Skip to content

How this works

Goldfish Lab does not score your deck. It plays it — shuffles, mulligans, makes land drops and casts spells, thousands of times — and reports what happened. This page says exactly what that means, and just as importantly what it leaves out.

What a goldfish is

A goldfish is a game against nobody. There is no opponent, so there is no removal, no counterspell, no blocker and no clock. That is a limitation and it is also the point: it isolates the one question your deck can answer on its own — how reliably does it do the thing it is built to do, and by which turn?

Concretely, the engine does not model:

  • • Opponents and interaction. Nothing is countered, killed or discarded.
  • • Combat, beyond its triggers. Every creature that can attack does, each turn, so that what it does as it attacks or deals combat damage to a player is played. Nothing blocks, no damage is added up and no creature dies. Your own life still moves: Necropotence, Dark Confidant and Phyrexian mana are paid for.
  • • Printed keywords. Flying, deathtouch and trample are shown on a card because a person reads them. They change no number here, and we never count them as if they did.
  • • Politics and table dynamics, which are most of what a Commander game is.

Where a card asks about the opponents, the engine assumes a table of three: each plays a land in each of their turns, and from turn 4 an opponent controls three creatures. That is enough for a card that checks it - Claim Jumper, Keeper of the Accord, Defense of the Heart - and it is all we assume about them.

So a deck that goldfishes beautifully can still lose every game. What this tells you is whether the deck's own plan is reliable — not whether the plan is good.

The opening hand, and how we know it is right

A Monte Carlo simulation is only worth reading if it agrees with mathematics wherever mathematics has an answer. For the first seven cards it does: the number of lands in an opening seven follows the hypergeometric distribution, which has a closed form.

The simulated distribution is checked against the exact one, computed with math.comb: the test suite draws 200,000 opening hands on every run and fails if they diverge beyond sampling error. The report used to print both side by side; it left the page because it answered a question nobody had asked, and the check did not leave with it.

The kept hand has no exact answer, because no closed form describes it — the mulligan rule throws the tails back and reshapes the distribution. The report's "Mulligans" shows what the rule did instead.

The command zone

The commander starts in the command zone and is cast from there when the mana is there, two more each time it was cast before. A deck may have two commanders - two partners, a creature that chooses a Background and its Background, or the Doctor's companion and the Doctor. Then each pays the tax for its own casts, the cheaper one that can be paid is cast first, and the report counts the games with at least one commander out and, in a row of its own, those with both.

A deck without a commander is played with an empty command zone, and its report says so. Every rule here is a Commander rule: 40 life and 100 cards. A deck built for another format is played by them all the same.

The mulligan rule

London mulligan: draw a fresh seven, then put cards on the bottom. In multiplayer Commander the first mulligan is free, so the first one bottoms nothing.

Because it costs nothing, the first seven needs one land more: it is kept with 3 to 5 lands, or with 2 lands and at least two cheap accelerants. A two-land seven without them goes back for free, as in Frank Karsten's mana tables.

Every hand after it is kept when:

  • • it has 2 to 5 lands; or
  • • it has exactly 1 land and at least two cheap accelerants.

A hand with 0 or 6 or more lands goes back. We stop after three mulligans, at five cards.

This heuristic decides every number that follows it, which is why it is written here in full rather than described as "a sensible keep rule". It is deliberately simpler than a good player: it cannot read your hand's plan, only its shape. One consequence is worth knowing — a typical deck keeps its first seven only about half the time, because a two-land seven goes back for free. The hand that replaces it starts with more lands more often, which is what the free mulligan is for.

The first draw, and how a turn makes mana

In multiplayer Commander nobody skips the draw on their first turn (Comprehensive Rules 103.8c), so that is the default. Skipping it is the two-player rule, and it is there as an option for a 1-v-1 game.

Every mana source is tapped once, at the start of the main phase: lands, mana rocks and creatures that tap for mana. A creature cast this turn is not in that pool yet, which is also what summoning sickness says - and so is a rock cast this turn, which in a real game could often be tapped at once. That second part makes the early turns read slightly lower than they play.

A source's cost is paid: a Signet turns one mana into its two colours, so it nets one. A card that does not untap (Mana Vault) makes its mana once. One that sacrifices itself (Lotus Petal) is used once, when it lets something be cast that otherwise could not. Phyrexian mana is paid with 2 life only when the mana is needed elsewhere, and never below 25 life.

What you drew

After every turn of every game, the engine counts the cards you have drawn so far: the opening hand you kept and each draw. A card a tutor found is not a draw and is not counted, and neither is a land a spell fetched. Cards a mulligan put on the bottom are not counted, and neither is your commander. Counting draws no random number, so the games are exactly the games they would have been without it.

Categories — ramp, card draw, removal and the rest — come from what each card does, read off its abilities, together with the community's card tags, or from what you said about a card. A spell that can only fetch lands, such as Cultivate, is ramp rather than a tutor, and a Treasure that comes beside a spell's card draw, as on Deadly Dispute, makes it no ramp spell. A card has one main role and at most a second one, or none at all, like most lands; strategies and mechanics such as a sacrifice outlet, lifegain or tokens come on top, as many as the card has. The chart shows the share of games with at least one by each turn, because that is the question a deck is built around: do I see ramp by turn two. Card types are shown as an average count instead: almost every game sees a creature by turn one, and how many is the part that differs between decks. The mana value bars count spells only — every land is mana value zero and would drown the curve.

Building blocks are the strategies and mechanics in one chart, the most often seen first. It draws eight lines at most; the rest are named under it, and its numbers hold every one.

On the deck page, Built around names the themes the cards point to, such as Aristocrats, Landfall or a creature type, read from the community's card tags. A theme has pieces: the cards that pay off, and the cards that feed them. It is strong, medium or hinted by how full its pieces are, the first counting double, and a piece with no card keeps it at hinted. A deck with no theme gets none: we do not guess.

The deck page's bars and filters sort the cards the same way, so a card you moved to another category moves there too. The faint band beside a bar is a common template for a hundred-card deck — about ten ramp, ten card draw, eight to ten removal, two to four board wipes, and Frank Karsten's 35 to 38 lands. It is a starting point many players use, not a rule, and no deck is marked for leaving it.

Under each category bar, “main” counts the cards whose main role it is. There every card counts once, so the main roles and the cards with none — lands, and the cards that serve the plan — add up to the deck.

Reading the report

Mana and lands, turn by turn

The median is what usually happens. The tenth percentile (p10) is the game that goes wrong — which is the one a deck gets rebuilt for, and the one an average hides completely.

What the mana was made of

How much of each colour the mana could pay. This is the column that says whether a mana base works: five mana on turn three is a good turn, and five mana of which four is the wrong colour is a turn spent doing nothing.

Means, and only means. A land or a rock that makes one of several colours — a dual land, Command Tower — counts toward every colour it offers, because the engine keeps that choice open until a spell is paid. So the colours of a turn can add up to more than its total.

Milestones

The share of games in which each milestone was true by the end of that turn. They are read from the cards' categories, the ones you can correct on a card's page, so a deck built differently gets the same chart without anything here changing. A card that draws counts as a draw engine only when it draws again and again. The chart draws the eight most frequent; a rarer ninth stays in the numbers under it.

Stress test

A goldfish meets no opponent, so the report adds two moments when the table strikes back: your commander destroyed once, in the opponents' turns right after it first comes down, and every creature destroyed before your turn 5. Each is played on a copy of the game taken at that moment, so the game with it and the game without it are the same game until then, and the difference is the disruption alone. The rest of the report is untouched by it.

Both are assumptions, named as such: nobody has published when real tables remove a commander or cast a wipe, so we do not pretend to know. A destroyed commander goes back to the command zone and costs two more; a wipe takes tokens and mana creatures too, while artifacts, enchantments and lands stay. Indestructible and regeneration are not played. The stress test runs in the first 2,000 games of a run, and the wipe only in runs of six turns or more.

On a report each blow is a card with one number, how fast the deck comes back: for the commander, the share of games it was back by the last turn; for the wipe, the mana on that turn. "The numbers" below them set the same games side by side, with the blow and without it.

Speed and the Commander Brackets

Since October 2025, Wizards of the Coast describes each Commander Bracket by how long a game is expected to last: at least nine turns in Bracket 1, eight in Bracket 2, six in Bracket 3 and four in Bracket 4; a Bracket 5 game can end on any turn. For every combo of your deck that ends the game on its own - a win, infinite damage, life loss or mill, endless turns - the simulation notes the first turn its cards were all where they need to be. A combo together by the end of turn five ends a six-turn game too soon.

A bracket fits when fewer than half your games have such a combo together before its turn. Only combos your deck already holds count, never one a card is missing for, and only the turns the run played: a short run leaves the lower brackets unchecked. Speed is one rule among several - Game Changers, mass land denial and extra turns decide a bracket as well - so this is a check, not a rating.

Why it is built this way: each bracket is checked against its own turn rather than one average win turn, so you see which brackets fit instead of a number to interpret. The share is given for every turn, because a deck that wins early in a third of its games plays differently from one that never does. A combo counts only when its cards are in the zones it needs, so a combo the board cannot use yet is not credited. And it is counted in every game of your run rather than in a sample, so the shares are as exact as the run is large.

What a simulation does not model

Some limits are not bugs and more games will not fix them. They are the difference between a goldfish game and a real one, so the report names which of your cards they touch, and the field on each card that answers them. The list is read from the deck as it stands now rather than from the run: these are properties of the cards, not measurements a simulation made.

When a card's reading is changed after a run, that run is not recalculated — its numbers are a record of what the engine actually saw — so the report says so, and running the deck again shows the difference. How exact a number is depends on the number of games: the report gives its main shares with a ±, explained under “How many games is enough”.

What the engine could not read

Every result says how much of the deck the engine could read from the cards themselves, counting every copy: thirty Swamps are thirty of a hundred cards. A gap there is ours — a card whose text we could not turn into behaviour — and the result lists every one.

What it counts is only that. When several spells are affordable, the engine casts the cheapest first — its own rule, not a reading of your deck's plan, with a built-in order for a few dozen well-known cards such as rituals and fast mana. Your commander comes before all of them but fast mana and a few staples, as soon as it can be paid, because that is how a Commander deck is played; what is left that turn is still spent. We used to ask for a casting order card by card and scored decks on how much of it had been given; on the deck we measured, 161 of 217 gaps were that question, so the score mostly reported how much hand-annotation a deck had received. We stopped asking: this is a tool for statistics about a deck, not for steering a game.

A tutor is the exception, because the cheapest card in a library is a land. A tutor takes a spell: first a piece of one of the deck's combos or a card-advantage engine, then the biggest spell it can cast by next turn. Only a search that can find nothing but lands takes a land.

Wherever nothing could be established, the answer is a stated gap rather than a guess. A deriver that defaults a missing number is right often enough to look correct and silent when it is not.

A card the engine could not read marks its deck red — “3 cards need you” — until you have looked at each one: correct what it reads, or say it looks right. Either way the deck turns “Ready”. Simulating never waits for it. Every value on a card's page says where it came from:

Scryfall field
Structured data straight off the card. As close to ground truth as this application gets.
Community tagger
Rolled up from the Scryfall Tagger project. Community-maintained, deliberately broad, and occasionally an opinion rather than a fact.
Read from the card text
A pattern matched against English prose. The weakest reading this application makes, and the one most worth checking.
Built in
A default that ships with the application and applies to every deck.
You, for all your decks
A judgement you recorded.
You, for this deck
A judgement you recorded for this deck only.
The card's categories
Read off what the card counts as, under “Categories”: correct it there and the engine plays it that way.
Nobody — the engine's own rule
No one said, so the engine's fallback applies. Worth a look if the card matters to how the deck plays.

Combos

Combos come from Commander Spellbook, looked up for your actual deck. We never guess a combo by reading card text.

What the simulator adds is the part nobody else can: how often a combo actually comes together, and by which turn. A combo counts as assembled only when every card is in the zone the combo needs it in — on the battlefield, in the graveyard, in hand. Not "drawn", and not "seen". Measuring drawn cards would be easier and would inflate every number on the page.

For a combo your deck is one card short of, the question is hypothetical, so we say so and measure it honestly: on a second deck that is yours plus that card, nothing cut, and the page names the card. Choosing which card your deck can spare is not a judgement this application makes.

Expect small numbers, and read them carefully. Two specific cards in a 99-card singleton deck is about 1.6% before anything has to be cast. A result under one percent is printed as "under 1%" rather than rounded to "0%", because 0% reads as never next to a combo that did come together.

Engines and synergies

Engines and synergies are read from what the deck is built around, by fixed rules over Scryfall's Tagger tags and our own reading of the cards, never by an AI. A deck with no theme has none, and a stray card outside a theme pairs with nothing.

A synergy is two cards one of the theme's rules joins: one serves the other, as a sacrifice outlet serves a card that pays off each death. Its strength is the rule's weight, times how rare the tags are that put each card in its part of the plan (a tag on few cards is a stronger sign than one on thousands), times how full the theme is in this deck.

An engine is one card for each part of a theme, two to four cards that make value together again and again: at least one of them has to work more than once, an ability you can activate or one that waits for something to happen. It carries its theme's level. An engine a part short names what is missing.

You can lock an engine or a pair, or dismiss it. We store only that, by its cards: a locked one stays first, even if the rules stop finding it; a dismissed one is not suggested again, and you can bring it back. When one of its cards leaves the deck, it rests until the card is back. Everything else is read again from your list on every visit.

A simulation also measures by which turn each engine runs: at the end of a turn, every part of it has a card out, any card of the deck that fills that part, not only the ones the engine names. A permanent counts on the battlefield, an instant or a sorcery in hand, ready to cast; the commander counts once it is cast. Up to five engines, one per theme, locked ones first; one a part short or resting is not measured. It is read in the run's own games, so it costs nothing extra.

With deck summaries switched on, the summary also says how each engine plays, in a sentence or two under it: which card to get out first, what to protect, what stops it. Mistral AI writes that with the rest of the summary, from the engine's cards, so it costs nothing extra and can be wrong. The engine itself is still found by the rules.

How many games is enough

Ten thousand games pin a share of 50% to about ±1.0 percentage points at 95% confidence, and one of 80% to about ±0.8; two thousand games to about ±2.2 and ±1.8. Past ten thousand you are buying decimal places that do not change a decision, which is why this application meters turns simulated rather than selling you iteration counts.

The ± beside a share in a report is that interval for the games the run played: 1.96 × √(p × (1 − p) / games). Two runs whose shares are further apart than their ± together differ for a reason, not by chance.

Four lands on turn 4, at the top of every report and on the precon pages, is the share of games with at least four lands on the battlefield at the end of turn 4: every land drop made, the commonest way a Commander deck stumbles early. A run of fewer than four turns does not have it.

In short, at the top of a report, sets three of its numbers next to every Commander precon's, each played the same way: ten thousand games, on the draw. A faint tick is a precon, the tall mark is your deck, and "the middle" is the precon in the middle. A precon's own page leaves that precon out, and with fewer than five precons played there is nothing to compare with.

Compared with your last run sets the same three numbers of the deck's run before next to this one's. Two runs of the very same deck differ a little by chance alone, so a number only counts as better or worse when the difference is larger than chance would make it 95 times in 100; otherwise it says "about the same". A run keeps a fingerprint of the card list, not the list, so the report can say that the deck changed in between, but not which cards.

A run is split into chunks across background workers. A chunked run is deliberately not bit-identical to a single run of the same seed; it replays exactly and agrees with one within sampling error, and both halves of that are asserted by tests.

Checked against Karsten

Frank Karsten's How Many Sources Do You Need to Consistently Cast Your Spells? (2022) is the standard answer to how many sources of a colour a deck needs to cast a spell on curve. For a 99-card deck he assumes 41 lands, the free first mulligan and a draw on turn 1. A spell counts as cast on curve when enough sources of its colour have been drawn by the turn equal to its mana value, among the games that drew that many lands. "Consistently" means 89% plus the mana value: 90% for a one-drop, 93% for a four-drop.

We ask his question twice, counting the way he does. With his mulligan rule, the count reaches his number of sources for every mana cost in his table and falls short with one source fewer: the counting is his. With the engine's shuffle, mulligan rule and draws, the ones every report uses, the answer is in the table below.

Mana cost Karsten Engine
5C 14 14
4C 15 15
3C 16 17 (+1)
2C 18 18
5CC 20 20
1C 19 19
4CC 22 22
C 19 20 (+1)
3CC 23 23
4CCC 26 26
2CC 26 26
3CCC 28 28
1CC 28 28
2CCC 30 30
CC 30 30
1CCC 33 33
1CCCC 36 36
CCC 36 36
CCCC 39 39

Sources of one colour among 41 lands, in Karsten's order. "C" is one coloured mana symbol, the digit the generic part: 2CC is a four-drop with two pips of the colour.

The engine needs the same number for 17 of the 19 costs. For the others it misses Karsten's target at his number by less than a tenth of a point, about the sampling error of his own count, and so asks for one source more. Until engine v38 it needed one or two more for seven costs, because it kept a free seven with two lands, which Karsten throws back. A game that starts from two lands and still finds its lands in time usually has just enough of them, where a three-land start has some to spare, and every spare land is one more chance at the colour. Since v38 the engine throws that hand back too.

Not checked yet: gold cards, for which Karsten adds one source to each colour, and decks with mana rocks or card selection, which his table leaves out too. The test suite replays these checks.

Where this comes from

If a number on this site ever disagrees with this page, the page is the bug. Say so and it gets fixed.