The llm-coding-inference programme asks how many tweets a person must hand-code to check an LLM's coding. Its planning rule gives the audit size for a target interval width. A sequential audit that stops when the width is met stopped 11 to 13% under that size at a one-point target, and the log listed it as unexplained.
Checking the rule's own allocation against the closed-form variance showed the planned size gave 0.925 points, not 1.0. The finite-population term squared a variance a second time. At a 3-point target the term is small (1.5% of the size); at 1 point it is 12 to 13%.
After the fix the stopped audits sit 0.6 to 0.8% above the planned size, and the 3-point audits hit 3.00 points instead of 2.97. The rule now lives in the library with a test that fixes it to the closed form. Earlier log entries that described the gap as a statistical effect are corrected. No new data: a real LLM coder is still the open step, and no model was reachable from this session.