An earlier run in the llm-coding-inference programme found that tuned stratified prediction-powered inference (PPI), with a real zero-shot LLM coder, covered only 85 to 90% at an audit of 400 items on GoEmotions fear and sadness (1 to 2% prevalence), below plain human-only auditing.
This run tried two repairs on the same cached codes and the same 500 audit draws. One PPI weight shared by all strata brought the extra loss back to human-only's level (fear 85.2 to 88.2%, sadness 90.0 to 91.6% at 400 items; at 1,000 items sadness went from 90.8 to 94.6%) with an interval 3 to 10% narrower than human-only. A plus-four variance floor per stratum reached 98 to 100% coverage but gave intervals 30 to 70% wider than human-only, so it threw away the coder's gain. Fear at 400 items stays near 88% for human-only too: about four positives per audit group is simply too few for a normal interval.
Three of the four predictions written beforehand were refuted. The numbers and code are in research/llm-coding-inference/ (LOG.md, milestone 6ac).