Draft Sim

How each event runs

Pick the event and the draft runs the way it does in real life. Arena Quick Draft: you and seven bots, three packs, one card at a time, passing left, right, left, untimed (Wizards calls it an "untimed bot draft"), so it has no clock here either, then best-of-one games until 7 wins or 3 losses. Arena Premier and Traditional Draft: eight people and a clock on every pick; Premier then plays ranked best-of-one until 7 wins or 3 losses (Contender Draft drafts the same way), Traditional three best-of-three matches. Arena's clock: Wizards said at the Ikoria launch (April 16, 2020) that early picks get 75 seconds, shrinking to 30. The steps in between aren't published, so the clock here starts at 75 seconds on a full pack and falls 5 seconds for each card already taken from it, down to 30 [estimate]; a 13-card Lorwyn Eclipsed pack starts at 75 too. Arena Pick-Two Draft: four seats, and every time a pack reaches you, you take two cards: with 14-card packs, seven picks a pack, 42 cards, and your first pack comes back at your fifth pick with six cards (confirmed in 17Lands' public Pick-Two data). Lorwyn Eclipsed's 13-card Arena packs give 39 cards, a last pick of a single card, and five cards back at your fifth pick. The clock runs per pass. Then best-of-one until 4 wins or 2 losses. Paper booster draft (Magic Tournament Rules, effective February 27, 2026): eight players, three Play Boosters each, with the token or ad card taken out first, so 14 picks a pack (MTR 7.7), passing left, right, left, and you play within your pod (MTR 7.6). At Competitive REL (RCQs and up) the pick clock is the table in MTR Appendix B: 40 seconds for a pack of 15 or 14 cards, 35 for 13, 30 for 12, 25 for 11 or 10, 20 for 9 or 8, 15 for 7, 10 for 6 or 5, and 5 for 4, 3 or 2. You may not look at your picks except in the review periods, 60 seconds after pack one and 90 after pack two, and then you get 25 minutes to register and build. Here your picks stay face down until each review period. At Regular REL (store drafts, Friday Night Magic) the clock is up to the store, so it starts off, and you can see your picks. Paper Pick-Two: four players, two cards a pass; the tournament rules set no clock for it, so its switch borrows the booster draft table. Sealed: six packs. Arena Sealed and Traditional Sealed build untimed; paper Competitive REL gives 20 minutes to register the pool and then 30 to build (45 at a prerelease), and the build clock here is the 30. Cubes from Cube Generator, in seven formats: booster draft and Pick-Two for 2 to 16 players with the packs each and pack size you choose (without a choice, as many players as the cube has 45 cards for, up to eight, three packs of 15); Rochester (packs opened one at a time, face up, picked around the table and back from the opener, the openers rotating); Rotisserie (the whole cube face up, picks snaking around the table; you pick first unless you choose another seat or a random one, and a very large cube shows each color's first 150 cards by mana value with the rest a tap away, the search reading every card left); Winston and Grid for two players; and sealed. The other seats there are modeled drafters: they take a card with chance in proportion to e^(its cube rating plus the pull toward their colors), the booster table's rule applied to whatever is on offer [MODEL]; in Winston the other player takes a pile when its cards add up to 2 or more above the cube's median card on that scale, or when it holds 4 or more, and in Grid it takes a row or column with chance in proportion to e^(its cards' summed values) [ESTIMATE: nothing published measures how people choose piles or lines]. Two cases the published Winston rules don't cover are ruled here: with the deck empty, the last pile with cards has to be taken; and if adding a card to the third pile empties the deck, that pile is taken. A cube too small for the night can't start, and the page says by how much. Once a draft starts the address drops the cube's settings, so a reload brings back the draft, finished or not, with your deck as you left it. A cube edited during a draft: adding cards or renaming it changes nothing in the draft; taking out a card the draft holds (in a pack, on the table, in a pile or a grid) sets the draft aside with a note, as it does for every format. Every clock is a switch, on by default only where the real event has one. Its deadline is saved with the draft, so a reload doesn't give you fresh time; switching the timer off pauses it, and on again resumes with the time you had. When a clock runs out, the sim takes the card you last looked at, or the first card in the pack if you looked at none; Magic Online takes the card you reserved, and what Arena takes isn't published. The last card of each pack (the last two in Pick-Two) is handed to you, as at a real table, and the pick review marks it as no choice. Not modeled on their own: Magic Online's queues and leagues (they draft like the paper table, but their pick timers aren't published), Rochester outside a cube and other custom drafts, and chaos drafts.

The packs

Each pack is dealt from the set's card list to its Play Booster layout in Wizards' "Collecting" article for the set, then checked against real Arena packs in 17Lands' public data. Reality Fracture ("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, one rare or mythic (about one in six a mythic), one traditional foil of any rarity, and one land (a common dual 54.5% of the time, otherwise a basic). The three echoed cards are one pair, two versions of one character at one rarity, plus one other echoed card; the pair is uncommon 91.7% of the time, rare 6.9% and mythic 1.4% (Wizards' 88.9%, 6.7% and 1.4%, with the borderless versions folded in). The pairs are matched by hand from the card list, 33 uncommon, 7 rare and 3 mythic, as Wizards' 66, 14 and 6 echoed cards require, so a pack's two linked cards are the same character (Gideon's Memorial with Gideon the Oathless, Thalia, the Survivor with Geist of Saint Thalia). Two groups are inferred, not published: Jace, Reality Sculptor with Tam, the Possibility (the only two rares left once every same-character pair is matched), and the five pairs of "Way of the" enchantments, matched by their mirrored designs. The pick screen names the echoed pair in each pack and says when its match is inferred. Special Guests (one pack in 55) are left out until their ten cards are in the site's list, and Arena packs are taken to match paper until they can be checked after September 29.

Checked against real packs: 20,000 simulated Arena packs a set against 15,000 real first-pick Arena packs a set (17Lands' public data, read September 25, 2026), card by card. Every card of the six past sets turns up at its real rate within sampling error: summed over each set's cards, the squared differences in units of their sampling error come to 0.8 to 1.1 a card, where sampling alone gives about 1. The Hobbit's wildcard is Wizards' literal 74.17% common and 3.9% uncommon. This page used to read those two as swapped, but 130,950 real Arena packs average 8.86 commons and 3.35 uncommons, the literal reading predicts 8.84 and 3.33, and the Arena packs here come out at 8.86 and 3.34. Secrets of Strixhaven now has its own layout: six commons, three uncommons, a wildcard (39.1% common, 39.1% uncommon, 19.5% rare, 1.9% mythic), a rare or mythic, one Mystical Archive card in every pack (87.5% uncommon, 9.6% rare, 2.9% mythic), a foil that can also be an Archive uncommon, and a land. On Arena the Archive card's odds are the measured ones (86.5% uncommon, 10.2% rare, 3.3% mythic), and a foil is an Archive uncommon 3.7% of the time; the Arena packs here hold 1.04 Archive cards and 7.42 commons, as the 15,000 real ones do. Lorwyn Eclipsed packs on Arena leave out the basics-only land slot, so they hold 13 cards (39 picks, 78-card sealed pools); paper packs hold 14. Teenage Mutant Ninja Turtles' land slot is a common dual 60% of the time in paper and 55.4% on Arena (64,302 real packs). In real Arena packs a Legendary Turtle common or uncommon turns up no more often than any other card of its rarity, so on Arena its Turtle slot deals any card of its rarity; there the four heroes' own rares (Leonardo, Cutting Edge and the like) turn up 1.35 times as often as other rares and their own mythics (Leonardo, Sewer Samurai and the like) 0.44 times as often as other mythics, and every rare or mythic taken on Arena here is weighed that way [measured; why real packs do this isn't known]. Commons: in real packs every card of the commons sheet turns up equally often, mono-colored, gold and colorless alike (The Hobbit: 0.139 copies a pack each). Until September 25 the packs here dealt one common of each color first, which put gold and colorless commons at 0.36 to 0.49 times their real rate and mono-colored ones 2% to 22% high. Now a pack's commons are a run of cards in a row on a sheet where the five colors take turns, gold and colorless commons spaced evenly between them, and a wildcard or foil that comes up a common is the next card of the run; the rest of the packs draw their commons at random. Both ways deal every common at the same rate. A common land that doesn't tap for two colors is on the commons sheet too (Hobbit Hole, Rumble Arena, Evolving Wilds, Terramorphic Expanse, Surveillance Room: in real packs each turns up as often as any common), while the common duals share the land slot; Turtles' Escape Tunnel is dealt with its duals, as its real packs show. Color spread: real packs rarely miss a color. All five colors among the mono-colored commons: real Hobbit packs 79.3%, Secrets of Strixhaven 72.5%, Marvel Super Heroes 100% (8,000 real first-pick packs each), against 79%, 69% and 100% here, where drawing commons at random would give 27%, 9% and 27%. The share of packs dealt from the striped sheet is fitted per set to match (The Hobbit 70%, Strixhaven 80%, Marvel Super Heroes 100%, Avatar 55%, Lorwyn Eclipsed and Turtles 50%) [estimate]; Avatar, Lorwyn Eclipsed and Turtles were fitted on second picks (real 59.4%, 60.0% and 63.0%, here 60%, 59% and 62%). First packs here show all five colors: Avatar 62%, Lorwyn Eclipsed 59%, Turtles 63%, and Reality Fracture 55%, which has no real packs yet and uses 70%, about the six sets' average. The fit is rough in its details: on four of the six sets the packs here have fewer four-color and more three-color packs than real ones, and real Strixhaven packs split more sharply (17.1% show three colors or fewer, against 11% here). Duplicates follow the set: never in The Hobbit, as in real packs, while Turtles packs repeat a card 9.8% of the time here (8.2% in real packs) because their foil slot draws on its own. Bonus sheets: Special Guests and source-material cards are in, at the rate real Arena packs carry them (per pack on Arena: Avatar 0.039, Lorwyn Eclipsed 0.019, Marvel Super Heroes 0.043, Turtles 0.035, Strixhaven one in 55; paper packs use the published 0.04, 0.02, 0.05, 0.09 and one in 55), each card at its rarity on its own sheet, as 17Lands' data lists them. Arena rates: on Arena the wildcard's rarity odds and the rare slot's mythic chance are solved from the 15,000 real packs so each rarity's copies per pack match (the wildcard keeps its published split between rare and mythic): wildcard common, uncommon, rare, mythic and the rare slot's mythic chance, The Hobbit 74.5%, 3.6%, 19.4%, 2.6%, 12.8%; Avatar 7.3%, 65.8%, 23.1%, 3.8%, 16.9%; Lorwyn Eclipsed 19.7%, 54.2%, 23.7%, 2.4%, 18.9%; Marvel Super Heroes 12.4%, 65.9%, 19.5%, 2.2%, 19.4%; Turtles 20.2%, 65.7%, 12.5%, 1.6%, 28.8%; Strixhaven keeps its published wildcard with a 15.3% mythic chance and a foil that is an Archive uncommon 3.7% of the time. Each rarity then comes within 0.01 cards a pack of the real packs. Paper packs keep the published odds: Avatar's wildcard is 4% common, 74% uncommon, 18% rare, 3% mythic; Lorwyn Eclipsed 18%, 60%, 20%, 2%; Marvel Super Heroes 12%, 67%, 19%, 2%; Turtles has two uncommons and a Legendary Turtle slot of its own. Special-frame versions in a slot count as that card's rarity. A set without an article here would use the general Play Booster shape (seven commons, three uncommons, a wildcard at 18%, 58%, 19%, 2%).

The pick model

Every seat at the table picks with a choice model fitted to real drafts (draftmodel.py, on 17Lands' public Premier Draft data). Each card has a pick value, and a card gets a bonus, beta, when it fits the drafter's colors: every color it needs is in one of the drafter's top-two color pairs (a pair whose smaller count is at least the largest count outside it; colorless cards always fit; a land needs the colors it makes; a hybrid symbol needs either color). Beta grows through the draft, in six phases by overall pick (1 to 3, 4 to 8, 9 to 14, 15 to 22, 23 to 30, 31 to 42), from about 1.6 to 2.1 logits at the start to about 3 to 3.4 in pack three. The table samples from the model's odds rather than always taking its favorite, so no two tables draft alike, the way people don't. Strengths and beta are fitted twice: on all users, and on top users (a game win rate of 62% or more).

chance of taking card c proportional to e^(value[c] + beta[phase] x fits(c))

The data behind each set:

• Lorwyn Eclipsed: 12,195 drafts (426,825 picks) sampled from 99,290 in the file of 18 May 2026; on drafts it never saw, the model names the card taken 55.7% of the time, against 38.8% for the old pick-order rating and 18.5% for a blind guess
• The Hobbit: 11,937 drafts (465,543 picks) sampled from 43,650 in the file of 03 Sep 2026; on drafts it never saw, the model names the card taken 60.0% of the time, against 41.1% for the old pick-order rating and 17.3% for a blind guess
• Marvel Super Heroes: 12,071 drafts (470,768 picks) sampled from 68,633 in the file of 27 Jul 2026; on drafts it never saw, the model names the card taken 59.0% of the time, against 42.8% for the old pick-order rating and 17.3% for a blind guess
• Secrets of Strixhaven: 11,820 drafts (460,980 picks) sampled from 107,502 in the file of 05 Jun 2026; on drafts it never saw, the model names the card taken 58.5% of the time, against 41.1% for the old pick-order rating and 17.3% for a blind guess
• Avatar: The Last Airbender: 12,054 drafts (458,052 picks) sampled from 102,793 in the file of 10 Jan 2026; on drafts it never saw, the model names the card taken 57.8% of the time, against 41.4% for the old pick-order rating and 17.6% for a blind guess
• Teenage Mutant Ninja Turtles: 12,134 drafts (460,932 picks) sampled from 32,151 in the file of 08 Apr 2026; on drafts it never saw, the model names the card taken 60.2% of the time, against 42.4% for the old pick-order rating and 17.7% for a blind guess

What the accuracy does and doesn't show. Most of the model's gain over the old rating is the color term: the old rating plus a fitted color term comes within 1.6 to 3.4 points of it, and the pick values alone are only 0.4 to 1.4 points more accurate than the old rating. The top-user fit predicts top users only 0.2 to 0.6 points more accurately than the all-users fit. 17Lands users win more games than the average Arena drafter (about 56%), so a table built from them is modeled on drafters with a higher game win rate than a random Premier pod.

Making the model draft like people

A model fitted one pick at a time doesn't have to behave like people over a whole draft, and this one didn't. Sampled pick after pick, it commits too soon: in simulated Secrets of Strixhaven pods (one run of 300 drafts) its drafters locked their colors at pick 6.9, against 9.9 for real 17Lands drafters, and took a rare with the first pick of a pack 43% of the time, against 48%. So every table scales the color pull in pack one and in pack three, and adds a small pull toward a rare on the first pick of each pack, fitted in simulation to each set's measured color lock (the first pick from which a drafter's two main colors never change again), pack-three picks outside the final colors, and first picks. The fitted pack-one pull ends at between 12% (Lorwyn Eclipsed) and 49% (Secrets of Strixhaven) of the model's own value, pack three at 1.10 to 1.24 times it, and the first-pick rare pull at minus 0.03 to plus 0.46 logits [calibrated].

How close the typical table comes [simulated: 16 seeds of 400 Strixhaven drafts each; each figure is the average, then the range across seeds, and a new seed can land a little outside it]: the seven other seats take 49.9% of commons (49.5% to 50.3%), 69.2% of uncommons (67.7% to 70.1%) and 84.5% of rares (83.2% to 86.2%) that pass through them (real pods: 49.0%, 70.9%, 85.9%); colors lock at pick 9.9 (9.6 to 10.2; real 9.87); a first pick takes a rare 47.6% of the time (46.7% to 48.5%; real 47.9%); 23.6% of pack-three colored picks fall outside the final colors (22.9% to 24.2%; real 23.4%). The Hobbit, Marvel Super Heroes and Avatar lock at 9.4 (9.1 to 9.7), 11.1 (10.8 to 11.2) and 12.8 (12.4 to 13.4) over 8 seeds of 300 drafts (real: 9.60, 11.08, 12.98). The rare take rate runs about 1.4 points low on average.

The three tables

Typical players are the calibrated model above. About one seat in six (the share of each set's drafts in the top win-rate bucket, 12% to 20%) picks with the top-user pick values and beta. Paper tables use the same model: no paper pick data exists.

Arena bots. The only public Quick Draft pick data is Crimson Vow's (2021, about 7,700 drafts). Measured there with the wheel method (the pack after your pick against the same pack when it comes back), the seven bots took 99.5% of rares and mythics that passed them, against 91.1% for people; commons and uncommons at about the same rates as people (49.7% and 74.8%, against 49.7% and 73.7%); take rates further from a coin flip (a card's take rate sat 0.325 from 50% on average, against 0.259 for people); and a pick order that ranks cards like people's with a correlation of 0.81, with large differences on single cards. On Phyrexia, every rare went earlier with bots, and a Quick Draft showed about 4 rares against about 7 in Premier (Sierkovitz, MTG Arena Zone). The bots here are the typical-player model plus a pull of 3.0 logits toward rares and mythics, a fixed quirk for every card drawn once per set (spread 0.7), and a lean of 0.3 toward one color per bot, all divided by a temperature of 0.5, tuned in simulation to those numbers [simulated: 16 seeds of 400 Strixhaven drafts; average, then range]: rares 99.8% (99.6% to 99.9%; Crimson Vow 99.5%); commons 48.9% (48.3% to 49.6%) and uncommons 68.5% (67.2% to 69.5%), against the typical table's 49.9% and 69.2%; take rates 1.30 times as far from a coin flip as people's (1.22 to 1.37; Crimson Vow 1.25); rank correlation with the typical table 0.82 (0.79 to 0.84 across seeds and two card-count cutoffs; an independent re-run counting cards seen 60 or more times found 0.80 to 0.84; Crimson Vow 0.81). With bots you see 7.1 rares a draft (6.9 to 7.3) against 14.8 with people (14.6 to 15.1), a ratio of about 0.48 (Phyrexia's was about 0.57). Limits: these are 2021 bots, though the same pattern shows in 2019 (War of the Spark) and 2023 (Phyrexia); Arena doesn't publish its bots; the quirks are random, not the real bots' card-by-card differences, which can't be sourced for current sets (17Lands shows Quick Draft data on its site, but only its public files may be used elsewhere); the color lean is an estimate from Sierkovitz's observation of a red-leaning bot. The bots lock their colors at pick 8.4 here (8.2 to 8.7); their real lock isn't measured.

bot utility = (pick value + quirk + 3.0 if rare or mythic + 0.3 for its color + beta x fits) / 0.5

The pro table is modeled on top 17Lands players, and sees nothing a real player couldn't. Measured: pros take cards with the top-user fit's chances, so there's no added pull toward rares, and mythics come less often (in Strixhaven, top players took mythics they saw 50.9% of the time against 56.0% for players under a 50% win rate). Estimated [every number in this paragraph]: they pick at a temperature of 0.6, so with less randomness. They choose a lane: each color pair scores its pool value (the pro's cards in that pair, counted above the set's median common), plus 1.2 times how much its colors are flowing (for each color, how far the highest-value card needing it that reached the pro sits above the median common, later picks weighted up to eight times more, packs one and three kept apart from pack two), plus, in a settled format and for three of the seven seats, 0.8 times how much top-player value the pair's commons and uncommons carry (a stand-in for archetype win rates, which the site doesn't have yet). A card fits as well as the top-scoring lane it belongs to, scaled by e^(minus the lane's gap to the top-scoring lane over tau), with tau 8, 4, 2, 0.9, 0.6 and 0.45 by phase, so pros stay open early and commit late. In packs two and three they fill counts: 0.45 more for a creature while creatures are under 15 of 23, for a two-drop creature under 6 of 23, half that for removal under 4, and 0.45 less for a five-drop over 4 of 23. They make denial picks (hate drafting) only where you play within the pod (paper), in pack three: 0.6 times the chance of meeting that player (3 of 7; 2 of 3 in a four-player Pick-Two) times how far the card's value tops their highest-value on-plan card by more than a point, for cards in the colors of the seat they pass to.

pro utility = (top-player pick value + top-player beta x lane fit + counts + denial) / 0.6

What the pro table changes for your seat [simulated: 16 seeds of 400 Strixhaven drafts, a drafter with the top-user fit in your seat, the same packs at each table, pools valued by top-player pick value (an average common is 0); average, then range]: the 23 highest-valued nonland cards your seat ends with average 0.73 (0.68 to 0.76) at the pro table, against 0.89 (0.85 to 0.92) at the typical table. Among the eight pools at the table, yours ranks 6.3 of 8 at the pro table (6.1 to 6.5), 4.3 at the typical table (4.1 to 4.5) and 4.7 with bots (4.5 to 4.9); it is the highest-valued pool 3% of the time (1.5% to 5.5%), against 14% (12% to 17%) and 11% (8.5% to 14.5%), and in the lowest two 59% of the time (53.5% to 64.5%), against 23% (19% to 26.5%) and 27% (24% to 32%). A Competitive REL pro pod, where pros may make denial picks, ranks your pool 6.3 (6.1 to 6.5). The pros' own pools average 1.08 (1.07 to 1.09), against 0.84 (0.83 to 0.86) for typical players, and they lock their colors at pick 11.4 (11.0 to 11.6), against 9.9 (in 17Lands' data, mythic-rank drafters lock at 10.7 and bronze at 9.3). A negative result: signal reading, as modeled, doesn't raise the value of a pro's own pool. A lone pro at a typical table ended with 1.264 with it and 1.264 without (one run of 800 drafts), and with heavier reading slightly lower. What it changes is when pros commit and how often they end in your colors; the difference in their pools comes from card values and less random picks. That fits 17Lands' own simulation, where reading raised win rates at tables that commit early and not at tables full of readers.

Format age

Settled format: the fit over the whole format, all weeks. First week: typical players get 1.3 times the pack-one color pull and 0.12 more pull toward rares and mythics, so in simulation they lock their colors 1.05 picks sooner on average (0.71 to 1.38 across 16 seeds of 400 drafts; an independent re-run with other seeds found 0.90 to 1.63, average 1.19). In Strixhaven's first week real drafters locked at pick 9.3, against 10.4 later, and took a rare first slightly more often: 46.9% against 45.1% [estimate of the mechanism; the target is measured]. The pro model drops its lean toward particular pairs, since those aren't known in a new format. Arena's bots don't change: they don't adapt, which is why their card-by-card differences last all season.

Sets without data yet

17Lands publishes draft data about two weeks after a set reaches Arena, and game data about three; Reality Fracture reaches Arena September 29, so its data is expected around October 13. Until a set has a model, the table rates its cards from your Eval Sheet grades (once 40 are graded), a 17Lands table you imported in Card Eval, or a reading of the card text (rarity, removal, stats for the cost, evasion, card draw), turned into logits at 1.513 per rating point. The pull toward colors is the average of the six fitted sets' "rating plus color" fits (beta rising from 1.52 to 2.69), with the six sets' average calibration. The page says which source it's using; text ratings are estimates from the text alone.

pick value without a model = 1.513 x rating (0.5 to 5 scale)

Pick review and the signals check

The review states what the models give, not whether a pick was right. Each pick is replayed through three drafters in your seat, with your picks so far: the top-player model (the pick model fitted on 17Lands drafters with a 62% or higher game win rate), the typical-player model (all 17Lands users), and the model behind the table you chose. The row shows how often top and typical players take your card in that spot; the top-player model's first choice and, when it isn't yours, how often the model takes it over yours given just those two cards; what this table's model would take; the card in the two colors you ended in with the highest top-player value, marked above or below your pick; what came back when the pack wheeled (a card seen at pick N returns at N plus the number of seats); and how long you took. The summary counts the picks that matched the top-player model's first choice, and the picks that were in your final colors and the highest top-player value there (an off-color pick never counts). For a set with no fitted model (Reality Fracture today) or when you pick another source, every row, chip and note names that source instead: the text-based estimate (the site's rough reading of the card text), your Eval Sheet grades, 17Lands' average pick order or a table you imported. Those are ratings, not data about how people draft the set, and the page never calls them top players or measurements. In Pick-Two both cards of a pass are listed, the second against the pack after your first. A card handed to you as the last of a pack is listed but not counted. With the ratings switched on, each card shows the chance this table's model in your seat would take it, and screen readers hear it as part of the card's name. The signals check, after pack one, asks which one or two colors had the most high-value cards passed to you, then counts the cards above the set's median common (by top-player pick value) that the two players passing to you took, by color (one player at a two-seat cube table); the two colors they took the fewest of are marked, and your call is compared with them. It also shows where the highest-ranked card of each color that reached you from upstream ranks in the set (100 is the set's highest).

The build put together by counts

A build the page puts together by counts, not a recommendation. For each pair of colors, the 23 cards with the highest top-player pick value that the pair can cast (colorless cards included; a hybrid symbol like {W/U} can be paid with either color, and a two-faced card counts its front); the pair whose 23 add up to the highest total is used. If the pair has fewer than 23 castable cards, the third color whose cards add the most value fills the rest (a splash). Up to three nonbasic lands that make both colors come along, and basics fill to 17 lands, split by the colored mana symbols on the fronts of the deck's cards; a hybrid symbol counts toward the colors the deck plays, so it never adds a basic of a color the deck doesn't. When the pool can't reach 23 spells, the line says how many it found and how many are missing (for example 14 of the 23 spells a 40-card deck with 17 lands needs; add 9 more). A deck under 40 cards says how many it's short. Tap any card to move it between the deck and the sideboard. The table tab shows every other seat: its model, its colors, the pick its colors locked, the rares it took and its four cards with the highest top-player value.

The Admiral Phase (Captain drafts)

From the Captain league charter (Chad, September 29, 2026): "Admiral Phase first: snake-draft Captains before anything else." On a cube, switch on "Admiral Phase first" and choose the Captains each (8 to 12, 10 unless chosen). The opening part of the draft deals only legendary creatures that can take the Seat (a legendary creature by its front face, in the Captain card pool), then the main draft runs as usual from the rest of the cube, where the legends the phase didn't deal still turn up. The setup states how many legends the phase uses against how many the cube has, and won't start short.

• Booster draft and Pick-Two: one round of packs of legends first, one pack a player, the pack size set to the Captains each, so each player takes exactly that many.
• Rochester: one round of face-up legend packs of that size, picked around the table and back.
• Rotisserie: the first players x Captains picks of the snake come from the legends only, all face up.
• Winston: a stack of twice the Captains each in legends first, drafted by Winston's own rules (about that many each; the split depends on the piles), then the main 90.
• Grid: the fewest legend grids that give each player at least the Captains each (a grid gives the first picker 3 and the second 2 or 3), then the main grids.
• Sealed: a separate pool of legends of that size beside the six packs.

legends used = players x Captains each (Winston: 2 x Captains each; Grid: 9 x grids); legends left for the main draft = the cube's legends - legends used

Grid: grids = the fewest g with floor(g / 2) x 3 + ceil(g / 2) x 2 >= Captains each

The other seats take a legend with the rule of the rest of the draft (chance in proportion to e^(its cube rating plus the pull toward the seat's top two colors)), less 0.5 on that scale for each Captain the seat already holds with the same color identity, so a stable spreads over identities [ESTIMATE: nothing published measures how people draft Captains]. In Comp (every cube unless it was built as a Cracked cube), a legend on Captain's banned list, the Reserved List or Banned as Captain is never dealt in the phase. A Captain draft skips the signals check (its first round is legends), and the pick review labels the phase's picks A1, A2 and on.

A Captain deck from the pool

In Captain Limited any creature in your pool may be your Captain. A Captain draft's deck tab starts built around one: of your Admiral Phase Captains, the one whose forty adds up to the highest card values (one that keeps the deck Comp first), and "switch Captain" lists every creature in the pool with the spells and lands its forty would hold. Any other pool has "build a Captain deck from this pool", with the same list and a preview before it opens in Decks.

the forty = the Captain + the 22 spells with the highest card values inside its color identity (one of each) + 17 lands: up to 6 of the pool's lands in its colors and 2 colorless ones (4 for a colorless Captain), then basic lands split by the colored mana symbols of the Captain (counted twice) and its spells; Wastes for a colorless Captain

A pool with fewer than 22 spells inside the Captain's colors says how many short it is, and basic lands fill the forty; nothing is filled silently. The rest of the pool is the sideboard, as in any limited event. The deck is saved as Captain Limited with the Captain in the Seat, named short: "Carnage (Captain draft, Oct 1)".

Pass and play

A cube night for two can be drafted by two people on one device: Winston, Grid, Rotisserie, or a booster or Pick-Two draft with two players. Each turn ends on a hand-off screen that shows neither pool; the next player taps to see their turn. In a booster draft each player takes their card from the pack in front of them, and the one holding the device picks first from the next pack, so the device changes hands once a pass. No clock and no signals check run. At the end both decks are built (around a Captain in a Captain draft), both pools go to past drafts, and a switch shows either deck. Logging a match can log it for both decks, the result reversed for the other.

Two devices can't share one live draft: the site has no accounts or server to carry picks between them.

The daily pick

One first pick a day from a real pack of a set with a 17Lands pick model (The Hobbit, Marvel Super Heroes, Secrets of Strixhaven, Teenage Mutant Ninja Turtles, Lorwyn Eclipsed, Avatar: The Last Airbender, one a day in turn). The pack is an Arena pack dealt by the same pack builder as the drafts, from the calendar date where you are, so everyone on the same date gets the same pack. You pick; then the page shows, for every card in the pack, 17Lands' average pick (the pick number real drafters took it at, over every pick of the card in the file, with how many times it was taken and the file's date) and the pick model's odds for this exact pack as a first pick, for all drafters and for top players, with where your card sits in each order, and no grade.

model odds for card c = e^(pick value(c) + beta at picks 1 to 3 x on2(c)) / the same summed over the pack; with nothing picked, on2 = 1 unless no one color pair holds what the card needs

Pick value is the model's fitted value for each card (for top players, their own fit); a card with too few picks in the data takes its rarity's average.

Your picks, the day streak (days in a row with a pick, up to today or, before today's pick, up to yesterday) and the history stay in this browser. "Copy the result" gives the facts as text to share.

Ask friends: a link carries a pack's card names, the set, who asked and, if they choose, their own pick, shown to a friend only after the friend picks. The friend's answer link carries their name and pick; opened in the asker's browser it is counted with the pack, once (a newer answer under the same name replaces the older).

Saved drafts and seeds

A draft saves after every pick, with the names of the cards its numbers point to. If a set's card list changes while you're drafting (a new model, a new bonus sheet), your draft carries on, matched card by card; a card the set's list has dropped comes back from the site's card list for that draft, and only a card the site no longer knows at all would set it aside. Drafts saved by the version before September 24, 2026 carry on too, on the packs they had dealt, as a paper store draft with no clock against typical players, the closest event to how they ran. Each seed is saved with the version of the pack builder that dealt it; packs are built differently since September 24, 2026, so an older seed deals different packs now, and the page says so when you draft one again.

A draft in a link (rate my draft)

A finished booster draft of a set can ride whole in a link, for someone to review. The link is the page's address with #rd= and a code after it. Browsers never send the part after the # to a server, so nothing is uploaded and the link never depends on the site keeping anything. What rides in it: the set and a fingerprint of its card list, the event, the table you drafted against, the seats, the cards a pass and the picks a pack, the seed, the ratings source the review used, the day the link was made, every pack you saw with the card you took from it (both cards of a pass in Pick-Two), how long each pick took, which picks were handed to you or timed out, your deck (its spells, basics and two colors), a note for the reviewer (up to 1,000 characters) and your notes on single picks (up to 280 characters each). Notes are typed beside each pick in the pick review and saved with the draft. The code is that record as JSON, compressed with the browser's own deflate (CompressionStream, deflate-raw) where the browser has it, then written in URL-safe base64; a typical draft makes a link of about 1,400 characters.

Left out: cube drafts (a cube lives only in its maker's browser, so the link couldn't be opened anywhere else), Sealed pools (no picks to review), pass-and-play drafts, Captain drafts with an Admiral Phase, and drafts not finished. The other seats were the site's modeled drafters, never people, so no one else is named.

Opening a link shows the same pick review, read only, worked out again with the opener's copy of the set's model; nothing on that screen is saved, and the opener's own draft in that browser is left as it was. If the drafter's ratings came from something only their browser holds (their Eval Sheet grades or a table they imported), the review uses the set's first source instead and says so. If the set's card list has changed since the link was made, the page says a card may show differently. A link longer than 24,000 characters, or one whose cards fall outside the set's list or whose taken card isn't in its pack, isn't opened, and the page says why. A pasted second link in the same tab opens fresh.

With friends, by link

Rotisserie by link: the whole cube face up, one pick at a time in snake order. Each link carries the draft: its players, picks each, the cube's fingerprint (two 32-bit hashes of the sorted card list) and every pick in order (each pick is its place in the sorted list, packed in as few bits as the cube needs), with a checksum. The first link also carries the card list: each card as its number in the site's card index (the gaps between the sorted numbers, Rice-coded, about 9 bits a card), with the index's version, and by name only for a card the index doesn't have under that name. A link from before a site update whose numbers no longer match the list's fingerprint says so and asks for a fresh one. Measured October 23, 2026: a 360-card cube's first link is 612 characters, under Discord's 2,000. A link that adds picks, ends the draft or opens an ended one again shows what it changes and applies it only on a tap, with undo. A link is compared with what this browser holds: newer adds the picks, older is kept back and says so, and two links that pick differently from the same point show both and let the group choose. Nothing is sent to a server; each browser keeps the drafts it opened. Discord holds 2,000 characters a message (4,000 with Nitro), so the page says when a link is past that and offers a file.

Sealed pool

Packs are opened from MTGJSON's booster sheets (MIT license; the file's date is on the page): each pack picks a booster layout by its weight, then each slot's cards by their sheet weights, so a seed opens the same packs again. Basic lands from the packs are set aside, since sealed basics come from outside the pool. A pasted pool reads counts, set codes and foil marks; names the site doesn't know are listed. The first 40 is built by count: the 23 highest-rated cards one color pair can cast, then 17 lands split by colored mana symbols; card values are 17Lands' pick order for a set with data, otherwise the site's text estimate, and the page says which.

Sources

• Wizards of the Coast, "Collecting Reality Fracture" (Mike Turian and Jubilee Finnegan, September 8, 2026), and the "Collecting" articles for Secrets of Strixhaven, Teenage Mutant Ninja Turtles, The Hobbit, Avatar: The Last Airbender, Lorwyn Eclipsed and Marvel Super Heroes: the pack layouts. Vice, "Echoed Pairs Explained": what a pair is.
• Magic Tournament Rules, effective February 27, 2026 (7.2, 7.6, 7.7, Appendix B): paper draft and sealed procedure, pick clock, review periods, build times. Wizards, Pick-Two Draft.
• Wizards, MTG Arena draft guide, the Marvel Super Heroes and Reality Fracture Arena event schedules, and ComicBook's coverage of Arena's draft timer at the Ikoria launch (April 16, 2020): Arena events and clocks.
• 17Lands public datasets, CC BY 4.0: Premier Draft data for the pick models and the pack checks, Crimson Vow Quick Draft data for the bots, Pick-Two data for its shape. Data from 17Lands; 17Lands doesn't endorse this tool. Only the public files are used. 17Lands blog: "Do the Bots Send Signals?" (2019), "Learning from the Best" (2022), "Simulating Draft Strategies".
• Sierkovitz, "Phyrexia: All Will Be One Quick Draft guide", MTG Arena Zone (2023): bots and rares. Draftsim, "Quick Draft vs Premier" and "Pick-Two Draft" guides.
• Ben Stark, "Drafting the Hard Way" (ChannelFireball, 2013) and Limited Resources 507; Paulo Vitor Damo da Rosa, "Drafting the Correct Way"; Reid Duke, "Signals in Booster Draft" (2015); Melissa DeTora, "Drafting 101: Understanding Signals" (2016): the pro doctrine behind the pro table. Frank Karsten: the land and curve counts.
• Henry N. Ward, Daniel J. Brooks and Dan Troha, "AI solutions for drafting in Magic: the Gathering" (2020, arXiv 2009.00655): the DraftsimBot this page used before, and the accuracy yardstick (44.5% for DraftsimBot on its own data).