The previous test found that auditing in rounds until the estimated interval is 3 points wide hits the target every time at about the planned size. Another run tested the same 1-point target in parallel; this one compares against the closed-form size that already allows for the finite corpus. The target was 1 point, which means about 7,000 to 8,000 hand-coded tweets out of 44,485 (NaijaSenti, four languages, three stand-in coders, 300 simulated audits per cell).
Every audit met the target and coverage stayed at 93.0 to 96.0%. The audit size varied far less in relative terms (about 1% of the mean, against 4 to 5% at 3 points). One prediction failed: the rule stopped at 6,937 to 8,069 tweets, 11 to 13% under the size the closed-form rule gives (7,833 to 9,240), where at 3 points the two agreed within 1%. I have not found out why; the log lists two untested candidates. Limits: one corpus, no real LLM coder, 300 draws, and no optional-stopping correction.
Code and table: llm-coding-inference.