01 A draw with no memory
Every grid carries six criteria: three for the rows, three for the columns. The generator draws them at random among those the network can play, 138 for the metro, with a seed taken from the date: a day's grid is always the same, and once created it never changes.
The draw only had three constraints:
- each cell has at least two stations that match both criteria;
- a grid keeps at most two criteria per family (place, history, trivia, name) and six from the network;
- no cell can end up blocked because its only answers would have been used elsewhere.
Then it starts over, up to 100 times, to land in the day's difficulty: very easy on Monday, very hard on Wednesday. But it never looked at the grids of previous days.
02 What I measured
On the last 73 metro grids, that is 438 criteria drawn:
| Measure | Result |
|---|---|
| Criterion already present the day before | 38 times |
| Pairs of grids less than 7 days apart sharing a criterion | 262 |
| Different criteria used | 121 out of 138 |
| Grids with the criterion Named after a public figure | 24, one in three |
So one day in two, you found at least one criterion from the day before. And a few criteria came back much more than the others. Named after a public figure, for example, is true for many stations: it pairs with almost everything and helps aim for an easy grid, so the difficulty filter often kept it.
03 Not all values weigh the same
The second flaw is more discreet. The draw was uniform per criterion: each of the 138 had the same chance. But some types of criteria exist in many values: 20 arrondissements, 18 nearby places, 16 lines, 10 letters. These four types make 64 criteria out of 138, and so took nearly half of the draws, while 39 types have only a single value (automatic station, terminus, concert hall…) and shared the rest.
Over the 73 grids, lines took 65 spots and letters 64: together, nearly one criterion in three.
04 Rule 1: a week of rest
Before drawing, the generator now rereads the criteria of the grids of the same network over the previous 7 days and sets them aside. A criterion played on Monday can't come back before the following Tuesday.
The weekly grid follows the same rule with its own grids: it reuses no criterion from the previous week's grid.
Setting aside up to 42 criteria still leaves about a hundred candidates, which almost always suffices. If no grid is possible without them, on a small network for example, the generator relaxes the rule in steps:
| Attempt | Criteria set aside |
|---|---|
| 1 | those of the previous 7 days |
| 2 | those of the day before only |
| 3 | none |
So a grid is always created. And the draw stays reproducible: the same date, with the same grids before it, gives the same grid.
05 Rule 2: a weight per type
Each criterion now gets a weight that depends on the number of values of its type:
weight = 1 / √(number of criteria of the same type)
A criterion alone in its type weighs 1. Each of the 16 lines weighs 1/4: together, the lines weigh 4, as much as four unique criteria. The square root is a compromise: a 1/n weight would give all the lines together the same chance as a single criterion like terminus, and the lines, which are the heart of the game, would become rare.
| Type | Values | Chance before | Chance after |
|---|---|---|---|
| Arrondissement | 20 | 14.5% | 6.3% |
| Near a landmark | 18 | 13.0% | 5.9% |
| Metro line | 16 | 11.6% | 5.6% |
| Contains the letter | 10 | 7.2% | 4.4% |
| A single-value type | 1 | 0.7% | 1.4% |
These chances are those of drawing a criterion, before the grid's constraints and the difficulty filter, which still rule out some candidates. The four big types go from 46% to 22% of the draws, and each unique criterion has twice the chance of coming up.
The weighted draw keeps the day's seed: each criterion gets a randomly drawn key, larger the stronger its weight, and the criteria are sorted by key. The first three make the rows; the criteria compatible with these three rows are sorted the same way, and the first ones that form a valid grid make the columns.
06 What doesn't change
Grids already created stay as they are: their criteria and answers are frozen on the day they are created. The rhythm of the week, the limits per family and the two-answers-per-cell rule don't move. Only upcoming grids benefit from the two rules.