A previous step of the llm-coding-inference programme sized the human audit of AI-coded text from a small pilot, and found it hit the target interval width only about half the time. This step tests the alternative: hand-code in batches and stop the first time the estimated half-width reaches 3 points.
On the NaijaSenti corpus (four Nigerian languages, three stand-in coders, 1,000 simulated audits per cell), starting from 20 tweets per language and adding batches of 30 split across languages by the current variance estimates:
- all 1,000 audits in every cell stopped at or under 3.0 points;
- the mean audit was 898, 1,074 and 919 tweets against 902, 1,083 and 927 for a rule that knows the true error rates;
- coverage of the 95% interval was 94.2 to 95.2%, with no correction for stopping early;
- the audit size varied less than under pilot planning (sd 34 to 48 against 79 to 119 tweets).
Batches of 90 cost about 30 more tweets. The price is operational: the audit has to be hand-coded in rounds. Limits: one corpus, one 3-point target, stand-in coders and no real LLM; the coverage result says optional stopping did not hurt here, not that it never will.