Accuracy
We publish our numbers. Every figure on this page is measured against real, community-labeled decks, not hand-picked showcases. Each block is dated, and the rating engine keeps moving between measurements, so read the date as part of the number.
of ratings land within one bracket of the community's label
Measured on 599 labeled Commander decks, as of 2026-09-14
None of them was in the model's training data: every one was first rated after 2026-08-05 and withheld from the fit, and decks rated before 2026-09-02 are left out too because a calibration pass had read their labels. That is the out-of-sample slice of the 5,709 labeled decks we hold.
The honest breakdown
When our rating disagrees with a deck's label, it almost always reads the deck as stronger: 31.2% of decks rate above their label versus 14.5% below. Before calling that an error rate, consider the direction. Commander players are famous for underselling their own decks - "it's just a 2" is practically a table ritual - and the official bracket definitions do not bend to modesty. A deck holding a two-card infinite combo, four or more Game Changers, mass land denial, or chained extra turns has a minimum bracket under WotC's own chart, whatever its owner calls it. We enforce those definitions; self-ratings often skip them. We measured it: of the 187 decks we rate above their label, 48.1% carry contents whose official minimum bracket already exceeds that label - those labels are not possible under the chart. We publish the miss anyway, because the remainder is real and we would rather measure it than hide it.
By label source
Not all labels are equally trustworthy, so here is every cohort, including the ones that make us look worse. Accuracy tracks label quality almost perfectly - which is what you should expect from a rater that is measuring something real.
Players on partner sites who saw our rating and confirmed it. Anchored, so read gently.
Every deck here is one someone disagreed with, by definition. Low exact-match is the point of the cohort.
How to read these numbers
The ground truth here is human judgment: in this slice, players on partner sites who saw our rating and either confirmed it or corrected it. Human labels are noisy in a particular direction - self-ratings skew low, and disagreement clicks over-represent people who felt over-rated. The same deck gets called a 2 by one table and a 3 by the next. Against that baseline, an exact-match score can never approach 100% for any rater, human or machine, and a rater that matched sandbagged labels perfectly would be reproducing the sandbagging. That is why the honest headline is the within-one number, with exact-match shown plainly above. Decks with a partner commander are rated on the primary commander in this corpus, so partner-pairing effects are not fully captured.
One more rule we hold ourselves to: the labeled corpus is also what the win-turn model learns from, so grading the whole corpus would grade the model on its own homework. Earlier versions of this page did that. Since 2026-09-14 the numbers above come only from decks the model never saw, withheld from its training set and rated after the last calibration pass, and they will be re-measured on a fresh slice whenever the model changes. Tournament cEDH lists we imported ourselves are left out of this slice: our cEDH staple data is drawn from the same events, so rating them would test recall rather than judgment.
By bracket
| Bracket | Decks | Exact | Within one |
|---|---|---|---|
| 1Exhibition | 9 | 0% | 66.7% |
| 2Core | 94 | 51.1% | 87.2% |
| 3Upgraded | 339 | 49.6% | 99.4% |
| 4Optimized | 107 | 69.2% | 95.3% |
| 5cEDH | 50 | 70% | 100% |
Bracket 1 (Exhibition) is rare by design - decks must show active evidence of being unfocused - so its sample is small and its labels are the noisiest in the corpus.
How the rating works
CommanderBracket does not count cards against a checklist. The engine models how fast a deck actually threatens to win, then maps that speed to the official WotC bracket definitions:
- Win-turn modeling. Ratings derive from an estimate of the turn a deck can credibly threaten a win, calibrated against thousands of real labeled decks. Two decks with the same staples can land in different brackets because they play at different speeds.
- WotC bracket rules enforced. Game Changer counts, mass land denial, chained extra turns, and two-card combos apply the official bracket restrictions as hard floors, exactly as WotC defines them.
- Prerequisite-aware combo detection. Combos are checked against a 40,000+ combo database including their setup requirements. A pairing that needs a five-creature board before it does anything is not treated like a turn-one win.
- Reasons on every rating. Every bracket ships with the factors that produced it, so you can see why - and push back when you disagree. Those disagreements feed the next calibration pass.
Ask the Judge
Our rules assistant is graded the same way: against the comprehensive rules themselves, not against vibes. On a 150-question difficulty-stratified eval (June 2026), it answered 96.5% of questions correctly with zero invented rules - every wrong answer cited a real rule and reasoned incorrectly about a hard interaction, and the assistant abstains with a "ask your local judge" warning rather than guess when it is unsure.
That eval predates our current rules model, and re-grading found a handful of questions where our own answer key was wrong, which the scoring counted in our favour. Treat the percentage as approximate. The zero-invented-rules result held across every question and is the part we stand behind. Unofficial rules guidance, not a tournament ruling. For tournament-level answers, consult a certified judge.
Or run any list through the free Commander bracket checker - pasted decklists and deck-site links both work.