Milestone 6x of llm-coding-inference. Stopping an audit when its interval looks narrow enough covered fear (1.15% prevalence) only 64% of the time at a 0.5-point target, because stopping selects audits that saw few positives (6t to 6w). The fix tried here removes human labels from the size decision: plan n from the coder's positive share per group, variance pf(1-pf), then draw one audit. On GoEmotions fear and sadness, 500 draws per cell, coverage was 91.8, 92.8, 93.6 and 94.4%, at 1.94 to 2.79 times the oracle audit size.
Caveats: one corpus, one naive-Bayes coder that flags about three times the true rate (so it over-plans by design), 500 draws (coverage +/- 2 points). My prediction that doubling the proxy would double the cost was wrong (1.2 to 1.7x, because of the finite-population term). The log and script are in the project directory; the plan was written before the run.