01 The first-day problem
An answer is rare when few players give the same station in the same cell. But the first player of the day has no one to compare with, and a cell that has just opened has no statistics. So I start from a baseline: an estimate of the probability of thinking of each station, which the players' real choices then gradually shift.
02 The weight of a station
A busy station comes to mind more easily, and a connecting station is better known than a station on a single line:
weight = ridership ^ 0.75 × (1 + 0.5 × (metro lines − 1))
The 0.75 exponent flattens the gaps, otherwise Saint-Lazare and its 33 million entries would crush everything. Each line beyond the first adds 50%.
| Station | Ridership | Lines | Weight |
|---|---|---|---|
| Saint-Lazare | 33.1M | 4 | 1,091,667 |
| Châtelet | 8.4M | 5 | 466,033 |
| Abbesses | 1.5M | 1 | 43,159 |
| Botzaris | 0.67M | 1 | 23,401 |
Saint-Lazare has 50 times more entries than Botzaris, but only 19 times more weight from ridership, then 47 times more with its 4 lines.
This weight is corrected by history: a station that players find more often than expected over the last 180 days sees its weight go up, so it becomes less rare and earns less. A neglected station becomes rarer. Over time, the game diversifies.
03 The share in a cell
In a cell, the share of a station is its weight divided by the sum of the weights of all valid answers. Take a cell whose answers are the 8 stations of line 7bis:
| Station | Weight | Share |
|---|---|---|
| Pré-Saint-Gervais | 12,258 | 2.9% |
| Buttes Chaumont | 14,645 | 3.5% |
| Bolivar | 14,929 | 3.6% |
| Danube | 15,910 | 3.8% |
| Botzaris | 23,401 | 5.6% |
| Louis Blanc | 65,830 | 15.8% |
| Place des Fêtes | 89,132 | 21.4% |
| Jaurès | 180,742 | 43.4% |
This share is frozen when the grid is created, along with its answers: a past grid never depends on the current station data.
04 The points
Points don't come from the percentage itself but from the answer's rank in its cell: the rarest is rank 1, ties share the best rank.
points = 15 + 100 × (n − rank + 1) / n
with n the number of valid answers in the cell. The rarest earns 115 points, the most common 15 + 100 / n. The three rarest tiers add a bonus: +30 for mythical (1% or less), +20 for very rare (2% or less), +10 for rare (5% or less).
| Station | Rank | Rank points | Tier | Total |
|---|---|---|---|---|
| Pré-Saint-Gervais | 1 | 115 | rare, +10 | 125 |
| Buttes Chaumont | 2 | 103 | rare, +10 | 113 |
| Botzaris | 5 | 65 | uncommon | 65 |
| Jaurès | 8 | 28 | very common | 28 |
A cell where more than 4 answers are rare promotes its rarest to mythical, even above 1%, so that it always has its trophy. A game's score is the sum of its points, minus 20 per error and 50 per restart, capped at 900.
05 When the players take over
The percentage shown next to an answer mixes the baseline and the real choices, with Bayesian smoothing:
share = (station's choices + 30 × baseline) / (total choices in the cell + 30)
The 30 are virtual players spread according to the baseline. With no real player, the share is exactly the baseline; at 30 players, baseline and players weigh the same; at 300, the baseline counts for only 10%. So the first player is not at 100%, and a few isolated choices don't make the percentages jump. As long as a cell has fewer than 30 players, the percentage is shown rounded with “≈”: it still comes mostly from the estimate.
06 How I chose the settings
The 0.75 exponent, the 0.5 bonus per line and the 30 virtual players were chosen by replaying the generator on 60 grids (540 cells, 4,659 answers). With an exponent of 0.5, the baseline was too flat; with 1, the big stations crushed the big cells and “mythical” no longer meant anything. At 0.75, 9% of answers are mythical and almost no cell is “flat”, where even the most obvious answer would not exceed 5%.