Smuggler's Cove: The Mathematically Best Bottle Selection

written by Zach Morrissey on 4/12/2026

Smuggler’s Cove by Martin Cate came to me in the deepest depths of the COVID pandemic, where I was looking desperately for things to do with my time. I’d been given an extra copy by my sister-in-law who loved making drinks from it. I was quite intimidated. It’s beautiful, it’s expansive, it’s uncompromising.

So I started with making the Mai Tai. Then next, the Planter’s Punch. Then the Zombie. After just those three drinks, I’d amassed an impressive collection of bottles. Looking into the rest of the recipes in the book, I had no idea where to go. Buying good rum and ingredients is a spendy endeavour, and I want to know what I’m in for here.

Skip to the bottle recs here if you don’t care about methodology

What to buy?

Let’s narrow down the problem. As an example, a selection of 10 bottles from all of those mentioned in the book yields roughly 536 billion possible selections of bottles. And that’s just liquor bottles, not anything else. So we’ll have to narrow it down further.

The stats

Smuggler’s Cove has 133 ingredients used across its 115 recipes, of which 72 are bottles of alcohol in some form (liquor, liqueurs, bitters, beers, wines, amari, etc). I used the Smuggler’s Cove Index that was posted on reddit a few years back, and did some data cleanup + added metadata for each of the ingredients within.

Limiting the selection to just bottles of alcohol gives us a meaningful reduction in search space: Most of the time the alcohol is the most expensive portion of a recipe and the least easily obtained. The rest of the ingredients are largely purchased at a supermarket or, less often, ordered online at a specialty shop.

Just choose the most frequent ingredients? Nope.

A simple approach would be to find the bottles that appear the most often. Unfortunately, that doesn’t guarantee good results. We want to choose bottles that optimize for the number of additional drinks we can make. Statistically, it’s better to choose bottles with high co-occurrence.

Let’s say you’ve got these drinks, using some made up data here:

# drink -> list of ingredients
drink_1: ["A", "C", "D"]
drink_2: ["A", "E"]
drink_3: ["B", "F"]
drink_4: ["B", "G", "H"]

Ingredients 'A' and 'B' both occur in 50% of the recipes, but have zero co-occurrences. They’re the most popular, but if you bought them both you’d have zero drinks you could make. On the contrary, if you’ve got a bar with bottles 'A', 'C', and 'D' in it, adding 'B' doesn’t give you any additional drinks you can make, but adding 'E' would. Most frequent elements may give you an optimal subset of ingredients, but they’re not guaranteed to.

Finding a solution: enter the CP-SAT solver

Naively brute forcing all 500+ billion combinations of bottles ain’t gonna work.

I opted to use Google’s OR-Tools, which is a toolkit that bills itself as “open source software for combinatorial optimization, which seeks to find the best solution to a problem out of a very large set of possible solutions”. Most specifically, I use the CP-SAT solver in Python.

This allows one to express a problem in terms of variables, constraints and objectives, and the solver will do the hard work of figuring out the best solution for you. No hand-rolling your own optimizations required! Using the example from the docs, it looks like:

Variables are used to designate your search space. Imagine three variables x, y, and z that can be either 0, 1, or 2.
Constraints are a way to phrase how you’d like to maximize for a specific criteria. We want x != y.

The model will give you every combination of values for x, y, and z that follow the constraint x != y.

x=1 y=0 z=0
x=2 y=0 z=0
x=2 y=1 z=0
x=2 y=1 z=1
x=2 y=1 z=2
...etc

This, however, just gives you all of the options available. So how do you describe the optimal solution? You can add an objective function to get that. Let’s say you want the highest value of all options when added together.

# set the objective function for the model so that it gives you the best output.
model.Maximize(x + y + z)

A little napkin math will tell you that the result is 5, but you can also plug in more complex functions of x, y and z to meet your use case.

The Best n Bottles to Buy

So how would one model the constraints for Smuggler’s Cove in a way that it finds us the best selection of bottles? Let’s plug in our criteria. This first example will solve for the optimal set of n bottles to make the largest number of cocktails, where n is a number of my choosing.

Ingredients

Ingredients can be modeled as booleans. They’re present (1) or not (0) within a drink recipe.


# first, create the model
from ortools.sat.python import cp_model
model = cp_model.CpModel()

# create a dictionary comprehension of all ingredient names -> variables
ingredient_vars = {x: model.NewBoolVar(x) for x in ingredients}

# require that the model only have the number of ingredients that we've asked it
# for. so if we want the best 5 bottles, it'll give us the best 5 bottles.
model.Add(sum(ingredient_vars.values()) == n)  # n = number I want

Drinks

Each drink can also be modeled as a boolean, but we need to make sure that they’re constrained such that being able to make a drink requires every ingredient used.

drink_vars = []
for drink in drinks:
    required_ingredient_vars = [ingredient_vars[x] for x in drink.ingredients]
    drink_var = model.NewBoolVar(drink.name)
    model.AddBoolAnd(required_ingredient_vars).OnlyEnforceIf(drink_var)
    drink_vars.append(drink_var)

Optimization function

Once your relationship between drinks and ingredients is found, you can use the number of drinks with each subset to become your objective.

# higher sum of drinks = better outcome
model.Maximize(sum(drink_vars))

However…

Not all drinks are created equal

The problem? That’s just solving for the most drinks you can make, which isn’t really how people use recipes in a book. Making many different drinks isn’t a useful outcome if you don’t like any of them. The goal must aim for the best drinks in the book, where it weights competing subsets by how many of the best drinks in the book each can make.

And how did we determine that? A little scraping of a variety of misc reddit threads where users have ranked their favorites over time and then using those to build out a ‘scores’ sheet of the various drinks.

Here’s a peek into the top 10, by ranking:

So to solve for this better, this uses a blended objective of the number of drinks as well as the highest drink scores. In python pseudocode, that objective looks like:

# alpha used in this analysis = 0.3
# a configurable value to blend between maximizing number of cocktails vs. score
# alpha 1.0 = only consider number of cocktails
# alpha 0.0 = only consider cocktail scores
objective = (alpha * num_drinks) + ((1 - a) * cocktail_score)

Determining the Right Number (n) of ingredients

OK so is 10 the right number to aim for?

So this will give you the best available subset of ingredients to maximize how many drinks one can make given a subset size, but how does one know how many to aim for? We can run the script for every n ingredients to the maximum and see if there’s a natural inflection point:

…so, not really. You start seeing diminishing returns around 35 bottles or so, but the trend’s mostly a “1 new ingredient = 1 new possible drink” trend.

Let’s get an actual bottles recommendation

OK, so using the original problem (which 10 bottles do I buy?), and the default settings you’ll get the following result, which is very close to what I’d recommend:

Bitters (Aromatic)
Bitters (Herbstura)*
Liqueur (Allspice Dram)
Liqueur (Falernum)
Liqueur (Maraschino)
Rum (Black Blended Overproof)
Rum (Black Blended)
Rum (Blended Aged)
Rum (Blended Lightly Aged)
Rum (Column Still Aged)

* This is defined as a 50/50 blend of Angostura and Herbsaint. IMO could easily be anything anise-like e.g. pastis, absinthe, etc.

Which allows you to make 28 of the most popular drinks, actually!

Barbados Rum Punch
Bombo
Boo Loo
Cora Middleton
Corn And Oil
Daiquiri No. 1
Demerara Dry Float
Hot Buttered Rum
House Spiced Rum
Hurricane
Jet Pilot
Kaiteur Swizzle
Lei Lani Volcano 2.0
Mary Pickford
Pampanito
Paniolo Old-Fashioned
Planter’s Punch
Puka Punch
Pupule
Queen’s Park Swizzle
Rum Flip
Rumbustion Punch
Sidewinder’s Fang
Smuggler’s Rum Barrel
The Twenty Seventy Swizzle
Tiki Bowl
Top Notch Volcano
Zombie

Now that’s a good lineup of drinks! And let’s be real, that’s just if you’re very strict about adherence to the correct category of rum there. The perennial favorite, the Mai Tai, would fare just as well with one or a mix of the rums designated above.

Go forth and get tropical. If you liked this, code’s here on Github.