How accurate are Commander bracket checkers?
Every bracket checker says it is accurate. Almost none of them say accurate against what.
That second part is the whole ballgame, and it is worth explaining why before I put our own numbers up.
There is no answer key
Brackets are not a measurable property of a deck like mana value. They are a judgment call, written down by WotC as a set of definitions, and then applied by humans who disagree with each other.
So when you build a corpus to test a bracket rater, you are not collecting ground truth. You are collecting other people's opinions. Most of those opinions are self-reported by the deck's owner, and deck owners are famously modest. "It's just a 2" is practically a table ritual.
This means an exact-match score measures two things at once. It measures whether the rater is any good, and it measures whether the labels are any good. A checker tested on 36 hand-picked decks with clean labels will post a beautiful number. That number tells you almost nothing about your deck.
Our numbers
Measured 2026-08-03, on 3,585 community-labeled Commander decks. Nothing excluded, nothing hand-picked.
- 96.6% land within one bracket of the label
- 56.7% are an exact match
- 40.0% are off by one
- 3.4% are off by two or more
If you only look at the exact-match number, 56.7% looks mediocre. Keep reading, because the split is the interesting part.
Accuracy tracks label quality almost perfectly
Here is the same corpus broken out by where the label came from, best sourced first.
| Label source | Decks | Exact | Within one |
|---|---|---|---|
| Tournament cEDH lists | 64 | 95.3% | 100% |
| Ratings users agreed with | 971 | 81.3% | 99.8% |
| Official preconstructed decks | 142 | 75.4% | 98.6% |
| Self-rated community decks | 1,111 | 57.7% | 94.4% |
| Disputed ratings | 1,197 | 30.3% | 95.6% |
On decks where the label is beyond dispute, because the deck showed up in competitive results, we hit 95.3% exact and every single rating lands within one bracket. On decks where somebody rated their own creation, we hit 57.7%.
The bottom row is there on purpose. Every deck in it is one where a person told us we got it wrong. Low exact-match in that cohort is the definition of the cohort, not a surprise.
The precon row needs a footnote of its own. All 142 of those decks are labeled bracket 2, because that is what a precon was assumed to be when the labels were collected. WotC has since decoupled precons from any single bracket, and our data says the same thing independently: 19 of the 142 ship contents whose official minimum bracket is 3 or higher, straight out of the box. Some precons are solid 3s. The label was a category, not a measurement, and a quarter of that row's "misses" are us reading the actual cards.
That gradient is the actual evidence. A rater measuring something real should get more accurate as the labels get more trustworthy. A rater that scored the same across all five rows would be telling you its corpus was doing the work.
The misses lean one direction
When we disagree with a label, we almost always read the deck as stronger. 33.4% of decks rate above their label. 10.0% rate below.
Before you call that a bias, look at what is in the decks. The bracket definitions have hard contents floors. A two-card infinite combo, four or more Game Changers, mass land denial, chained extra turns. Any of those sets a minimum bracket regardless of how the deck feels to play.
We checked. Of the 1,196 decks we rate above their label, 21.5% carry contents whose official minimum bracket already exceeds the label they were given. Those labels are not achievable under the chart. The deck was mislabeled before we ever saw it.
The remaining misses are real, and we count them as misses anyway.
Where we look worst
Bracket 1. 49 decks in the corpus, 14.3% exact, and when we miss we nearly always rate the deck higher than its label.
Read that one the same way as the precon row. Bracket 1 is not "a weak deck" or "a cheap deck". It means a deck with no real win condition and nothing holding it together, and that is genuinely hard to build on purpose. Most decks somebody files as a 1 turn out to be a functioning bracket 2 on a small budget. The bar is doing what it was designed to do.
The corpus is thin there too, 49 decks against 1,810 at bracket 3, so a handful of decks swings that percentage several points.
I am not going to pretend 14.3% is a good number. It is the weakest cell in the table and we publish it as the weakest cell in the table.
The number is supposed to move
The rating is not a checklist with a score attached. The engine estimates how fast a deck can realistically win and maps that turn onto a bracket, which is why two decks with the same tutor count can land in different places. That estimate is tuned against this corpus.
The corpus grows every time somebody tells us we got their deck wrong. There is a thumbs up and a thumbs down on every rating, and the corrections go into the labeled set that the next round is measured against. The disputed-ratings cohort in that table, all 1,197 of them, is that pipeline.
What I am not going to tell you is that this makes it more accurate every month. I do not have a clean before-and-after to show you, and the entire point of this post is not making accuracy claims you cannot show. What I will tell you is that the number is dated, it gets re-measured against the whole corpus rather than a frozen test set, and when it moves the wrong way that is what goes on the page.
What to ask any checker
Including this one:
- How many decks did you measure, and can I see the number?
- Where did the labels come from?
- What does accuracy look like split by label source?
- Do you publish the brackets you are worst at?
If a tool answers the first question with a number under a hundred, the accuracy claim is a vibe.
Our full breakdown, per bracket and per cohort, with the methodology and the parts that make us look worse, lives on the accuracy page. It updates when the engine does, and every figure on it is dated.
If you just want your deck rated, that is free and takes a decklist paste.
Card art is the property of Wizards of the Coast, images courtesy of Scryfall. CommanderBracket is unofficial Fan Content and is not endorsed by Wizards.