How sure is the desk?

Regular-season games, 2025–26, from the MLB Stats API. Log attendance on promotion flags, with club-season, weekday, month, day/night, temperature and rain controls. Errors clustered by club. Ranges are 95% intervals.

The effects mostly held from one season to the next: 11 of 12 2026 estimates fell inside their 2025 ranges (outside: minor league bobbleheads). They don't make single-game forecasts much better, though: fitted on 2025 and scored on 2026, the typical error moved from 4,807 to 4,767 fans (MLB) and 912 to 914 fans (Minor leagues). Promotions shift the average; weather, opponents and everything else still dominate any one night.

How the plan compares. “Your 2026 way” places exactly the same nights as the Matchday plan, on the weekdays the club used in 2026, so only the dates differ. 2027 start times aren't published yet, so weekdays a club played mostly in daylight in 2025–26 (usually Sundays) count as day games and get no fireworks.

Labels are the presenting sponsors clubs published for their promotions. A ranked brand counts when its name matches a real presenter. Top 300 brands per club, 2,000 bootstrap resamples of clubs.

• The pre-registered test failed. On 114 clubs and 3,065 real 2026 sponsors, the Desk ranking found fewer than the assistant alone (-112, range -158 to -66). It beat a placebo by 105 (range 82 to 131), so the taste signal is real; most real sponsors are local banks, hospitals and dealers that an assistant can name and Qloo doesn't cover.
• The two find different sponsors, so a hybrid test was written down first and run on a different season. On 110 clubs' 2025 sponsors the hybrid found 29 more than the assistant alone at the same list length (range 12 to 48). That is why the sponsor board shows both lists.
• The live board leaves alcohol off. The tested lists included beer and wine brands; the board you see drops alcohol, gambling, tobacco, cannabis, crypto and firearms brands.

Written down before it ran (docs/eval-plan.md section 3, with an amendment made before any outcome existed). 354 theme nights in 2025–26 were built on 26 family-safe subjects that three or more clubs used, such as Star Wars, Bluey and Harry Potter. Each night's crowd is measured against what the attendance model expects for a theme night at that club on that date, and compared only with nights on the same subject elsewhere, so national popularity drops out.

It didn't show. How the area around a ballpark leaned toward a subject didn't reliably predict which of its theme nights drew more (+0.08, range -0.058 to 0.216), and neither did the assistant's ratings (-0.05) or the town-taste score (+0.03). Single nights are noisy: even the placebo landed at -0.15. The desk still suggests theme ideas this way, as a list of things each town likes, and every club page says the ranking is untested.

• Lifts are observed associations; clubs choose when to run promotions, so they are not controlled experiments.
• Sponsor labels show who bought, not who should. A suggested brand may never sign.
• Many sponsors renew every year, so the 2025 replication overlaps the 2026 labels; it is a second season, not an independent sample.
• The assistant may remember published sponsor deals from its training data, which favours it in both tests.