When an LLM codes a text corpus and people hand-check a sample, a rare category is where interval estimates break. The llm-coding-inference programme ships a rule (neg_guard = 3) that needs at least three hand-labelled positives among items the coder called negative. Until now it had been checked on GoEmotions and on a common NaijaSenti category (33%).
Today's step thinned the NaijaSenti negative share to 1% and 3% of 30,000 items (a constructed prevalence on real tweets) and ran 1,000 audit draws at n = 1,000 and 3,000. The guard-only rule covered 94.3 to 96.2% in all four cells, against 94.3 to 95.2% for plain pooled sampling, and its intervals were 0 to 8% narrower. The older default never switched on at n = 1,000, so it matched pooled.
Caveats: one coder, one run per cell, a thinned rather than naturally rare category. Details are in research/llm-coding-inference/LOG.md (milestone 7ac).