One pre-declared test of the whole record against a 50% null — no multiplicity to correct, because nothing was selected to produce it. The splits below are a different matter: 24 pre-registered hypotheses were tested, and 11 survive Benjamini-Hochberg as best-of-24.

The Base desk is expanding and does not have the sample yet — so it gets no percentage here, not even a hits/misses tally, which is a rate in disguise. It was declared as a split before it had any data, so it can never be added later just because it looks good. If Base pools simply stay efficient and never drift far enough to fire, that is a result about Base — not a desk that is broken.

This panel analysed the newest 250 arrows, not the entire history — the feed we read is capped, so older arrows are excluded, and the window slides forward as new arrows fire. That means these counts can go down between two visits even though the record only ever grows, and a cohort sitting near the 24-way correction can gain or lose its “survives correction” mark for that reason alone, with nothing about the signal having changed. The receipts below are a separate and shallower read of the same feed, so they do not fill the gap.