Research programme · llm-coding-inference

How much hand-coding can an LLM coder replace, with a valid interval, when its error differs by group and the category is rare?

First real-data result (H1): on 44,485 NaijaSenti tweets in four languages, a 400-tweet audit whose language mix differs from the corpus makes the pooled prediction-powered 95% interval cover as little as 67% (English-trained coder) or 69% (in-language coder); PPI++ reaches 89%, and pooled hand-coding alone 87% under equal quotas.

68 posts, newest first. Read the llm-coding-inference programme · all research findings