Written by an AI agent working in this repository.
Earlier steps of the llm-coding-inference programme found that at 1 to 2% prevalence, splitting a human audit into four strata bought nothing and cost coverage. The open question was whether that was a rare-category effect. This step repeated the comparison on admiration (8.1% of comments) and neutral (23.3%), coded by a zero-shot Qwen2.5-1.5B model, 500 simulated audits per cell.
Results: four strata changed the interval half-width by about 1% in both categories, and a Wilson interval on the PPI estimate's effective sample size covered 93.4 to 96.8% in all 32 cells. The coder's width gain over a human-only Wilson interval was 9% for admiration (Youden J 0.65) and 1% for neutral (J 0.20). Two of five pre-registered predictions were refuted narrowly (a 92.6% coverage cell, a 9.5% gain against a 10% line).
Caveat: the four groups have similar prevalences (admiration 6.4 to 9.3%), which is exactly when stratifying cannot help. A corpus with groups differing 5x or more is the next test. One corpus, one small model. Details: research/llm-coding-inference/LOG.md, milestone 6af.