In llm-coding-inference we ask how few hand-coded items you need when an LLM codes the rest. Three zero-shot Qwen2.5 models (0.5B, 1.5B, 3B) agreed with the human label 66 to 69% of the time on NaijaSenti, yet prediction-powered intervals came out only 0.4 to 2.2% narrower than hand-coding alone at n = 400.
Earlier steps blamed error structure. The new step (5h) shows a simpler cause. Negative tweets are 32.7% of the corpus, so a constant "not negative" code already agrees 67.3%. The 3B model's sensitivity is 0.072 and specificity 0.989: Youden's J of 0.06, nearly a constant. A synthetic coder with each real coder's sensitivity and specificity and independent errors reproduces the real width ratio to within 0.006 in all six cases, while a symmetric coder at the same agreement (J 0.40) gives 0.937. A 20% gain needs J around 0.5.
Caveats: one corpus and prompt, the best-J cut is chosen on the items it is scored on, and a 7B or fine-tuned coder is untested. Report J, not agreement, when you pick a coder for this. Table: research/llm-coding-inference/experiments/h5h_informedness.txt. Written with AI assistance.