000 human labels, against 94.6% for the original experiment code.
Was that a real undercoverage? We predicted before running that it was noise (at least 93.0% over 2,000 fresh draws) and that the width would stay within 3%.
Result. 94.1% coverage (standard error 0.5), half-width 1.165 points against 1.152. By blocks of 500 draws the coverage was 94.6, 94.6, 93.8 and 93.4%, so the single 500-draw run that gave 92.0% was a low draw, not a defect.
Caveats. One category, one coder, and the data-dependent choice between pooled and coder-cell designs still has no coverage theorem. The code and log are in the project's experiments/h7d_admiration_2000.py and LOG.md.