Written by an AI agent working in this repository.
The llm-coding-inference programme builds confidence intervals for survey-style estimates when an LLM codes most of the data and humans hand-code a small audit. Its library path (plan_audit, then audit_interval) was checked against an earlier experiment on 14 corpus, coder and audit-size cells with 500 draws each. One cell, GoEmotions admiration at n = 1000, read 92.0% coverage against 94.6% in the earlier experiment.
At 500 draws, two independent coverage estimates differ by about 1.4 points (one standard deviation of the difference), so 2.6 points is roughly two sd, and it was one of 14 comparisons. A re-run with 2,000 draws gives 94.1% (standard error 0.5), half-width 1.165 points. The four blocks of 500 read 94.6, 94.6, 93.8 and 93.4%, so a single block can drift by more than a point.
The lesson is small: a lone low cell among many at 500 draws is weak evidence either way. The screening rule still has no coverage theorem, and the check covers one cell and one coder. The paper and README now state both the 92.0% and the 94.1%.
Details: research/llm-coding-inference/LOG.md, milestones 7c to 7e.