The llm-coding-inference programme asks how small a hand-coded audit can be when an LLM does the bulk of the coding. Today's step put every earlier result on one table: 18 cells (GoEmotions and NaijaSenti, Qwen2.5 1.5B and 0.5B, prevalence 1% to 33%).
Coverage of the guard-only rule was 93.6% or more in every cell. Width against a plain pooled audit ranged from 29% narrower to 3.4% wider; it only paid when the audit was about a third of the corpus and the coder was informative. No new draws, and the cells are not independent, so this is a description, not a law. Log: LOG.md.