Salt Scores: Reading the Table Before You Sit Down

Every deck you build or import gets a salt average on a 0–4 scale. The engine walks your commander, your partner if you have one, and every copy of every non-land card in the 99, assigns each card a number, and takes the plain arithmetic mean. That average is the headline; the analysis also returns the single highest-scoring card in the deck and a list of top offenders.

The labels are fixed cutoffs on that average: Very salty at 2.5 and up, Salty at 1.8, Moderate at 1.2, Mild at 0.8, and Low salt below that. The bar fills to average ÷ 4, so a deck that has earned "Very salty" fills roughly 63% of it. Nothing you can legally build fills the bar — that would require all ~63 of your non-land cards to be Armageddon-grade.

Lands are dropped before anything else happens. A card whose type line contains "land" is skipped entirely, which keeps 37 Islands from dragging a stax deck's average down toward the floor.

Honest bug, documented rather than hidden: because the land filter runs before the score lookup, two entries in our own salty-card data can never fire. The Tabernacle at Pendrell Vale (2.8) and Maze's End (2.5) are both lands, so they are discarded before anyone checks the table. If your Tabernacle deck reads as low salt, that is why, and your opponents will disagree.

Three tiers, checked in this order:

1. The curated table. 125 exact Scryfall names with hand-assigned EDHREC-style scores, ranging from Mystic Remora at 1.7 up to Armageddon and Ravages of War at 3.8. If a card is on this list, the list wins and nothing else is consulted.

2. Eight oracle-text patterns. If a card isn't listed, its rules text is tested against eight regexes and the highest match wins: destroy all lands 3.5, players can't cast spells 3.2, take an extra turn 3.0, each player sacrifices all 3.0, extra turn 2.8, skip … draw/untap/combat 2.8, search your library for a card 1.4, and counter target spell 1.2.

3. The floor. Everything else scores 0.4. Not zero — the assumption baked in is that any card played against you is mildly irritating on principle.

Those patterns are narrower than people assume, and the narrowness is load-bearing. destroy all lands matches Armageddon, whose entire text is that sentence. It does not match Jokulhaups, which reads "destroy all artifacts, creatures, and lands" — Jokulhaups only scores 3.5 because it's on the curated list. search your library for a card catches Demonic Tutor and Vampiric Tutor but misses Rampant Growth and Cultivate, which search for a basic land card. counter target spell catches Counterspell and Arcane Denial but misses Swan Song, which counters target enchantment, instant, or sorcery spell.

The patterns also produce a predictable false positive: skip … draw/untap/combat doesn't care whose step is being skipped. Solitary Confinement skips your own draw step as a maintenance cost and gets tagged 2.8 as stax. Necropotence has the same line, but it's on the curated list at 2.4, so the list overrides the heuristic and it lands where it should.

Mass land denial owns the top of the scale. Armageddon and Ravages of War at 3.8, Decree of Annihilation and Jokulhaups at 3.5, Catastrophe and Obliterate at 3.3, Sunder at 3.0, Fall of the Thran at 2.9, Wildfire at 2.6.

Resource locks sit just under it: Stasis 3.5, Winter Orb 3.3, Nether Void 3.3, Static Orb and Contamination 3.2, Smokestack and Braids, Cabal Minion 3.0, Back to Basics 2.9, Blood Moon 2.5.

Stax and taxes — the pieces that let the game continue but make it expensive: Grand Arbiter Augustin IV 3.4, Opposition Agent 3.3, Drannith Magistrate 3.0, Trinisphere 2.9, Sphere of Resistance 2.6, Rule of Law 2.5, Null Rod 2.3, Rest in Peace 2.0.

Extra turns are the tightest cluster: Time Warp 3.2, Nexus of Fate 3.0, Time Stretch 2.8, Temporal Manipulation and Alrund's Epiphany 2.6. Anything unlisted that says "take an extra turn" picks up 3.0 from the pattern layer, which is why an off-list card like Temporal Mastery still scores like the staples.

Card denial — discard, draw denial, hand attack — is represented only by named cards: Narset, Parter of Veils 3.1, Notion Thief 2.8, Tergrid, God of Fright 2.4. There is no text pattern for it. An unlisted symmetrical discard enchantment like Bottomless Pit or Necrogen Mists scores the 0.4 floor, the same as a two-mana bear.

Read the top of the table and the calibration becomes obvious: it is ranked by time, not by lethality. Armageddon at 3.8 outscores Thassa's Oracle at 3.0. The Oracle ends the game; Armageddon doesn't end anything, it just guarantees that the next forty minutes are worse for everyone including the person who cast it.

These 125 numbers are one developer's calibration of community sentiment, not a survey we ran and definitely not data from your pod. Mana Crypt sits at 2.9 and Jeweled Lotus at 3.7 because players resent losing to a turn-one commander, not because either card is unfun in the abstract. If your table plays proxied fast mana every week and nobody blinks, the number is simply wrong for you — and it has no way to know that.

The average is taken over every non-land card, and that denominator is brutal. Take a 100-card deck with 37 lands: 62 spells plus the commander leaves 63 cards to score. If nothing is listed and nothing matches, the average is exactly 0.40 — Low salt, floor.

Now load it up. Rhystic Study 1.8, Mystic Remora 1.7, Smothering Tithe 2.0, Cyclonic Rift 2.1, Seedborn Muse 2.6, Mana Crypt 2.9, Consecrated Sphinx 2.9, Time Warp 3.2. Eight of the most complained-about cards in the format, totalling 19.2. The other 55 cards contribute 22.0. The deck averages 41.2 ÷ 63 = 0.65 and the meter still says Low salt.

Working backwards, and assuming your offenders average about 2.8 each, here is what each label actually costs you out of those 63 slots: roughly 11 offenders to reach Mild, 21 to reach Moderate, 37 to reach Salty, and about 55 to reach Very salty. "Salty" therefore means more than half your non-land cards are famous offenders. That isn't a normal deck with a mean streak — it's a dedicated prison list.

The practical consequence: for any deck that isn't a stax deck, the label is nearly useless and the offender list is where the information lives. That list is sorted descending, keeps everything above 1.0, and is capped at eight. It also always keeps the top five entries regardless of score — so a genuinely clean deck will show you five cards sitting at 0.4 that have done nothing wrong. That's the panel refusing to render empty, not an accusation.

The salt average is computed inside the same analysis that produces your power level and it is displayed next to it, but it contributes zero weight to that 0–10 number. The score is speed 25%, consistency 25%, combo threat 20%, interaction 15%, mana base 15%. There is no salt term. The two numbers are neighbours, not relatives.

They come apart in both directions. Isochron Scepter and Dramatic Reversal aren't on the curated list and match none of the eight patterns, so each scores 0.4 — identical to a vanilla creature. Our bracket detection meanwhile checks your exact list against Commander Spellbook and finds a two-card combo, which places the deck at Bracket 4. Two cards that end the game on the spot, scored as politely as Grizzly Bears.

Devastation is the same story in reverse gear. It destroys all creatures and lands, so the bracket engine's mass-land-denial detector catches it and forces Bracket 4. But it isn't on the salt list, and destroy all lands needs those three words adjacent, so its salt contribution is 0.4. The bracket engine uses a much broader set of land-denial patterns than the salt engine does — separate detectors, written for separate jobs, and this is where you can see the seam.

Going the other way: a Persistent Petitioners deck runs thirty-odd copies of a card that isn't listed and matches nothing. Each copy is scored individually, each one at 0.4, so the deck reports an average of 0.40 — the absolute floor — while milling three opponents out over forty minutes. And a Grand Arbiter Augustin IV prison deck can grind the meter up toward Salty while its power score sits mid-pack, because zero detected combos means the combo-threat subscore contributes nothing to 20% of the total.

One clean overlap worth knowing: Winter Orb reads "players can't untap more than one land during their untap steps," which trips the bracket engine's land-denial detector and pushes the deck to Bracket 4 on its own, while also scoring 3.3 salt. Stasis, at 3.5, is saltier and doesn't trip that detector at all — it skips untap steps rather than touching lands. Same effect on the game, different rules text, different answer from each system. See methodology for how bracket detection works in full.

Because of the dilution math, the number worth acting on is the maximum and the offender list, not the average. Cutting Armageddon from that 63-card pool moves the average by (3.8 − 0.4) ÷ 63 ≈ 0.05 — invisible on the meter — while completely changing what happens at your table. Optimising the bar is the wrong game.

The question the scale is really asking is: how many turns of not playing Magic does this card create for the other three people? That framing explains why Rhystic Study sits at 1.8 and Winter Orb at 3.3. Rhystic Study taxes attention and is genuinely obnoxious, but everyone keeps taking turns. Winter Orb takes the turns away. Cards in the first category are fine in quantity; cards in the second are a decision about what kind of pod you want.

Three specifics that hold up in practice. Mass land denial is defensible when it's a win condition — Armageddon with a lethal board is a finisher; Armageddon as a mid-game control lever is thirty minutes of topdecking for everyone. Extra turns are fine in ones and twos and stop being fine at three, which is also where our own bracket rule fires: three or more extra-turn cards forces Bracket 4 for chaining potential. And stax pieces read very differently depending on whether your deck can actually close — a lock with no clock is the single most resented thing you can present.

If you want to see this in motion, the salt meter updates live in the deck builder as you add cards, and published decks carry their salt average on the browse card over at the public deck index, which is the fastest way to calibrate what a 1.2 or a 2.1 actually feels like across real lists rather than in the abstract.

Two more limits worth stating plainly. The salt engine reads only a card's front-face rules text, so a split or modal double-faced card whose miserable half lives on the back gets the 0.4 floor unless it's in the curated table — Boom // Bust is only scored at 2.7 because someone typed it in. And the whole thing is a forecast, not a measurement: it's built from what the format complains about in aggregate, and it has never met your playgroup. Treat it as weather, check the offender list, and then have the conversation before shuffling rather than after turn six.