The league schedule generator can plan three games a night so that no team waits between its games. For odd leagues it stores seasons a SAT solver found. Each one is a round robin plus a time slot for every match, with each team's slots in a row. At nine time slots those runs stalled at 19 teams and up: 21 teams on four courts ran past 20 minutes, 23 teams past 25. On the page that meant 21 teams on four courts had 102 empty slots between someone's games over three test seasons, and 23 teams had 184.
The trick is to stop asking for the whole season. Take the round robin where round r is round 0 with every team number shifted by r (mod n). Shifting all teams by one turns any night into another night with the same shape. If the night made of rounds 0, 1 and 2 can be laid out, so can rounds 3, 4 and 5, with the same rows and different names. So the solver only has to lay out one night, about 30 matches instead of a few hundred. Then a short search fits copies of that night together to cover the season.
That gives this:
| Teams | Courts (9 slots) | Waits before (3 seasons) | Waits now | Solve |
|---|---|---|---|---|
| 21 | 4 | 102 | 0 | 0.2 s |
| 23 | 4 | 184 | 0 | 221 s (5 starters) |
| 27 | 5 | 231 | 0 | 0.3 s |
It worked at six slots too: 27 teams on seven courts, which the whole-season run had left after 40 minutes, went from 201 waits to 0. The tenth starter tried had one night pattern, and nine shifted copies of it make up the whole season.
The plain circle-method round robin worked at once for 21 and 27 teams. For 23 it didn't: no night of it lays out on four courts. The circle method is one starter, a way of pairing up the numbers 1 to n − 1 so that every difference turns up once, and any starter gives a round robin with the same shifting property. The fifth starter tried had two night patterns that lay out, and they fit together.
Where it doesn't work. When every slot of the night is booked (19 teams on three courts, 25 on four), no night of any starter I tried laid out. At 19 teams that was 42 starters and every pattern of each. The old whole-season solver can still find one such night on its own, in 33 seconds, so a wait-free season for 19 teams on three courts isn't ruled out. What's ruled out is this shortcut, for those starters. The same thing shows at six slots, where the fully booked 13 and 17 team cases need the whole-season model.
So the method is good when a night has a few spare slots, and it found the stored 15 and 17 team designs again in 0.1 and 22 seconds, against 51 and 245 for the whole-season run. The script, its measurements and what's still open are in the README and eval/odd-cyclic.txt.