The llm-coding-inference programme asks how to get a valid confidence interval for a rare category's prevalence when an LLM codes the text and humans audit a sample. The rule that survived the last dozen milestones keeps one safeguard: use the coder's labels to split the audit into cells only when the coder-negative cell holds at least three pilot positives. It had only been checked with one model, Qwen2.5-1.5B.
I scored the same 6,683 GoEmotions test comments for fear (1.15%) and sadness (2.3%) with Qwen2.5-0.5B-Instruct, a coder with a Youden J of about 0.26, and ran the same four cells (two categories, audit sizes 500 and 2,000, 500 draws). I wrote the prediction first: coverage 93% or more everywhere.
It held. Coverage with the guard alone was 95.4 to 96.6%. At n = 2,000 the intervals were 9.6% (fear) and 6.0% (sadness) narrower than the pooled design; at n = 500 they were 1.9 and 3.4% wider. The gain is smaller than with the 1.5B coder (17% and 29%) but not gone.
Limits: two categories, one corpus, one model family, one run per cell, so coverage is only good to about 2 points. The library default is unchanged. A second corpus is the next check.
Code, outputs and log: research/llm-coding-inference/ (milestone 7aa).