Win Rates and Balance: Is It Really Too Strong?
A win rate only means something next to its range, so ask whether the whole range sits above your target, not whether one number looks high. With many sides or weapons, widen every range, because one of them will look strong by luck alone.
How do you read a table of win rates?
Tidecall Skirmish is an invented two-player strategy game with four sides, and these numbers are made up for this example. In 620 matches, each with two different sides, every match gave one win and one loss, so all the sides together won exactly 620 times.
The maker's target: every side should win 50%, give or take 5 points.
| Side | Wins | Plays | Win rate | Very likely between | Verdict |
|---|---|---|---|---|---|
| Storm | 12 | 19 | 63% | 36% and 84% | Cannot tell |
| Reef | 239 | 420 | 57% | 51% and 63% | Above target, maybe within limits |
| Kelp | 183 | 391 | 47% | 41% and 53% | Cannot tell |
| Gale | 186 | 410 | 45% | 39% and 52% | Cannot tell |
Start with the most tempting row. Storm has the highest win rate and the least to say. Its 19 plays leave a range from "weak" to "broken". About 560 more plays would settle it. Weakening Storm now would be reacting to a good week.
Reef is the only side with real news. Its whole range lies above 50%, so it very likely wins more than half the time. But it may only be at 51%, which is within what the maker accepts. So the honest verdict is: above target, not proven too strong.
Kelp and Gale look weak, and they could be fine. Notice also what a table like this hides: in a two-player game, if Reef wins more, others must win less. All sides together average 50%, simply because of how the sum works. So Kelp and Gale may be low only because of Reef, and strengthening them too would be fixing one problem twice.
Why are the ranges so wide?
For two reasons. First, the number of plays: pinning a win rate down to 5 points either way takes about 385 plays, and to half of that, about 1,540.
Second, and easy to miss: there are four sides. If you test enough things, some will look extreme by chance. Statisticians call this the multiple comparisons problem. Run 100 tests at the usual level, where nothing is really different, and about 5 will still look like a finding. The usual cure is the Bonferroni correction: you divide the allowed risk by the number of tests. So with four sides, each range is held to 1 in 80 rather than 1 in 20. Without that, Reef's range would be 52% to 62%.
When should you change a side?
- Set the target first. A boss should not win 50%. Fair means whatever you decided, in advance.
- Touch one side, by a little. The maker cut Reef's starting health by 3, and left Kelp and Gale alone.
- Wait for the number. Decide how many matches you need, then compare the two versions.
A side at exactly 50% can still be dull, so read weapon and item usage too. The win rate check gives your AI all of this as a finished tool.
Paste this to your AI
Tell me if any side in my game is too strong or too weak, and how sure you are. The play list is in [file or folder].
1. Ask me what fair means here: the win rate I want, and how far from it is fine. Write that down before looking at any results.
2. Draw one match as a Mermaid state diagram from choosing a side to the result. Check that every match in the data has exactly one result per player, and leave out mirror matches. Tell me how many you removed.
3. Use get_recipe with win-rate-check if Spielerando is connected. If it is not, use Wilson ranges, widened for the number of sides with the Bonferroni method, and say so.
4. Show a table: side, wins, plays, win rate, range, verdict. Allowed verdicts: too strong, too weak, fits, leans, cannot tell.
5. For each "cannot tell", say how many more plays it would take.
Suggest a change only for a side whose whole range is outside the target. Say what the change would do to the other sides. If nothing qualifies, say that no change is justified yet.
Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP
How to check it worked
- Add up the wins. In a one-against-one game, all wins must equal the number of matches.
- Look at the plays column first, before the win rate.
- Split the data by week. A real imbalance shows up in both halves.