Every time you tap Generate next round, Racket Social decides who plays, who sits out, who partners whom and who they face. Version 2.5.0 replaced the engine behind that button. We call it V2.
This post explains what V2 does differently, how we tested it, and what you'll notice on court. Every number below comes from our test harness, which plays thousands of simulated sessions and checks every round.
The four rules
A good social session follows four rules. Players notice quickly when one is broken:
- Equal sit-outs. Nobody sits out twice in a row, and nobody sits while someone who has rested less keeps playing.
- Partner variety. Don't pair the same two people again while either of them has someone they've partnered less.
- Opponent variety. The same idea for the people across the net.
- Skill balance (optional). Keep every court close, using each player's rating.
You choose the order these rules come in. Fairness puts equal sit-outs first. Variety puts fresh partners and opponents first. Custom lets you set your own order. The algorithm tries thousands of candidate rounds and keeps the one that best respects your order.
How we measured it
"Feels fairer" isn't good enough, so we built a harness that judges every round against each rule. It counts a round as a break only when the break was avoidable: a better round existed given everything that had happened so far. For groups of up to 12 players, it checks this by listing every possible round (up to 155,925 of them) and comparing.
We ran 40 simulated sessions per scenario (5 for the slowest 12-player ones). Each session lasted as many rounds as there are players, up to 12. The harness is seeded, so the same run always gives the same numbers. First we tested each rule on its own, then all of them together in every priority order.
What changed
The harness found two reasons V1 broke its own rules, and V2 fixes both.
1. V1 measured the wrong thing for partners, opponents and skill
V1 scored repeat partners and opponents by how recently two players last met. It treated "last round" the same as "three rounds ago", then averaged across all pairs. Two long-ago meetings could cancel out two immediate repeats. V2 counts what players actually notice: how many pairs are repeats while a fresher option was available. How recently a pair last met only matters as a tie-breaker. For skill, V1 averaged the gap across courts. V2 looks first at the most uneven court, because that's the one people complain about.
2. V1 stopped searching too early
The algorithm uses simulated annealing. It starts by trying wild changes and becomes pickier as it "cools". V1 stopped while it was still accepting clearly worse rounds, so on big groups it often settled on a good-but-not-best round. V2 finishes colder. That costs about 75% more work per round, and still takes well under a tenth of a second.
The results
Share of rounds where each rule was broken although a better round existed, with each rule tested on its own. Lower is better.
What it means on court
Five or six friends, one court
With few players, opponents come round again quickly, so variety matters most. One V1 round from our tests had six players on one court. V1 put two players back against the same two opponents they had faced in the previous round, their third meeting. Meanwhile a round existed where nobody faced anyone more than once. V1's recency score actually rated the repeat round as better, because two long-ago pairings in it pulled the average down.
Before: an avoidable repeat opponent in 28.8% of rounds (6 players) and 22.5% (5 players). After: 0%. Repeats still happen when every option is a repeat (with 5 players that's from round 2, with 6 from round 4), but never when a fresher matchup was possible.
A busy club night: 12 to 23 players
On a big night the most visible problem is the bench. Someone sits out again while a player who has rested less keeps playing. With 23 players on 4 courts, V1 did this in 42.5% of rounds, and by the end of a session some players had sat out three more times than others. With 12 players on 2 courts it happened in 13.3% of rounds.
After: 0% in both. Everyone ends the session within one sit-out of everyone else, in every roster size we test, from 4 to 23 players. With 12 players on 3 courts you also play longer before getting a repeat partner: about round 11 instead of round 9, out of a theoretical best of 12.
Late arrivals and early leavers
Real sessions aren't tidy. People turn up at half past and leave before the end. Racket Social judges sit-outs by rate (rounds sat out ÷ rounds present), so a latecomer isn't benched to "catch up" on sit-outs they weren't there for. In our 12-player test with two late arrivals and one early leaver, V1 still got the bench wrong in 4.0% of rounds. V2 got it right every time.
Mixed-ability groups with skill balance on
With Skill-balanced matchmaking on, V1 averaged the gap across all courts. It would accept one lopsided court if the others were very even. Testing skill on its own with 8 players of mixed ratings, V1 left a court more than one rating point less even than it needed to be in 12.5% of rounds. V2, which looks first at the most uneven court, got it to 0%.
When the rules compete
In a real session all four rules are on at once, and your priority order decides who wins. We replayed 19 tricky mid-session situations under all 24 priority orders, 10 times each. Under the default Fairness order, out of 190 runs:
| Rule (Fairness order) | V1 breaks | V2 breaks |
|---|---|---|
| 1. Equal sit-outs | 6 | 0 |
| 2. Partner variety | 16 | 0 |
| 3. Opponent variety | 56 | 20 |
| 4. Skill balance | 33 | 23 |
Lower-priority rules can still give way. That's what a priority order is for. But the rules at the top now hold. Across all 24 orders, the rule you put first was broken in 701 of 4,560 runs on V1 and 71 on V2.
All single-rule results
Rounds with an avoidable break, each rule tested on its own. "1 court" means one doubles court. Rosters with spare players sit the rest out.
| Rule · roster | V1 | V2 |
|---|---|---|
| Equal sit-outs · 4, 5, 7, 10 and 14 players | 0% | 0% |
| Equal sit-outs · 12 players, 2 courts | 13.3% | 0% |
| Equal sit-outs · 12 players, late arrivals + leaver | 4.0% | 0% |
| Equal sit-outs · 18 players, 4 courts | 0.6% | 0% |
| Equal sit-outs · 23 players, 4 courts | 42.5% | 0% |
| Partner variety · 5 players, 1 court | 6.0% | 0% |
| Partner variety · 6 players / 8 players | 0% | 0% |
| Partner variety · 12 players, 3 courts | 16.7% | 0% |
| Opponent variety · 5 players, 1 court | 22.5% | 0% |
| Opponent variety · 6 players, 1 court | 28.8% | 0% |
| Opponent variety · 8 players, 2 courts | 37.5% | 0% |
| Opponent variety · 12 players, 3 courts | 65.0% | 20.0% |
| Skill balance · 8 players, mixed ratings | 12.5% | 0% |
| Skill balance · 8 players with one star / 10 players | 0% | 0% |
What we haven't fixed yet
- Opponent variety with 12 players on 3 courts is down from 65% to 20%, not zero. A better round exists in those cases but the search doesn't find it in time. Better search is next.
- Skill balance at the top of a Custom order can still be outvoted by partner variety below it, in about 13% of rounds. We plan to make priority strict, so a lower rule can never outweigh a higher one.
V2 is live for everyone; you don't need to change anything. If a round ever looks unfair, tell us in the WhatsApp community. Real sessions are the best test data we have.
About the numbers
-
Source: Racket Social rule harness, V1 and V2 runs of 18 October 2026, with the internal
write-up in
docs/801-algorithm-review/848-rule-harness-report.md. Reproduce from the app's source withnpm run test-harness:rules -- --algo v1and--algo v2. Runs are seeded and give identical numbers each time. - Single-rule runs: 40 seeded sessions per scenario (5 for 12 players on 3 courts), each lasting min(players, 12) rounds. Partner, opponent and skill breaks are judged against every possible round for up to 12 players. Sit-out breaks use an exact formula for any roster size.
- Combination runs: 19 mid-session situations × 24 four-rule priority orders × 10 seeds. Timing measured on a laptop. Phones are slower, but well within the 2-second budget.