01 What I count
A grid is hard when its cells have few answers. So I start from the number of valid stations in each of the 9 cells. But not all answers are equal: “Line 4” can be read off the map, “Young neighborhood” can't be guessed. So every criterion has a knowability weight, between 0 and 1, and each cell counts its effective answers:
effective answers = valid answers × weight of the row criterion × weight of the column criterion
| Weight | What it is possible to know | Examples |
|---|---|---|
| 1 | visible on the map or in the name | line, terminus, arrondissement, starts with a vowel |
| 0.6 | general knowledge | named after a public figure, opened before…, ghost station |
| 0.35 | specialist knowledge | next to a museum, a nightclub |
| 0.3 | local data | price per m², abstention |
| 0.15 | specialist local data | young neighborhood |
These weights are not pulled out of a hat: they start from the share of games that fill each cell, measured on the game. “Young neighborhood” is found in only 15% of games, “Starts with a vowel” in 92%.
02 The raw score
Three factors, each between 0 (easy) and 1 (hard), computed on these effective answers:
| Factor | Weight | Equals 1 (hard) when… | Equals 0 (easy) when… |
|---|---|---|---|
| Average | 50% | the average is 3 answers or fewer | the average is 12 answers or more |
| Minimum | 35% | the poorest cell has 3 answers or fewer | the poorest cell has 8 answers or more |
| Poor cells | 15% | 4 cells or more have 3 answers or fewer | no cell has 3 answers or fewer |
The average gives the general level; the minimum counts because a single almost-empty cell is enough to ruin a grid; the poor cells, because several are worse than one. In the code, that's all there is to it:
def raw_difficulty(answer_counts: Sequence[float]) -> float:
average_factor = 1 - _clamp((mean(answer_counts) - 3) / 9)
minimum_factor = 1 - _clamp((min(answer_counts) - 3) / 5)
low_factor = _clamp(sum(count <= LOW_ANSWERS for count in answer_counts) / 4)
return round(average_factor * 0.5 + minimum_factor * 0.35 + low_factor * 0.15, 3)
Three grids whose criteria can all be read off the map:
| Grid | Answers per cell | Average | Minimum | Cells ≤ 3 | Raw score |
|---|---|---|---|---|---|
| Easy | 10, 14, 7, 9, 12, 8, 20, 11, 6 | 10.8 | 6 | 0 | 0.208 |
| Medium | 12, 8, 5, 3, 20, 6, 2, 9, 15 | 8.9 | 2 | 2 | 0.598 |
| Hard | 4, 3, 2, 6, 3, 5, 9, 2, 7 | 4.6 | 2 | 4 | 0.914 |
Weighting changes a lot. A metro grid with “Vote nearby” and “Young neighborhood” scored 0.598 without weighting (easy); with it, its cells drop to 1.2 or 1.26 effective answers and it climbs to 0.890.
03 From the number to the level
The raw score is frozen when the grid is created, along with its answers: it never moves again, even if the station data changes. It is then sorted into five bands, from “Very easy” to “Very hard”. Their thresholds (0.67, 0.83, 0.90, 0.97) are the quintiles of 1,000 randomly drawn metro grids: each level contains about one grid in five.
| Level | Raw score |
|---|---|
| 1 · Very easy | less than 0.67 |
| 2 · Easy | from 0.67 to 0.83 |
| 3 · Medium | from 0.83 to 0.90 |
| 4 · Hard | from 0.90 to 0.97 |
| 5 · Very hard | 0.97 or more |
This is the level the game page shows to the right of the title. In the archives, the rating is relative: all grids are sorted by raw score and each one gets a rank from 1 to 5. So it can change when new grids arrive.
04 The rhythm of the week
The generator does not take the first grid that comes along. Each day aims for a difficulty range:
| Day | Target raw score |
|---|---|
| Monday | very easy |
| Tuesday, Friday, Sunday | medium at most |
| Thursday | easy or medium |
| Saturday | anything but very hard |
| Wednesday | very hard |
It draws a candidate grid from the date, computes its score and starts over as long as it falls outside the range, up to 100 candidates. At worst, it keeps the closest one: a grid is always created, and it is always the same for a given date. It takes 5 candidates on average for Monday and Wednesday, fewer than 2 for the other days.
The weekly grid, for its part, is always drawn from the very hard band.
05 What I'm improving
This score is still an estimate, and on the games of October and November 2026 it separates grids poorly: the grid of 6 November was lost in 86% of games. So I also measure, once the day has passed and with at least 10 games, the share of lost games of each grid. This measured level is not displayed yet: it is used to recalibrate the criteria weights so that the computed score matches how hard the grid feels.