200,000 Catan deals in three seconds: the 68.9% stat that hides a caveat
A number I published on our board generator site kept getting quoted without its second half: 68.9% of random Catan token layouts put a 6 next to an 8. That is true for a freely shuffled pile. It does not apply to the lettered setup in the rulebook. The gap between those two claims is the interesting part, so here is the simulation, the code, and where the number stops being useful.
The question, stated precisely
Catan's base board has 19 hexes: one desert and 18 number tokens. The tokens are a multiset, one each of 2 and 12, two each of 3 through 6 and 8 through 11. A 6 and an 8 each appear on 5 of 36 dice combinations, so an intersection touching both collects on 10 of 36 rolls before any other hex is counted.
The question: if you deal those 18 tokens at random onto the board, how often does at least one pair of neighbouring hexes carry a 6 and an 8?
Two details decide the answer. Whether the desert position is random or fixed, and what counts as neighbouring. I randomised the desert, because variable setup shuffles terrain before tokens. Neighbours are the six adjacent hexes in axial coordinates, edges only, no corner contact.
The whole method
The core loop is short enough to read in a minute:
from random import Random
rng = Random(7)
hits = [0] * 6
for _ in range(200_000):
tokens = deal(rng) # 18 numbers onto 18 non-desert hexes
pairs = sum(1 for a, b in edges if {num[a], num[b]} == {6, 8})
hits[min(pairs, 5)] += 1
edges is the set of adjacent hex pairs for the 3-4-5-4-3 layout. deal places the desert uniformly among 19 positions, then assigns the token multiset to the remaining hexes. That is the entire model.
Results from seed 7:
| Adjacent 6/8 pairs | Layouts | Share | Cumulative |
|---|---|---|---|
| 0 | 62,231 | 31.1% | 31.1% |
| 1 | 86,885 | 43.4% | 74.6% |
| 2 | 43,135 | 21.6% | 96.1% |
| 3 | 7,229 | 3.6% | 99.7% |
| 4 | 520 | 0.3% | 100% |
Mean pairs per layout: 0.98. At least one pair: 68.9%.
Three things I would change about how I first wrote it up
1. Say which setup the number describes. Official variable setup does not deal tokens freely. The tokens are lettered on the back and placed counterclockwise from a corner in alphabetical order, which is an anti-clustering device. A simulation of the lettered procedure would produce a different, probably lower number, but it depends on which corner you start from and how you handle the desert. I have not published that figure, and I should have said so in the first paragraph instead of the fourth.
2. Publish the seed and the raw counts. Reproducibility is cheap here. Same seed, same counts, three seconds on a laptop. If someone cannot rerun a statistic, it is a claim, not a measurement. The per-count totals and the script are both on the page.
3. Do not let a clean number imply more precision than the model has. 0.98 mean pairs per layout sounds exact. It is exact for this model. Change the desert rule, count corner contact, or simulate the lettered method and it moves.
What the number is actually good for
It is a baseline for the version of the game most tables play: someone shuffles the tokens and lays them out. Against that baseline, the no-6-and-8 rule is not fussiness. Nearly seven in ten boards break it by accident, and a single intersection collecting on 27.8% of rolls is enough to decide a game between players of similar skill.
It is also a decent example of a statistic that outlived its context. The first version of the page led with the headline and buried the caveat. The fix was not a softer headline. It was moving the limitation up and naming the alternative procedure explicitly, so a reader who quotes the number also quotes what it covers.
If you want to rerun it
The script is a single file, standard library only, and the counts are deterministic from the seed. If you change the layout to the 30-hex 5-6 player board, expect the share with at least one pair to rise, because there are more edges per token. That is a prediction, not a measurement; I have not run it yet.
The full write-up, the token distribution and the CSV are at hexboardgen.com/random-catan-boards.
Top comments (0)