Written by an AI agent working in this repository. Results are from research/llm-coding-inference (milestone 6ai).
Splitting an audit by the LLM coder's own label (coder-positive versus coder-negative) narrows the interval 15 to 25% when the allocation uses each cell's true variance. The previous step showed that estimating that variance from a 30% pilot starves the rare cell and collapses coverage. This step tried two repairs on GoEmotions (6,683 items, zero-shot Qwen2.5-1.5B, 500 draws per cell).
- A floor on each cell (20% or 35% of the audit) fixes the common category: admiration (8%) reaches 93 to 94% coverage at n = 1,000 and is 21% narrower than one pooled stratum. Fear (1.2%) and sadness (2.3%) stay invalid, 50 to 92%.
- An independent pilot (used only to plan, not in the estimate) is valid at admiration but the pilot items cost exactly what the allocation saves: no gain at all.
So for rare categories, pool; for an 8% category with an informative coder, use coder cells with a 35% floor and n of about 1,000. One corpus and one small model, so treat the floor as a lead. Log and code: research/llm-coding-inference/LOG.md, experiments/h6ai_floor_indep.py.