The league schedule generator uses simulated annealing to find round robins with little carry-over, which is the unfairness of always meeting a team right after the same opponent. Each step tries a small change to the season, prices it, and keeps it or not. On a 9-team league that is about 1.3 million priced moves per try (it anneals from more than one start), and the venture's test suite runs many tries.
The waste
Annealing keeps a worse move with probability exp((cur - v) / T). In code:
const v = cost(season);
if (v <= cur || rand() < Math.exp((cur - v) / T)) cur = v; // keep
else undo();
Rearranged, a move is kept exactly when v < cur - T·ln(u), where u is the random number that rand() is about to return. Late in the search most moves are rejected, and each one paid for a full count over every team's season to learn that.
If you know u before you price the move, you know the threshold, and you can stop counting as soon as the move can't come in under it. The carry-over count is a sum over teams, and every team still to be counted adds at least one per game step, so "counted so far + the least the rest can add" is a valid lower bound to stop on.
Same answers, not just similar ones
The catch is the random stream. The search is seeded so a share link rebuilds the same season, so the stream must not change. The original only calls rand() when v > cur. So we:
- added
peek()to the seeded generator, which returns the next value without drawing it; - stop early only when the partial count already exceeds the threshold plus 1, a rounding margin. That implies
v > cur, so the original would also have drawnuand rejected the move. We draw it too, and reject it; - price accepted moves in full, as before.
Every decision is the same, so every season is the same. We checked that, not assumed it: one SHA-256 over 214 searches (3 to 16 teams, cyclic, weighted, capped, home-and-away) matched before and after. A new test pins four of them. Breaking the bound or the margin on purpose turns that test red.
What it bought
In one 9-team search, 72% of the moves stopped early, on average after 5.4 of 9 teams.
| before | after | |
|---|---|---|
| the slowest test file, alone (two alternated runs) | 119.9 s, 113.1 s | 110.0 s, 103.4 s |
| nine weighted searches, 9 to 14 teams | 28.7 s mean | 26.8 s mean |
| six break-capped searches | 6.5 s mean | 5.8 s mean |
About 8% on the search-heavy tests, and the same for anyone generating a schedule on the page. The capped search had a second, free win: a move over the break cap was rejected without a draw, yet still had its carry-over counted. Now it isn't counted.
Two cleverer versions lost. Counting the moved teams first, so a bad move shows up sooner, cost more in bookkeeping than it saved. Starting the count at one moved team made no measurable difference. The whole-suite wall time, one run each, moved by less than its own run-to-run spread, so we report the alternated per-file numbers instead.
Details and the ranked next steps: docs/history/ci-perf-2026-11-17-league-priced-early.md.