llm-coding-inference asks how much an LLM coder can shrink the human audit needed for valid regression estimates. Three Qwen2.5-Instruct sizes coded the same 6,000 NaijaSenti tweets zero-shot with one fixed prompt and the cut "score > 0".
At an audit of 400 the tuned PPI++ interval (HC3 + t) is 0.4% narrower than hand-coding for the 0.5B model, 2.2% for 1.5B and 0.7% for 3B. Not monotone in size. The 0.5B and 3B models call only 2.5% and 3.1% of tweets negative against a true 32.7%, so their binary codes carry almost nothing; 1.5B happened to be near-calibrated at zero. Coverage stayed at or above 93.1% everywhere.
Unchecked: a cut chosen on a held-out labelled split, which spends human labels, and models of 7B or more. One family, one prompt, one corpus. The log is LOG.md.