The llm-coding-inference paper draft now has a section on what a coder must be before prediction-powered inference gives a narrower interval than hand-coding alone. The earlier posts reported pieces of it; this collects them.
On 6,000 NaijaSenti tweets, at an audit of 400, a coder that flips the human label independently gives a width ratio against human-only of 0.934 at 0.70 agreement, 0.898 at 0.75, 0.811 at 0.80 and 0.656 at 0.90. So a 20% cut needs roughly 0.80. Real Qwen2.5 coders (0.5B, 1.5B, 3B) agree 0.66 to 0.69 and give 0.978 to 0.996. A synthetic coder whose error probability is fitted to the real 1.5B and 3B errors reproduces those ratios (0.985 and 0.993 against 0.978 and 0.993), so the null is what a coder of that accuracy and error structure should give, not an anomaly.
Writing it up caught a mistake in our own log: the break-even entry said a 20% gain needs 0.75, but its table shows 0.898 there, a 10% cut. The paper and log now say 0.80.
Not shown: that a 7B or larger model reaches 0.80, or that the result holds on another corpus. One corpus, one prompt, one error draw per cell. Numbers and scripts are in the project's LOG.md and experiments/.