How the bracket estimate is made and checked
SpellControl reads a Commander decklist and says which of the five Commander Brackets it looks like. This page says how, what the estimate was tested against, how many decks that was, and where it is wrong. It is a conversation starter for Rule 0, not a verdict: it sees the card list and nothing about how the deck is piloted.
To try it on your own list, paste it into the bracket checker. No account is needed.
Every figure below is reproduced from the repository by the benchmark in
packages/deck-metrics/src/calibration/ (run it with
npm run calibrate in packages/deck-metrics). The test decks are
frozen in calibration/fixtures/corpus.json, built 2026-09-30, so the numbers
only move when the estimator does.
How the estimate works
The code is packages/deck-metrics/src/index.ts. Two parts, in order.
- Hard floors. Rules that set a minimum bracket from the cards alone. One to three Game Changers is Bracket 3; four or more is Bracket 4. Mass land denial is Bracket 4. A two-card combo is Bracket 3 if it assembles late and Bracket 4 if Commander Spellbook rates it Ruthless, the deck has enough tutors and fast mana to assemble it early, it is highly redundant, or it uses your commander. Three or more extra-turn spells are Bracket 3. Enough stax pieces are Bracket 3 or 4. With no floor, the deck starts at Core.
- Power signal. A 0 to 100 score from fast mana, tutors, low average mana value, interaction density and combo engines that set no floor. Below a floor of 4, a score of 66 or more moves the deck up one bracket. At a floor of 4, a score of 80 or more with at least four Game Changers reads as Bracket 5.
The Game Changers list is a copy of the official one, dated 2026-02-09
(gameChangersDrift.ts); a scheduled check flags when the official list moves.
Combo data and ratings come from Commander Spellbook. Where the estimate is a close call
between two brackets, the deck page shows both sides and asks you which your table plays.
What it was tested against
The benchmark holds 265 labeled decks in three sets, plus a second opinion on 70 of the
precons. Labels are defined in calibration/benchmark.ts.
| Set | Decks | Label | Estimate matched |
|---|---|---|---|
| Official precons | 197 | The higher of Core, the Game Changer floor, and Commander Spellbook's bracket tag | 197 of 197 |
| cEDH top-8 lists | 48 | Bracket 5 | 47 of 48 |
| Constructed examples | 20 | The bracket the official text gives a precon plus real cards | 20 of 20 |
| ScrollVault second opinion | 70 (precons) | ScrollVault's per-precon bracket | 62 of 70 |
Read those rows with care
The first three rows are weaker evidence than they look. The precon label is built from the same kinds of facts the estimator reads (Game Changers and Spellbook's tag), so 197 of 197 shows the estimator applies its own rules consistently, not that it agrees with an independent judge. The 20 constructed examples are made to fit the rules, so 20 of 20 is by construction. The cEDH set holds only Bracket 5, so a guesser that always says 5 scores 48 of 48 there; the one miss is a list read as Bracket 4 because its power signal sat under 80.
The only outside check is ScrollVault, which publishes its own bracket for each precon using its own method. That is the number to weigh:
| Predictor | Exact | Within one bracket |
|---|---|---|
| SpellControl estimate | 62 of 70 (88.6%) | 68 of 70 (97.1%) |
| Always guess Core (Bracket 2) | 55 of 70 (78.6%) | 70 of 70 (100%) |
That is seven more exact matches than a guess that never looks at the cards, on 70 decks. The 95% bootstrap interval on 88.6% runs from 80.0% to 95.7%. It sits just above the baseline's 78.6%, but it comes from one outside source and one 70-deck sample, so read the gap as real but small. Where ScrollVault says Core, the estimate agrees on 52 of 55. Where ScrollVault says Bracket 3, it agrees on only 10 of 15 (66.7%; Wilson 95% interval 41.7% to 84.8%), and fifteen decks is too few to say more than that.
Where it disagrees
The eight ScrollVault misses, grouped by cause (printed by npm run calibrate):
- Three decks where we floor Bracket 4 because a combo runs through the commander and ScrollVault does not.
- Two decks where ScrollVault floors a two-card combo that Spellbook rates Core or Exhibition.
- One deck where we floor Bracket 4 on a land denial ruling and ScrollVault does not.
- One deck where we count a combo Spellbook rates S or P and ScrollVault discounts it.
- One deck ScrollVault lifts on its own optimization score, which we have no floor for.
What the estimate cannot see
- Intent. Bracket 1 is a deliberate theme-over-winning build that a card list cannot show, so the estimate never says Bracket 1. A deck with no signals reads as Core. You can state your own bracket, and the page shows it beside the estimate.
- How it plays. A precon that plays above Core with clean contents is labeled Core. The estimate has the same blind spot.
- Combos. They come from Commander Spellbook, so a combo it lacks or rates differently moves the answer. Combo disagreements are most of the misses above.
- Speed. The simulated turn-by-turn clock is shown on the deck page but does not change the bracket: weighting it at 0 to 30 points moved no deck's bracket (2026-10-25), so none ships. A modeled win path exists for only 52 of the 265 test decks.
-
Thresholds. The 66 and 80 cutoffs were set by hand on precon and
Archidekt samples (comments in
index.ts). The highest power signal among the 197 precons is 49, so the 66 cutoff has never been tested near a real precon. -
No table data. No deck was labeled by the people who played it. Owners
can say whether an estimate looks right; those answers are kept for human review and do
not tune the estimate (
backend/src/routes/bracket-votes.ts).
How deck generation is evaluated
A cheap mechanical check, frontend/scripts/recall-benchmark.mjs, hides a few
cards from each precon and cEDH list in the same test set, gives a ranker the commander and
the rest of the deck, and measures how often it brings the hidden cards back near the top of
the full legal pool. It compares the generator's pick scoring against baselines such as a
random shuffle and EDHREC play rate, with bootstrap intervals. It is not a pass-or-fail gate
and no result from it is claimed here. Getting a held-out card back is not the same as
building a better deck, so it is a first look, not a verdict.
Free, no account required.