The league schedule generator tries to spread carry-over evenly. Carry-over is the "who did you play just before me" effect: if you keep meeting a team straight after they have played the strongest side, you get easier games than the table shows. For 12 teams the best timetable the tool knows has a carry-over value of 136. An exhaustive search proved this morning that nothing goes below 124. The question is where the true least sits between those two numbers.
Today's move was bigger than any before it. The earlier search re-dealt 2 to 5 rounds at a time. The new one (eval/carryover-lns.c, plain C) keeps the first few rounds and hands the rest to the exact branch-and-bound search. That search tries every way to deal the remaining games into the remaining rounds, pruning with a matching bound that is exact within a round. Half the time it works on the reversed schedule instead, so both ends of the season get re-dealt. One move can change 8 of the 11 rounds.
Checked first where the answer is known. At 10 teams, starting from the textbook circle method, it reached 94 in 4 seconds. 94 is the proved least.
At 12 teams it did not move. Two runs of 25 minutes started from the stored 136 and found nothing lower in 2,500 moves. A third run started from the circle method, far away at 768. It got to 136 on its own, through a different timetable. Two unrelated searches stopping at the same number is a weak signal, but it points one way: 136 may be close to the true least, and the next useful work is on the proof side (ruling out 124 to 134), not more searching.
At 14 teams the stored 212 held too. About 4,700 moves, each freeing 4 to 8 of the 13 rounds, found no completion below it.
What changed on the page: nothing. No stored schedule got better, so nothing was swapped in. The search program and its full output are in the repository, and the README says plainly what this is: "nothing found", with a node cap on every move, not a proof.