Does the answer depend on game length?
Each panel is one deck curve (the average mana cost of the non-land, non-ramp spells). The line is the best deck's land count if the game ends on that turn; the shaded band is every land count that gets within 2% of it. The pink strip is the “36-40 lands” advice. 20 and 48 lands are the ends of the tested range, so a line pinned there may want even fewer or more.
Table view
Explore every mana base
Each square is a group of decks with the same number of lands (going up) and the same amount of ramp (dorks plus rocks, going across), scored by its best split of dorks and rocks. Darker means closer to the best deck. The outlined area is everything within 2% of the best. Tap or hover a square for its numbers.
Is 36-40 lands necessary?
A setting is one curve, average cost, commander cost, draw scenario and game length. The advice is refuted in a setting when some deck with fewer than 36 or more than 40 lands scores within 2% of the best deck that has 36 to 40 lands. That is the test written down before any result (see Method).
One deck for every game: robust picks
You don't know how long the next game lasts or how much you will draw. The robust pick is the deck whose worst shortfall (its regret: how far behind the best deck it is, in %) is smallest across all draw scenarios and game lengths. It is the safe choice, not the best one in any single game.
Dorks or rocks?
Dorks cost 1 and tap the turn after. Rocks cost 2 and tap at once. For the best deck in each setting, how its ramp splits, and how much it would lose if all of that ramp were one kind.
Medians over every setting. “Loss if all rocks”: the best deck's lands and amount of ramp, but all of the ramp rocks, in % behind the best (when the grid has that deck). Same for all dorks.
Does commander cost matter?
How it works
- Exact, not sampled. Part 1 played thousands of random games per deck. Part 2 lists every reachable game state (cards in library, in hand and on the battlefield) with its exact probability, turn by turn. There is no sampling noise, so a gap between two decks is a fact of the model, not luck.
- The deck: 99 cards plus a commander. Untapped lands, 1-mana dorks (tap the turn after), 2-mana rocks (tap at once) and non-ramp spells. Lands 20 to 48, dorks 0 to 15, rocks 0 to 20, at most 25 ramp. One draw a turn from turn 1; scenarios P1-P8 add 0 to 7 extra cards on turns 5 to 8.
- Two spell curves. Tight: every spell costs one of the two whole numbers around the average (3.5 means half 3s, half 4s). Spread: a bell curve around the average folded onto three costs (average 3: some 2s, many 3s, some 4s).
- Two scores, ranked separately. Useful mana (V): mana spent on spells and the commander (not on ramp), with each turn counting 10% less than the turn before, because early mana matters more. Spells cast (N): how many spells (ramp excluded) and the commander you cast by the last turn. Where they disagree is itself a result.
- Game length. Every last turn from 3 to 15 is its own setting: “the game ends on turn T” means only turns 1 to T are scored.
- Near-optimal: within 2% of the best deck's score in the same setting. Every deck in that range is “as good as the best” here.
- Shared game graphs. The game tree under a fixed play rule does not depend on how many lands or dorks a deck has, only the draw odds do. So one graph is built per curve, commander and scenario, and every deck is scored by one pass over it. That is what makes millions of exact solves affordable.
- Checks passed before any result: land-drop odds match the hypergeometric tables to 1e-14; every game of part 1's own code reproduces to 1e-11; part 1's Monte Carlo averages agree within sampling error; probability adds to 1 and totals stay consistent. The spread curve uses a library shortcut measured at under 0.1% error.
- Pre-registered. The tests, the 2% margin, the grid and both scores were written down before the first run, with every later change logged: PREREG.md. Code and method: tools/sim/landmyth.
- Play rule: for now, part 1's fixed rule (keep 2-5 lands or 1 land and 2 ramp, a free first mulligan; cast ramp first through turn 6, then spend as much mana as possible). Phase 2 replaces it with the best possible play and mulligans, computed exactly.
What this model leaves out
- Play is a fixed rule, not the best play. A smarter player might want a different deck. Phase 2 computes optimal play and mulligans; this page will show both side by side.
- No opponents. No removal on dorks, no rock or land destruction, no board wipes. A disruption variant comes later and will always be reported separately.
- Only two curve shapes, and the spread one is folded onto three costs (the full bell curve is too large to compute exactly).
- Simple cards. No tapped lands, Sol Ring, colours, card draw that costs mana, or mana sinks yet.
- Coarse grid for now. Until the finer passes land, decks are tested in steps of 4 (or 2) cards, so a best deck or a range can still move by a step.