September 21, 2026

Reality Fracture draft primer

Reality Fracture comes out on Arena on October 2. This primer puts two very different kinds of evidence side by side:

• The simulation. Bots drafted the set 800 times (100 pods of eight), built decks and played 100,000 games against each other under the full rules.
• The experts. Nine set reviews from four outlets: Limited Resources (Marshall and Luis), Limited Level-Ups (Marc and Alex), Lords of Limited (Ben and Ethan), and the written reviews at TCGplayer (LSV and Martin Jůza). We summarize and paraphrase them, with credit. Listen to or read the originals for the real thing.

Neither has seen a real Arena draft yet. The bots have played the set a lot but play worse than people; the experts play better but have only played a little. Where they agree, trust it. Where they don't, this page says who has the better case.

This run is on the complete spoiler, and for the first time every card in the set is actually simulated. Fifteen cards — thirteen of them rare or mythic — used to be missing, because the engine could not play them. Two of the fifteen turned out to be among the best uncommons in the format, so if you read an earlier version of this page, the card advice below has changed.

Each claim is tagged with how much to trust it:

• Solid: a big effect, clear of the noise, and the kind of thing the simulator gets right on sets humans have played.
• Lean: real in the simulation, but small or in an area where the simulator is weaker.
• Unproven: likely to be wrong; see why.

Want a specific card or pack? The card ratings (with every host's grade on hover), pick helper and color pairs tools use the same data. Once the set is out, they switch to real 17Lands results by themselves. Drafting on Arena? The live pick helper runs on your computer and reads each pack for you.

The short version

1. Red-green is the best pair. The simulation has it at 53.2%, alone at the top, and it is also the experts' favorite. The next eight pairs sit between 49.1% and 51.0%, so outside red-green, being in the open lane matters more than the pair you pick. Solid.
2. Avoid blue-black. Worst in the simulation (45.7%) and the lowest-rated pair among the experts too, though only Lords of Limited scored it. Solid.
3. Don't trust either side on blue. The experts rate it the best color; the simulation ranks it last. Bots misplay flash, bounce and counters, so the truth is probably in between. Unproven.
4. The format is slow. Every show that has played it now says games go long, and the simulation rewards sturdy bodies over tempo. Lean.
5. Choose to play first. The player on the play won 53.4%. Solid.
6. Play plenty of creatures, but the effect is smaller than it looks: decks with 20 or more won 57.5% and decks with 13 or fewer won 46.4%, yet almost all of that is card quality rather than the creature count itself. Lean.
7. Green and red are the deepest colors. White is in 58% of bot decks and wins the least of the three you would actually want to be in. Lean.
8. Late picks worth watching for: both Garruks, Overwrite the Multiverse, Way of the Wildspeaker and Multiply by Zero. Solid.
9. Do not first-pick Emrakul. Ten mana is ten mana. Solid.
10. Keep the color hosers in the sideboard. Every show says so, and Flourishing Grapple wins 36.6% when drawn in bot main decks. Solid.

Colors and pairs

Each pair's win rate against the whole field, with 95% intervals.

Red-green (53.2%) is alone at the top, and its interval (52.6–53.8%) does not touch the next pair's. That is a real change from the partial-spoiler version of this primer, where the top four were inseparable. Black-red (51.0%), black-green (50.8%) and blue-red (50.5%) follow, then five pairs between 49.1% and 49.9%.

Blue-black is still the pair to avoid (45.7%), and it is now further adrift than before: more than three points below the next-worst pair. Only 4% of bot decks ended up there.

By color, red (51.0%) and green (50.8%) lead, black (49.8%) and white (49.4%) follow, and blue (49.0%) trails. The number that matters for your seat is not the win rate but the traffic: 58% of bot decks played white against 30% for blue. White is the most contested color in the format by a distance, so the white cards you see late are the ones everyone else already passed.

• Blue may be better than it looks. On The Hobbit, where real results exist, this simulator underrated blue decks. So "blue is last" here probably means "blue is playable". Blue-black's last place is the weakest call in this primer. Unproven.
• Red-green is a creature deck and the simulator is at its best judging creature decks against each other, which is part of why it is confident here. Expect the gap to be real but smaller once people are drafting it. Lean.

What the experts expect

We had each host's verdict on every pair and color scored from −2 (avoid) to +2 (the best deck), using only what they actually said, and averaged over the hosts who judged it. To draw that on the same scale as the simulation, the experts' average is stretched so its spread across the ten pairs matches the simulation's. So read the chart for order and gaps, not the exact percentage: the level is borrowed, the ranking is the experts' own.

Circles: simulated win rate with 95% intervals. Diamonds: the experts' expectation; the shaded band runs from the most to the least bullish host. The number after each pair is how many hosts judged it. Hover for each host's score.

The two sources agree on the ends and little else. Rank agreement across the ten pairs is ρ = +0.10, close to none. Both put red-green first (four hosts, average +1.5, and Marc calls it the best pair) and blue-black last. In between, the experts like the white pairs far more than the bots do: Azorius (+1.0) and Boros (+1.0, from Ben alone) sit near the top of the experts' list but at 49.1% and 49.5% in the simulation. Black-red is the most contested pair among the hosts: Ben, Marshall and Luis like it, while Marc, Alex and Ethan don't. The simulation has it second at 51.0%, but on only 5% of decks, which fits the hosts' "only when it's open".

By color it's worse: ρ = −0.10. Blue is the experts' best color (+1.2 over five hosts; Marc gave it the top score) and the simulation's worst. Black is the experts' worst (−0.6) but middling in the simulation, with two of the better removal spells (Extended Absence 56.1%, Multiply by Zero 54.2%). Red and green are near the top for both.

Who has the better case? On blue, probably the experts, partly: this simulator underrated blue on The Hobbit, where real results exist. On white, probably the simulation: the four white pairs all land around 49–50% while white is in 58% of bot decks, so white looks overdrafted rather than strong.

Which colors run deep

How many of each color's commons beat the set's average card when drawn, in the simulation.

Commons are what you see all draft. Green has 8 of its 11 commons above the average card, and white 8 of 15. Blue has only 4 of 14. That's another reason not to expect blue to flow, even if its best cards are better than the bots think.

The color pairs tool shows each pair's best commons and uncommons, measured in that pair's decks, and a sample deck.

How fast is it?

Early previews said fast: Limited Resources expected quick games where bounce shines. Every show that has played the set since (Limited Level-Ups' rare review, Lords of Limited's early guide, Martin at TCGplayer) says the opposite: slow and grindy, games running to turn six and beyond, seven- and eight-drops getting cast, mana sinks overperforming. You still want early bodies, to attack opposing Jace tokens and protect your own.

The simulation fits the slow read. Sturdy bodies overperform (Wrecking Gecko 56.9%, Thalia, the Survivor 60.7%) and tempo cards lag (Unsummon 47.9%). Lean, because bots are bad at tempo either way.

What to take

Commons

The three best commons in each color by win rate when drawn.

Bestial Incursion (a 4/4 trampler for four that comes back with flashback) and Blossom-Blessed Angel (a 2/4 flier with vigilance plus a one-mana +1/+1 counter) are the best commons in the set. Take either over most uncommons.

Green runs deepest: five of the twelve best commons are green. After Bestial Incursion come Wrecking Gecko (a 5/5 ward for five), Greenhouse Propagator and Arcane Amphisbaena (a two-mana deathtouch body plus a Jace). In black, Extended Absence (exile a creature for four) and Screeching Soulbreaker are both strong. In red, Heartstring Puller (a 3/1 trampler that brings a 2/2) and Tether Technician (a 4/5 with reach that can deal 2 when it enters) are the picks.

Uncommons

Kiora of Fire and Ashes (six mana for a 2/2 and a 5/5 flying dragon, 61.9%), Thalia, the Survivor (a 3/4 lifelinker that taxes their spells, 60.7%) and Tetsuko Umezawa, Pursuer (a 2/4 double striker, 60.6%) lead.

Then come the two the simulator could not play until now. Ghalta the Unstoppable (59.5%) costs {8}{G} on paper, but it costs {X} less for your greatest power, so in a green deck it lands for five or six as an 8/8 trampler that gives your whole board trample. Ghalta the Immovable (57.3%) is the stranger one: a 0/7 for {8}{W}, discounted by your greatest toughness, which lets every creature you control attack as though it had no defender and assign damage by toughness instead of power. A 0/7 that attacks for seven. Both were absent from the last version of this primer — the second one was even listed as a trap — and both are now top-seven uncommons. Craftwork Crusher (58.7%) and Twisted Fates (57.5%) round out the tier, and Your Fate Ends Here (54.6%) is still the best removal spell at the rarity.

Sleepers

These won far more than their pick position says, and their 95% intervals clear the set's average. Bots took them around pick 8 or later.

Garruk, Curse Breaker (61.5% at pick 8.4) is the standout, and Garruk, Veiled Butcher (56.4% at pick 8.3) is close behind: both are mythics, so you will rarely see them late, but the bots pass them when they do. Overwrite the Multiverse (56.8%) is a six-mana wrath that leaves you a huge Jace. Way of the Wildspeaker (55.2% at pick 10.7) makes a 7-loyalty Jace that can −4 for a 4/4 trampler the turn it lands. Multiply by Zero (54.2% at pick 10.7) is cheap interaction the bots keep letting go.

Way of the Warlord, a standout in the previous version, has come back to the pack at 52.4%.

Traps

These go early and don't win.

Emrakul, the Exigent Doom is the clearest trap in the set, and the bots fall for it completely: they take it with pick 1.0 on average and win 42.7% of the games they draw it. It is a 12/12 for ten mana with a plan for getting there — exile it from your hand for {3} to turn a land into a two-mana-producing land, then cast it from exile — and in a 40-card deck that plan still does not get there often enough. Take the playable card instead.

Proft, Sinister Mastermind (44.1% at pick 4.6) reads like a three-mana 5/5 menace, but you cannot cast it until your graveyard holds seven cards, which is late or never in Limited. Germinate Recruits (39.0%) does nothing unless you gained life that turn. Flickering Hound is a 2/2 for four whose trigger rarely matters. Return to the Light Realms costs seven and does not stabilise a board.

Where the experts and the simulation disagree

Card by card, they agree more than on pairs: across the 264 cards graded by at least two hosts and played in at least 300 simulated games, rank agreement is ρ = +0.43. That's real agreement, but with big exceptions, labelled here:

Each dot is a card: its average expert grade against its simulated win rate when drawn. Labelled: the five cards the simulation likes most relative to their grade, and the five it likes least.

The simulation's case (take these higher than the reviews say):

• Blossom-Blessed Angel (59.5% over nearly 19,000 games). Limited Resources and Limited Level-Ups both graded it C; Ben on Lords of Limited called it absurdly good after playing. Ben and the bots agree. Take it early.
• Thalia, the Survivor (60.7%). Graded C by Limited Level-Ups and B by Limited Resources. A lifelinking 3/4 that taxes their spells fits a slow format.
• Tetsuko Umezawa, Pursuer (60.6%). Alex graded it D and Marc C-. A 2/4 double striker is a strong body when boards stall.
• Wrecking Gecko (56.9%). D+ to C- from the shows, "garbage" in Lords of Limited's first impressions. Ward 2 on a 5/5 matters in a slow format full of removal.
• Ghalta the Unstoppable (59.5%) and Ghalta the Immovable (57.3%). Limited Resources graded them B- and C-/D+, Marc D+ and D-. Both cost far less than they look once you have one big creature (more below).
• Greenhouse Propagator (56.2%). Lords called it garbage at first; Ben moved it to C+ after playing. The simulation backs the revision.

The experts' case (the bots are probably wrong):

• Sphinx of False Conclusions (48.3%). Marc's best rare (A+), A from both Limited Resources hosts, LSV 4.5. Flash ambushes are exactly what bots misplay.
• Unsummon (47.9%). Limited Level-Ups calls it one of the best commons. Bots use bounce badly. Even so, the shows now say one is enough in a slow format.
• Edgar, Ancient Bloodlord (45.1%). B and B- from Limited Resources, 3.5 from LSV. Its value is in well-timed sacrifices the bots miss.
• Way of the Mind Sculptor, Perfected Theory, Rise of the Deathbringer: graded far above how they play in the simulation (Limited Resources gave the Way and Rise a B, though Perfected Theory only a D). The first two are partly simplified in the engine (see below), so don't read much into them.

Too close to call: Verdant Kraken (A or A- from four hosts, Martin up to 5.0 in sealed, but only B from Luis, who doubts it ends games through flyers and removal; 47.5% over 2,789 games) is probably a strong card rather than a bomb. Flourishing Grapple's 36.6% is what maindecking a color hoser looks like, which is why every show says not to.

The full list, with every host's grade and a paraphrase of what they said, sorted by the biggest disagreement:

Empower Jace and the Way cycle

Empower Jace is the set's main mechanic. It puts loyalty counters on your Jace token, creating one if you don't have one. Across its 33 cards it averages 1.3 points below the set's average card — slightly negative on its own, not the free upside it looks like. What decides a card is its payload. The clearest case is the uncommon Way cycle:

Win rate when drawn for each Way, with 95% intervals. The dashed line is the set's average card.

The experts can't agree on the best Way either. Alex on Limited Level-Ups and Ethan on Lords of Limited say Way of the Healer; Ben on Lords says Way of the Pyromancer, which Alex graded D+. The simulation has Healer dead average (50.3%) and Pyromancer at 45.1%. Bots manage Jace loyalty poorly, so all the Ways are probably better than these numbers.

Wildspeaker (55.2%) and Paradox (54.1%) are above average, Warlord (52.4%) and Deathbringer (52.1%) are just above, and Healer sits on it. Cryomancer, Mentor, Pyromancer and Necromancer are all below. Way of the Mind Sculptor looks worst of all at 31.7%, but its loyalty-removal draw is left out in the simulator, so don't read much into that one.

Two other patterns: flying creatures average +2.2 points over the set's average card, and surveil cards average −1.4 — which, like the blue result, is somewhere the simulator is known to be weak. Cards with prepare (creatures that carry a spell you can cast from them) sit at about average.

Building your deck

We compared the 800 decks' records against what was in them. We also held each deck's average card quality fixed, so "more creatures" doesn't just mean "better cards".

Deck win rate by number of creatures, with 95% intervals.

The raw trend is dramatic: decks with 20 or more creatures won 57.5%, decks with 16–17 won 51.6%, and decks with 13 or fewer won 46.4%. The average bot deck ran 15.

But hold card quality fixed and most of that disappears. Each extra creature is worth +0.26 points (±0.21) once you account for the deck's average card quality — a much smaller effect than the buckets suggest, and only barely clear of zero. This is a change from the previous version of this primer, which put it at +0.65, and the reason is that the deckbuilder got better at valuing cards: decks that used to be creature-light were also card-poor, and that confound is now smaller.

• Removal is not measurable here. +0.07 points per removal spell (±0.28), quality held fixed. The earlier +0.6 does not survive. That does not mean removal is bad — it means in this simulation good removal spells are just good cards, and the quality term is already counting them. Lean.
• Two-drops don't matter by themselves, same as before: −0.05 (±0.24). Don't take a weak two-drop over a strong four-drop to fix your curve. Lean.
• Choose to play first: 53.4% on the play (±0.3). Solid.

So: take the better cards. The old advice to force 16+ creatures is weaker than it looked — what those decks really had was better cards.

From the experts, on the mana: the gold uncommons are pushed but heavy. Triple-pipped cards like Hapatra, Twisted Fates and Mind Meanderer belong to their pair, not in a splash. Pick up dual lands, and count sources: Limited Level-Ups suggests 10–11 for double-blue three-drops and 12 or more for cards with four pips of one color.

The caveat from last time still stands: on The Hobbit, the simulator overrated decks full of big-bodied creatures, because bots don't punish a board of large creatures with tricks and evasion the way people do.

What the simulator can't tell you (yet)

• The whole set is in now, but 108 of the 285 cards are still simplified. Every card is playable and simulated; 108 have some piece of text the engine leaves out, marked † in the tools with the missing part on hover. A simplification only ever makes a card weaker, so treat those numbers as floors. Ferocity of the Hunt is the clearest case: without its return-from-death half, it's just a deathtouch aura.
• The bots' pick order is worse than their play. Emrakul is the proof: they take a ten-mana creature first pick, every time. Read a card's average pick as "what a card-counting bot thinks", not as what the pick should be, and be most suspicious of it on expensive cards.
• Combat tricks look terrible here (Charge the Sanctum and Academic Ascent are near the bottom of the commons). Bots use tricks worse than people, so these are underrated. Don't write them off.
• Blue and surveil are where the simulator has been wrong before. On The Hobbit it underrated both blue decks and instants. Discount the blue-black result accordingly.
• The experts haven't played much either. Most of the reviews were recorded before anyone could draft the set; the shows that have played it already changed their minds on speed. Their pair scores are opinions turned into numbers by us, not measurements.
• Nothing here has met a real draft deck. The bots have never played an Arena deck from this set, because none exist yet.

How this updates

The tools refresh on their own. Once a day, they fetch Reality Fracture's live 17Lands data. When real players have enough games with a card, its real win rate replaces the simulated one, and the tools say which numbers are real. After release, we'll check this primer against 17Lands in a new post.

Expert sources: Limited Resources #872 and #873 (commons and uncommons, rares and mythics), Limited Level-Ups (first impressions, commons and uncommons, rares and mythics), Lords of Limited #496–498 and TCGplayer's set reviews. Every take here is our paraphrase.

Raw data: fra-draft-2026-09-21.json. Reproduce with mulligan draft --set fra --pods 100 --games 100000. How far to trust the simulator in general: How it works.