Set Lab

Facts, not verdicts

Set Lab counts. It says how many removal spells a color has, how many copies of an archetype's commons the 24 packs of an eight-player table hold and how often one lands in the three packs you start with, and it sets each number beside the same number for past sets. It never says an archetype or a color is the one to draft, never calls a format quick or grindy, and never predicts what wins. What the numbers add up to is the reader's call.

The chart at the top is the curve table drawn: one panel for each color, multicolored and colorless, each with its count and average mana value, all on one scale, with the tallest column labeled. Columns of the pack table that are zero for the whole set, and removal methods no card in the set has, are left out and named under the table.

Three kinds of number

• Counted: plain counts of the set's card list, from each card's type line, mana value, colors and rarity.
• Read: from the card facts and the rules text, by the fixed rules below: what a card does to a creature and how, its keywords, and what it waits for or supplies. Every read number shows its rule and the cards behind it, each with the line it was read from.
• Estimate: copies per pack and every chance, worked out from each set's published slot rates. Real packs are printed on sheets and collated, which is close to random but not exactly random, so these are estimates.

Copies per pack

The page uses the pack layouts Draft Sim deals its packs from (one shared file, packs.js): each slot's rarity odds and the sheet it draws from, from Wizards' "Collecting" article for each set, checked against real Arena packs in 17Lands' public data. Reality Fracture's layout is Wizards' "Collecting Reality Fracture" (September 8, 2026): six commons, one uncommon, one slot that is a common 23% of the time and an uncommon otherwise, three echoed pairs cards, a rare or mythic, a traditional foil of any rarity and a land. The article counts 71 commons for the common slots, the common-or-uncommon slot and the foil's commons (the 70 nonland commons and Room of Refuge, the one common land that isn't a dual) and only the ten common duals for the land slot, and the page follows it. Its Special Guests (one pack in 55) aren't in the site's card list yet, and its Arena packs are taken to match paper until they can be checked after September 29.

For each slot, a card's chance is its rarity's odds times one over the size of that rarity's sheet. A slot that can't repeat a card already in the pack shares its chance among the cards still unused, each weighted by the chance it isn't in the pack yet; that is exact for a sheet of alike cards and very close otherwise. In the commons slot every card of the commons sheet is equally likely: every nonland common, and in five past sets and Reality Fracture also the one common land that doesn't tap for two colors. A wildcard or foil that comes up a common is a card of the same sheet. On Arena, past sets use the slot odds measured from real Arena packs (Draft Sim's explainer lists them).

copies of a card per pack = sum over slots of P(the slot's rarity) x 1 / (cards on that sheet)

a card of the commons sheet in the commons slot: n / N, n commons in the slot, N cards on the commons sheet

That last rule is measured. In 15,000 real first-pick Premier Draft packs of each of the six past sets (17Lands' public data, read September 25, 2026), mono-colored, gold and colorless commons each turned up at the same rate per card: in The Hobbit 0.139 copies a pack for each kind, in Secrets of Strixhaven 0.086, 0.086 and 0.084. Until September 25 Draft Sim dealt its commons one color at a time, which put gold and colorless commons at 0.36 to 0.49 times their real rate; since then it deals every common at one rate, and Set Lab reads its copies per pack from the same shared file. Copies at the table are 24 times the copies per pack.

Removal, evasion and tricks

The card facts say what to look for and the card's own text has to say it. Each way the facts give a card to answer a creature (its method, with the kinds of permanent it reaches) is checked against the card's lines, reminder text left out, and the line decides both whether it counts and how long it lasts. Removal: exile (not your own creature, and not one that returns or that its owner may cast again), destroy, damage, shrink that lowers toughness (-N/-N with N above zero, -1/-1 counters, base power and toughness 0/0), an edict whose sacrifice (or exile) names creatures or permanents, into the library, gaining control with no end, and can't attack or block, or doesn't untap, for as long as the effect lasts. Tempo, counted apart: bounce, tap, phase out, gaining control until end of turn, preventing its damage, and anything that lasts a turn (can't block this turn, -N/-0, base power and toughness until end of turn, exiled and returned, airbending). A method no line confirms is listed under the removal table as not counted, with the reason and the line, rather than guessed: reanimation read as taking a creature, an edict that asks for an artifact, a creature that loses its abilities but can still attack and block. A card with two ways shows on both lines and once in the total.

The damage a card deals is read from its text, mode by mode: the largest amount any of its lines deals to a creature or any target (a kicked or modal line included; an emblem's ability isn't the card's), and "varies" when every such line deals X, damage equal to a power, or a fight. The card facts carry one number per card, which is why the text decides. Conditions on the target (attacking creatures, creatures with flying) aren't weighed; each card list quotes the line.

Evasion is a creature with flying, menace, trample, skulk, shadow, fear, intimidate, horsemanship or landwalk among its own keywords (a keyword it only grants or gets under a condition doesn't count, and a two-faced card counts its front face only), or that the card facts' effect classes say can't be blocked (is_unblockable, conditional_evasion) or can't be blocked by some creatures (limits_blockers). A combat trick is an instant that gives a creature +N/+N, a keyword or protection until end of turn ("it gains" and "until end of turn, target creature gains" alike), or puts +1/+1 counters on one, and isn't counted as removal. A two-drop is a creature with mana value 2. Toughness is the printed toughness of the front face; a * is left out.

Color pairs

A card counts for a pair when its front face's mana cost can be paid with the pair's two colors alone: a hybrid symbol needs one of its colors, {2/W} and Phyrexian symbols need neither, and colorless cards count for every pair. So a mono-colored card counts in all four pairs of its color, and each pair's numbers are what a deck of those two colors can cast.

Signposts, enablers and payoffs

The signposts are each pair's gold uncommons: uncommons whose colors are exactly those two. Each one is read sentence by sentence with the Interaction Oracle's reader (synergy.js), reminder text left out, for what it waits for or counts: instants and sorceries you cast, noncreature spells (prowess), tokens, other creatures entering, many creatures, creatures dying or sacrificed, life gained, cards drawn, discarding or cycling, +1/+1 counters, noncombat damage to opponents, your graveyard, scrying or surveilling, lands entering, planeswalkers and loyalty, untapping, and creature or card types it names (a lord, a trigger, a count, a cost). The lanes follow the Oracle's pair engine: its producer and consumer table and the sentences it reads.

Then, among the commons and uncommons the pair can cast, an enabler is a card that supplies that thing and a payoff is a card that waits for it too, the signpost included. A card can be both. An enabler needs a line of its own text that does it (the card facts' tag alone isn't enough), and a payoff a line that triggers on it or counts it: a cost cut ("noncreature spells you cast cost {1} less") isn't a trigger, and a creature that only enters with a +1/+1 counter on itself neither supplies counters nor waits for them. Reality Fracture's Empower Jace makes or grows a Jace planeswalker token whose -1 surveils 1 and whose -3 draws a card: in the planeswalker lane that counts (the card makes a planeswalker); in the scry, graveyard, draw and token lanes it shows on its own line, "through the Jace token", never inside the enabler count. The rule for each lane is on the page, under each pair, and every count shows its cards with the line each was read from. The reader works by fixed patterns, so a card worded in a way it doesn't know is missed, and a pattern can catch a line that doesn't mean what it seems; that is why every card is listed.

Availability

For each enabler and payoff common: the copies in the 24 packs at the table; the copies in the three packs you start with; the chance at least one copy is in those three; and the copies that reach your seat. A card shows up at most once in a pack in these layouts (Teenage Mutant Ninja Turtles' foil slot is the one exception, handled apart), and packs are independent, so the three-pack chance is binomial:

P(at least one in your three packs) = 1 - (1 - p)^3, p = the card's chance per pack

Copies that reach your seat are a range, and both ends are exact. Over a round each of the eight packs reaches you once, with 0 to 7 cards already taken. If every drafter ahead of you takes the card on sight, you see only the copies in the three packs you start with; if none of them takes it, you see every copy at the table. Between the two the page shows one stated model, not a measurement: the drafters ahead take cards at random, so a card is still in a pack with chance (pack size - cards taken) / pack size.

copies that reach you: from 3 x copies per pack to 24 x copies per pack; random picks ahead: 24 x copies per pack x (1 - 3.5 / pack size)

For a group of commons, copies add up exactly. The chance at least one of the group is in your three packs can't come from a formula over the cards one by one: the commons of one pack come in color runs, so two commons of one color are less likely to share a pack than chance alone gives. So it is counted in packs Draft Sim's own dealer deals from the same slots, 40,000 per set and platform from a fixed seed (the page shows the same number every time), and then taken to three independent packs. With 40,000 packs the standard error of the three-pack chance is under 0.4 points.

P(one or more of the group in your three packs) = 1 - (1 - q)^3, q = share of the dealt packs holding one or more of the group

Past sets

The comparison counts every set Draft Sim supports the same way, from its own card list and its own pack layout: The Hobbit, Marvel Super Heroes, Secrets of Strixhaven, Teenage Mutant Ninja Turtles, Lorwyn Eclipsed and Avatar: The Last Airbender beside Reality Fracture. Bonus sheets (Mystical Archive, Special Guests, source material) are left out of these rows, so each row is the set's own cards.

How it was checked

Moving the pack layouts into the shared file changed nothing Draft Sim dealt at the time: 40 seeds of 24 packs on every set and both platforms came out card for card the same before and after. On September 25 the shared file changed how both pages treat commons and Arena slot odds (Draft Sim's explainer has the details). The arithmetic is checked against 30,000 packs Draft Sim deals per set and platform: every card lands within sampling error (a chi-square per card of 0.83 to 1.10, as sampling alone gives), and the copies per rarity Set Lab shows match the dealt packs within a hundredth of a card. Against the 15,000 real Arena packs per past set, the Arena figures match every card within sampling error, mono-colored, gold and colorless commons each within 5% as a group; Reality Fracture and paper packs have no real packs to check against yet. A browser test recounts the curve, creature counts, removal and tempo counts, damage amounts, the toughness shares and the availability figures independently from the built card data, sweeps the page for verdict words, and checks the page at phone width. An attack on the page (September 25, 2026) found removal totals that took in one-turn and until-end-of-turn effects, reanimation and an artifact-only edict (3% to 6.5% of removal per pack in every set), one damage number per multi-mode card, "enters with a +1/+1 counter" read as waiting for counters, Empower cards counted as scrying, a range for copies reaching your seat whose low end wasn't an end, and a page that jumped to the top on every switch; all are fixed, and the test pins each one with the cards it was found on.

Week by week

Real games splits each event's file into weeks counted from the file's first day: week 1 is that day and the six after it, week 2 the next seven, and so on; a short last week says how many days it holds. For each week: the games, the share won with its 95% Wilson range, and each two-color pair's share of that week's games (main colors as 17Lands records them, a splash counted with them). For one card at a time (the cards with 200 or more games in hand over the whole file, basic lands left out): its games in hand each week, the share won with its range, and how far that sits above or below all of the week's games:

points against the week = (won in hand / games in hand - the week's won / the week's games) x 100

The same 17Lands users play all through a file, and who plays changes as a format ages, so a week's rate is that week's players, not a fixed field. A week under 200 games in hand says its range is wide.

The decks of a color pair

For each of the ten two-color pairs: every run's deck (a draft, or a Sealed pool) as it was in its last game in the file, by game time, filed under that game's main colors. A pair's average deck adds those decks' copies card by card and divides by the runs; the average deck of the runs that reached the event's top result (7 wins; 4 match wins in Traditional Sealed) does the same over those runs alone. The counts come from Set Lab's own reading of each card, the same as the rest of the page: lands (basic lands and cards whose front face is a land), creatures, noncreature spells, removal by the removal rule above (a card counted as removal when a line of its text confirms a way the card facts give it to remove a creature), two-drop creatures, nonland cards by mana value, and the average mana value of the nonland cards. Splashed is the share of runs whose last game had a splash color. The 15 cards with the most copies per deck in the top runs are listed with their copies per deck in every run.

average count = sum over the pair's runs of copies of the kind in that run's deck / runs

Compare my deck puts one of your saved decks beside them, read the same way: its main deck as saved (sideboard and maybe cards left out). A card from outside the set is counted by its type line and mana value from the site's card data and not read for removal, and the page says how many; a name the site doesn't know is left out of the counts and said. Counts beside counts: what they mean for your deck is yours to judge.

A paper pack's value at store prices

An estimate of what the cards in a paper booster would cost bought one at a time, on average; never what a pack or box sells for (MTGJSON publishes no sealed product prices, and the site shows none). It reads two things only when you ask: the set's booster sheets from MTGJSON (MIT license; boosters.py turns its AllPrintingsCSVFiles into one file per set, MTGJSON's file of 2026-10-17), and each card's store price from the site's price layer (MTGJSON's AllPrices, by Scryfall id). MTGJSON lists, per booster, its variants with a weight each (the share of packs built that way), how many cards each variant takes from each sheet, and each sheet's cards with a weight each (how often a card turns up on that sheet) and whether the sheet is foil.

cards a pack takes from a sheet = sum over variants of (variant weight / all variants' weight) x cards that variant takes from the sheet

a sheet's value = sum over its cards of (card weight / the sheet's total weight) x price, price = the store's price (its foil price on a foil sheet), or 0 when the store has none or it is under your floor

pack value = sum over sheets of cards a pack takes from it x the sheet's value; box value = packs in the box x pack value

with a price = sum over sheets of cards taken x (weight of its cards the store prices / the sheet's total weight) / cards a pack

Stores: TCGplayer, Card Kingdom, Mana Pool, Cardmarket (in euros) and Card Kingdom's cash buylist (what that store pays). The floor is yours to set (0 unless you change it). The price layer keeps a printing only when one of its US prices is 25 cents or more, or Card Kingdom buys it, so a printing missing from the layer counts as 0. The slot table groups the sheets by name (Reality Fracture's many echoed pair sheets fold into the pair and the third card), each with the cards a pack takes, the cards on the sheet, the share with a price, its value a pack and its share of the pack's value; the ten printings adding the most are listed, each printing and finish on its own row (a card can show more than once), with their copies a pack, price and value a pack. A box's pack count, and any other item in it (not valued), is from MTGJSON's sealed product list. Only boosters sold as packs are read (play, collector, set, draft, Jumpstart); MTGJSON's prerelease, promo, sample and Arena sheets are left out. MTGJSON has booster sheets for 204 of the 875 sets it lists, and for 54 of the 59 paper expansion, core, masters and draft innovation sets released from 2020 to October 2026 (counted October 18, 2026).

Sources

• Wizards of the Coast, the "Collecting" article for each set on magic.wizards.com, and "Collecting Reality Fracture" (September 8, 2026): the pack slots and rates.
• The card DNA database (version 603b6c46): answer methods with their scope, keywords, and the effect classes Set Lab reads for evasion and for exile that holds a creature (setlab_build.py writes them for the supported sets).
• 17Lands public datasets (CC BY 4.0): real Arena packs, which the slot rates were checked against (the Draft Sim research brief). Data from 17Lands; they don't endorse this tool.
• The site's card facts (roles, answer methods, damage amounts, keywords, what a card does and triggers on) and the Interaction Oracle's sentence reader; card text, colors and rarities from Scryfall (read October 22, 2026).
• 17Lands public game data (CC BY 4.0) for Premier Draft, Sealed and Traditional Sealed: real games, week by week and the decks of a pair, summarized by gamedata.py into counts. Data from 17Lands; they don't endorse this tool.
• MTGJSON (MIT license): booster sheets and sealed product contents (file of 2026-10-17), and the store prices of the site's price layer.

Real games, by event

From 17Lands' public game data (CC BY 4.0) for three Arena events, switched at the top of the section: Premier Draft, Sealed and Traditional Sealed, each named in the credit and the tables (17Lands hasn't published Reality Fracture's files yet). The site summarizes each file into counts of games and wins and sums of deck copies, and nothing else. A card counts in a game when it was in the deck, in the opening hand, drawn later, in hand at all (the opening hand or drawn: GIH) or in the deck but never seen. Every rate is wins over games, with its 95% Wilson range; improvement when drawn is the win rate in hand minus the win rate when never seen, in points. Cards need 200 or more games in hand. The 60% group is 17Lands' own bucket of each player's overall game win rate. 17Lands users win more than the average player, so compare within a set and an event, not with 50%.

A run is a draft, or a Sealed pool. Premier Draft and Sealed are best of one, and a run reached 7 wins when its games in the file include seven won. Traditional Sealed is best of three: its rates are per game, a run won a match when it won more of that match's games in the file, and it reached 4 match wins when four of its matches were won. Games a player didn't track aren't in the file, so a match with an untracked game counts by the games that are.