How ranking and scoring work

Which is Saltier? shows two Magic cards that are legal in Commander and asks which one makes the table groan more. Your answers build the community salt ranking, and you score points for reading the community well.

A round

• 5 warm-up pairs first. They score no points: just give your honest take. The server picks them to teach the ranking the most, often a new card being placed against one near its current estimate.
• Then the scored pairs. Choose 15, 20, 25 or 30.
• Every round mixes the difficulty. The scored pairs range from a clear favourite (about 80% of the community on one side) to near coin flips (about 52%): the range is cut into as many equal slices as the round has pairs, and each slice gets one pair, in a random order. So every round has the same spread, and you aren't told how close a pair is until you've answered it.
• Change the length at any time. It applies to the pairs still to come (a pair already dealt keeps its own). A round where you changed it after scoring began isn't saved as a best score.
• Don't know the cards? "I don't know these cards" skips a pair (up to 3 per round). Skips never change salt scores.

Answers

Each pair is answered on a five-point scale:

The result works like a chess score, where a draw counts 0.5. "Much saltier" says there's a big gap, "saltier" a smaller one, and "even" no gap.

Points

Before your answer counts, the current salt scores give the community's odds that each card is the saltier one. Call the favourite's share P (between 50% and 100%). A pair's hardness is 1 − (2P − 1): 1 for a coin flip and 0 for a sure thing.

• Saltier: agreeing with the community earns 50 + 50 × hardness, so a correct call on a close pair is worth up to 100 and an obvious one about 50. Disagreeing costs 50 × (1 − hardness): almost nothing on a coin flip, and up to 50 on an obvious pair.
• Much saltier: the same, then +25 if you agree and the community sees a clear gap (a 70/30 split or wider), otherwise −25. Points are capped at 100.
• Even: 100 × hardness − 50: +50 on a coin flip, down to −50 on a sure thing.
• Warm-ups score nothing.

A round of N scored pairs scores between −75 × N and 100 × N. Some examples:

Your rank

• Every finished round is ranked against every finished round of the same length anyone has played (15, 20, 25 and 30 pairs are ranked separately). Rank 1 is the best, and equal scores share a rank.
• The server keeps the score. It adds up each answer's points as you go, so the ranked score is its own tally, not one sent from your browser.
• Not ranked: a round where you changed the length after scoring began (it isn't comparable, and isn't saved as your best either), and rounds whose answers were removed as manipulation.

Salt scores

• A card's salt score is an Elo rating, the system chess uses, built only from players' answers. Every answer is a head-to-head match between the two cards.
• How far scores move. Each card's score changes by K × (result − expected), the other card's the opposite way. Expected is what the two scores predicted before the vote, so a surprise moves scores more than an expected result. Calling a heavy favourite only "saltier" can even nudge it down, because the community already rated the gap bigger than that.
• Unsure cards move more. Every card has its own step K, set by how sure we are of its score. A card whose score is still a guess moves up to 64 points per unit of surprise; as it collects votes the step shrinks, down to 12 for a well-known card. Two cards in the same pair can move by different amounts.
• Starting values. The original 100 cards started between 1300 and 1700, from community salt-survey scores. Cards added later start from their own salt-survey score on the same scale when they have one, or the middle of the pool when they don't, and are placed by votes (see New cards). After that, only answers move a card's score.
• Records count recent votes. Win–loss–even records and the community split count the latest 10,000 votes. Older votes still count toward the score itself; a well-known card's small step lets the ranking keep up as opinions change.
• Abuse is removed exactly. If votes turn out to be manipulation, they are voided: their effect on every score and record is reversed exactly. (Later honest votes cast against the inflated scores keep what they did, and honest votes it pushed out of the window stay out, so reviewing quickly matters.)
• Community split. After you answer, a bar shows how recent answers on that exact pair split (left / even / right), once the pair has at least 3 votes.
• Agreement. The rankings page shows how often players' answers on each card agreed with the ranking at the time, by the same rule as points: picking the card rated higher, or calling a close pair even (the favourite at most 75%). It counts recent votes, like the records. Low agreement means players often see the card differently from where it's ranked. Tap a card to see it, with the up to 5 cards it has beaten and lost to most (counting only opponents it has a better, or worse, record against).

Which pairs you see

• Cards that need data come up more. The first card is weighted by how few recent votes it has. The second card favours pairs that have rarely or never been compared.
• Each scored pair's difficulty sets a target gap between the two cards' salt scores (see A round).
• Some randomness. One pairing in ten is random, so cards far apart in the ranking still meet.
• Left or right is a coin flip, so position tells you nothing.

New cards

A card new to the pool is being placed. It appears only in warm-ups and in the optional "Help place new cards" pairs after a round, against cards near its current estimate, which works like a binary search for its spot. One pair in ten matches it with a random ranked card instead, to tie it to the ladder. While its score is a guess it moves in big steps (up to 64). It joins the public ranking once it has had enough votes: 4 for a card seeded from its salt-survey score, 12 for one that started at the middle of the pool. Until then the rankings page lists it under "Being placed", and the cards closest to finishing come up most.

Where the cards come from

Every card in the pool is legal in Commander. The pool was built from:

• the current and yearly (2019–2024) lists of the saltiest cards from community salt surveys;
• the most powerful: Wizards' Commander Game Changers list;
• the most expensive Commander-legal cards (by Scryfall's paper prices);
• generally hated effects: extra turns, mass land denial and "hatebears" (Scryfall community tags);
• controversial cards: recent unbans and ban debates from the February 2026 Commander announcement (Biorhythm, Lutri, Worldfire, Sway of the Stars, Coalition Victory, Thassa's Oracle, Rhystic Study).

Fair play and privacy

• The server runs the round. It picks every pair, in order, and each one is fixed once dealt: the next pair only comes with your answer to this one (or one of your 3 skips). So nobody can shop for the cards they want to push, and each pair can be answered only once.
• Limits. Answers given within 300 ms of a pair being served are refused. Each player is limited to 40 answers a minute and 400 a day, and to 30 new rounds an hour and 120 a day. Votes sent from other websites are refused.
• Watching for manipulation. Unusual patterns, such as one player casting most of an hour's votes or a card's score jumping in a day, are flagged for review. Votes found to be manipulation are voided exactly.
• What's stored. Each answer records the pair, the answer, whether it was a warm-up, which side you favoured, the difficulty, timings (from the pair being served, and as reported by your browser), the round and position, and your network's provider number. A keyed hash of your IP address (of its /64 block for IPv6) is used for limits and to review and remove manipulation; the address itself is never stored. These hashes are kept for at most 180 days.
• Ranked rounds keep their length, score and finishing time. Like your answers, a round is linked to the hashed network it was played from only until that hash is deleted.