How Deck Power Level Is Scored

The power level page shows a bracket from 1 to 5 and a score from 0 to 10. The bracket is Wizards' official framework — a set of pass/fail checks on what a deck contains. The 0–10 score is ours. We invented it, we tuned the coefficients by hand, and no judge, rules committee, or LGS has ever heard of it. Do not quote it at a table as though it were official; quote the bracket.

The two move independently, which is the whole reason we show both. The bracket runs a short list of requirement checks: any Game Changer moves a deck to 3, more than three moves it to 4, mass land denial or a detected two-card combo forces 4, and Bracket 5 requires the full cEDH signature — at least 4 Game Changers and at least 3 pieces of fast mana and either 2+ combos or 5+ tutors. So a deck with 2 Game Changers and one two-card combo lands at Bracket 4: the Game Changers push it to 3, the combo forces 4, and it stops there because it has neither the fourth Game Changer nor the third fast-mana card. That same deck might score 8.6 or 6.1 depending on everything else in the 99. The bracket describes your ceiling; the score describes how reliably you reach it.

Speed is a subtraction problem. It starts at 10 and average mana value takes chunks out of it:

10 − max(0, avgMV − 1.8) × 2.6 + fastMana × 0.9 + min(3, ramp × 0.25)

Average mana value is computed over nonlands only, weighted by quantity. The 1.8 is the free-ride line — everything below it costs nothing, and every point of average mana value above it costs 2.6. A deck at 3.4 average is down 4.16 before it earns anything back.

Fast mana is a fixed 17-card list, not a heuristic: Sol Ring, Mana Vault, Grim Monolith, Chrome Mox, Mox Diamond, Mox Opal, Mox Amber, Lotus Petal, Lion's Eye Diamond, Simian and Elvish Spirit Guide, Ancient Tomb, Gemstone Caverns, City of Traitors, Jeska's Will, Dark Ritual, and Cabal Ritual. Mana Crypt and Jeweled Lotus are absent on purpose — they are banned in Commander and the list only contains legal cards. Note that lands can be fast mana (Ancient Tomb, City of Traitors) but lands are never counted as ramp; the ramp bucket only looks at nonlands that add mana or fetch a land.

Ramp saturates at 3 points, which arrives at 12 pieces. The thirteenth Signet is worth exactly zero. Fast mana does not saturate, so Sol Ring — which counts as fast mana and ramp — is worth 1.15 raw speed, about 0.29 of your final score, from one card.

What to change: cutting half a point of average mana value is worth +1.3 speed, or +0.325 final — which usually means trading one 6-drop for a 3-drop, not shaving the whole curve. With no acceleration at all, speed drops below 7 the moment a deck passes 2.95 average; add a Sol Ring and eight ramp pieces and that same deck can sit at 4.0 average and still score 7.18.

min(5, tutors × 1.1) + min(5, draw × 0.45)

Two halves, each capped at 5. Five tutors max the first half; twelve detected draw effects max the second. Everything past that is free deckbuilding but a wasted point.

A tutor is anything that searches your library for a card with no type restriction — Demonic Tutor, Vampiric Tutor, Diabolic Intent — or a typed search for a creature, instant, sorcery, artifact, or enchantment that has not already been classified as ramp. Land searches qualify on neither branch, so Rampant Growth and Cultivate count once, as ramp, and never inflate a tutor count — otherwise every green deck in the format would read as a tutor pile.

What to change: each of the first four tutors is worth 1.1 raw, or 0.275 of the final score — the highest-leverage card slot in the model outside of a combo. Each draw effect is worth 0.1125 final, up to twelve. A consistency stuck near 5 almost always means all draw and no tutors, or the reverse; the two caps punish going all-in on one half.

Honest limitation: draw detection matches the literal phrase "draw a/one/two/three/X/that many card", and that pattern has real holes. Necropotence never says it — it exiles cards face down — so the most efficient card-advantage engine in the format registers as zero draw. "Target player draws two cards" (Sign in Blood) misses on the trailing s, and "draw four cards" misses because four isn't in the list. If a deck that visibly draws fifteen cards a game posts a consistency of 6, this is usually why. Pattern lists have gaps; we would rather document this one than quietly inflate the number.

The flattest formula in the engine: interaction pieces × 0.75, capped at 10. Fourteen pieces max it out. A card counts once no matter how many boxes it ticks, and lands never count at all — a utility land's reminder text can say "destroy" without the land being interaction.

What counts is where this gets interesting. Three detectors run over full oracle text, including every face of split and double-faced cards: counterspells ("counter target"), board wipes, and targeted removal. Targeted removal is not just "destroy target". It also matches edicts — the pattern looks for a player as the immediate subject of a sacrifice, so "each player sacrifices a creature" on Fleshbag Marauder, "each player sacrifices a creature or planeswalker" on Plaguecrafter, and Chainer's Edict all register.

That matters more than it sounds. A mono-black deck answers the board for four turns without ever printing the word destroy, and a detector built only on destroy-and-exile scores it at literally zero interaction. The same blindness hits mass land denial: the Pox family forces every player to sacrifice a land, so the land-denial patterns span clauses to catch "each player loses 1 life, discards a card, sacrifices a creature, then sacrifices a land" — while stopping at commas so a Braids-style choice of artifact, creature, or land is not treated as Armageddon.

The exclusions are as deliberate as the inclusions. Sacrifice as a cost ("Sacrifice a creature:") does not match, because the pattern needs a player as the subject, and "each opponent may sacrifice a creature" does not match either — the word may breaks the adjacency, and optional punisher effects are not removal. Theft counts only when it is permanent: Control Magic matches, Act of Treason is filtered out by its until-end-of-turn clause. Also counted: bounce that does not return your own permanents, tuck effects like Condemn, Chaos Warp's shuffle, fight, and mass minus-X-minus-X like Toxic Deluge.

What to change: 0.1125 of final score per piece, straight line, up to fourteen. Not counted: static prison pieces and shutdown auras. Drannith Magistrate, Rule of Law, and Darksteel Mutation answer things beautifully and match no removal pattern, so a stax deck built on those will score lower here than it plays.

(combos > 0 ? 5.5 + min(3, combos × 1.2) : 0) + gameChangers × 0.35

There is a cliff here and it is intentional. With zero two-card combos this subscore is nothing but 0.35 per Game Changer. The first combo adds 6.7 in a single step — 1.34 of your final score — because a deck that can win from two cards in hand is categorically different from one that cannot, however similar the other 98 slots look. Three combos max the combo term at 8.5; Game Changers keep adding 0.35 apiece until the whole thing clamps at 10.

Combos are matched against Commander Spellbook using your exact decklist, then filtered to combinations of two cards or fewer. A three-card line wins games perfectly well, and you will find plenty of them in the lists on deck browse, but it feeds neither this subscore nor the bracket. Game Changers are synced live from Scryfall's tagging of the official list, so when Wizards revises it the score revises with it.

Honest limitation: that 5.5 step is a design choice, not a measurement. Two decks separated by a single card — one holding a catalogued two-card win, one not — differ by 1.34 points of final score even if they play identically in nine games out of ten. And Spellbook only knows the combos it has catalogued: an original two-card loop you found yourself scores zero until someone submits it. The cliff also puts a hard ceiling on honest casual decks: with no combos and no Game Changers, combo threat is 0 and the other four subscores can only carry you to 8.0, however perfect they are.

Two halves again, weighted evenly: land fit and color coverage.

landFit = 10 − |36 − lands| × 1.1

Exactly 36 lands scores a perfect 10. Every land away from 36 in either direction costs 1.1, so 40 lands and 32 lands both score 5.6, and by 27 or 45 the term is gone. This is the bluntest number in the whole model: it does not know about modal double-faced lands, Treasure production, or a 2.4-average-mana-value deck that genuinely wants 33.

Coverage is smarter. It takes each color's share of your colored pips as demand, then requires max(8, demand × 40) sources of that color, caps the ratio at 1, and averages across colors weighted by demand. Sources come from Scryfall's produced-mana data, so rocks and dorks count alongside lands, and anything that adds one mana of any color counts for all five. The consequences are concrete: a mono-black deck needs 40 black sources for full marks, while a color that makes up 20% of your pips needs only 8. The floor of 8 is a splash tax — one card with a single white pip still demands 8 white sources, and a deck that splashes on 4 sources will bleed coverage for it.

What to change: moving one land toward 36 is worth 0.0825 final. Fixing a starved color is usually worth more: if black is 60% of your pips and you run 18 black sources against a requirement of 24, you cover 75% of that color and lose 0.15 coverage — three quarters of a point of mana base. Swapping six colorless utility lands for Swamps buys all of it back, and buys it back at the table too.

Separately, the page reports hypergeometric mana math. Your tipping point is average mana value rounded and clamped to 2–6; the model looks at 7 + tipping point cards and computes the chance of seeing fewer lands than the tipping point (screw) and more than tipping point + 2 (flood). For 100 cards, 37 lands, and a 3.4 average: 10 cards seen, 20.7% screw, 10.8% flood, 68.5% sweet spot. Cut to 33 lands and screw jumps to 29.5%. Those percentages are diagnostics — they do not feed the mana base subscore, which only ever sees land count and color coverage.

A mono-black midrange list: 37 lands, 63 nonlands, average mana value 3.4, Sol Ring and Dark Ritual for fast mana, 9 nonland ramp pieces, 4 tutors (Demonic Tutor, Diabolic Tutor, Grim Tutor, Diabolic Intent), 8 detected draw effects, 12 interaction pieces of which four are edicts, 36 black sources, 2 Game Changers (Necropotence and Demonic Tutor), and one catalogued two-card combo.

• Speed: 10 − 1.6 × 2.6 + 2 × 0.9 + min(3, 2.25) = 9.89
• Consistency: min(5, 4.4) + min(5, 3.6) = 8.0
• Interaction: 12 × 0.75 = 9.0
• Combo threat: 5.5 + 1.2 + 2 × 0.35 = 7.4
• Mana base: 8.9 land fit × 0.5 + 90% coverage × 10 × 0.5 = 8.95

Weighted: 9.89(0.25) + 8.0(0.25) + 9.0(0.15) + 7.4(0.20) + 8.95(0.15) = 8.65, at Bracket 4.

Now move one card. Cut half of the combo and combo threat collapses from 7.4 to 0.7 — the deck loses 1.34 points, landing at 7.31, and falls to Bracket 3. Put the combo back and instead add two more removal spells and four more draw pieces: interaction hits its cap at 10, consistency climbs to 9.4, and the deck reaches 9.15 — a full half-point higher than the combo was worth, achieved entirely with cards nobody at the table will complain about. Notice also that Necropotence contributes 0.35 here as a Game Changer and, because of the draw-detection gap above, exactly nothing to consistency.

Ranked by final-score points per card slot, which is the only ranking that matters when you have one flex spot left:

• First two-card combo: +1.34, and one bracket
• Each of the first four tutors: +0.275
• Sol Ring specifically (fast mana + ramp): +0.29
• Each draw effect up to twelve: +0.1125
• Each interaction piece up to fourteen: +0.1125
• Each land moved toward 36: +0.0825
• Each Game Changer: +0.07, and the first one costs you a bracket

That last line is the honest summary of the whole system. Game Changers barely register in the score and dominate the bracket, because the two things measure genuinely different questions: the bracket asks what a deck is allowed to do, the score asks how often it gets to do it. If you are placing a deck for a pod, use the bracket. If you are comparing two of your own builds in the deck builder — same commander, different 99 — the score is the more sensitive instrument, and every coefficient it uses is on this page and on the methodology page.