Methodology

How BinderBrew builds a Commander deck

No black-box AI. Every card is placed by a traceable rule — and we publish the eval numbers so you can check our work. BinderBrew builds from a typed domain model, role-bucket templates, bundled card-cache candidates, EDHREC-style synergy priors, and the official WotC bracket system.

Try the builder Read the confidence report

Tier 0 legality pass rate

100%

Generated decks pass legality, color identity, singleton, banlist, and size checks before they are shown.

Commander prior coverage

117 / 1,998

43 curated snapshots plus 93 archetype-prior snapshots.

Corpus average Jaccard

25.5%

113 human reference decks from the 2026-05-14 corpus; median 24.8%.

Tier 1 floor compliance

94.7%

Share of corpus decks without role-floor LOW or UNREACHABLE diagnostics after generation.

Golden fixture pass rate

20 / 20

Golden decks passed hard invariants; average top-100 prior overlap was 20.7%.

Quality numbers you can check

Committed eval snapshot

These numbers come from data/eval/quality-snapshot.json, generated on 2026-05-31 by scripts/build-eval-snapshot.mjs. Regeneration runs the same offline eval harness used by tests; the page reads the committed JSON and does no request-time scoring.

Layered rules engine

BinderBrew validates in layers: Tier 0 legality, Tier 1 universal role floors, Tier 2 strategy/archetype role targets, then Tier 3 commander priors. The confidence score is capped when a lower tier is thin instead of pretending the list is stronger than the rules can prove.

Human-deck overlap

Against 113 human reference decks, average Jaccard overlap is 25.5% and average top-20 overlap is 7.7 cards. This is a directional quality signal, not a promise to clone any one public deck.

Best and weakest overlap

Reference corpus range

Best overlap · The Ur-Dragon

Jaccard 51.6% with 63/122 shared cards and 13/20 top-card overlap.

Source deck: archidekt-3062146

Weakest overlap · Atraxa, Praetors' Voice

Jaccard 4.6% with 7/152 shared cards and 1/20 top-card overlap.

Source deck: archidekt-3082019

Known limitations

Long-tail commanders without curated or archetype priors still rely on structural role floors and strategy fallback detection.
Jaccard overlap is intentionally conservative: human decks vary by budget, meta, pet cards, and collection history.
The reference corpus is a sampled offline benchmark, not a promise that every generated list matches a specific public deck.
The runtime generator remains deterministic and does not scrape EDHREC or call an LLM during deck generation.

Phase

Understand your cards

1. Collection parsing

Uploaded or pasted cards are resolved to Scryfall entries. Alternate printings, foils, and language variants are normalized so a card you own is recognized even when the suggested printing differs. Multi-format input is supported: ManaBox CSV, Moxfield export, Archidekt export, TCGplayer collection, MTGO, Arena, and plain decklist text.

2. Commander eligibility

Eligible commanders are filtered by the four format rules: legendary creature (or "can be your commander" rules text), color identity match, banlist status, and partner/background/doctor pairings. Partners and Backgrounds use the union of their color identities to broaden the deck.

3. Candidate discovery

Candidates come from three local-first sources merged at generation time: (a) the user's owned cards, scored by ownership and synergy, (b) a bundled Commander card cache refreshed out of band from Scryfall-style bulk data, and (c) curated staple/archetype seeds for role coverage. Live Scryfall calls are kept out of the hot generation path so deck building remains fast and predictable.

Phase

Build the deck

4. Role-bucket templates

Each deck variant — Aggressive, Midrange, Value — uses a different role-quantity template. Lands, ramp, card advantage, targeted removal, board wipes, and protection slots are filled in priority order. Basic lands are filled last using the deck color identity.

5. WotC Commander Brackets

BinderBrew implements the official 1–5 bracket system. Brackets 1 and 2 reject Game Changers, mass land destruction, repeatable extra-turn spells, and known two-card combos. Bracket 3 allows up to 3 Game Changers but no MLD, extra turns, or early combos. Brackets 4 and 5 are unrestricted. The Bracket diagnostics panel on the deck review explains exactly which cards drive the deck's bracket.

WotC Commander Brackets — magic.wizards.com/en/news/announcements/commander-brackets-beta

6. Synergy ranking

When a snapshot exists for the chosen commander, candidate cards are reweighted by EDHREC-style synergy priors. Cards in the commander's top-100 inclusion list get a multiplier between 0.5x and 2.0x. Unknown commanders fall back to a neutral score, so generation never fails for new releases.

7. Budget guardrails

Three budget modes: owned-only (the deck must come entirely from your collection), limited (a dollar cap on missing-card spend plus an optional single-card cap), and infinite (any card is fair game). Missing-card lists are scored by role impact and price so the most valuable additions surface first.

Phase

Prove it's good

8. Deck validation

Every generated deck runs through the validator: 100-card size, singleton rule, color identity, banlist, and — when a bracket is chosen — bracket violations. Invalid decks are returned as failed variants with structured error codes so the UI can guide the user (raise the budget, lift the missing-card cap, pick a different commander).

9. Deck confidence report

The Review screen scores the generated deck on legality, role balance, mana curve health, budget status, owned-card percentage, and bracket fit. The report highlights strengths, watch points, and next actions — built to give you confidence before spending money on the missing cards.

Common questions about the build

Is BinderBrew using AI to build decks?

Not for generation. Deck-building is fully deterministic: typed domain rules, role templates, and ranked candidates. Offline model-assisted archetype extraction may produce committed JSON snapshots, but runtime deck generation never calls a model or scrapes live deck data.

What is the WotC Commander Bracket system?

Brackets 1 through 5 describe how strong a deck plays at the table. Bracket 1 is ultra-casual, bracket 5 is cEDH. The bracket system controls which restricted-list cards (Game Changers, mass land destruction, extra turns, two-card combos) are allowed. BinderBrew lets you pick your bracket up front and reports the bracket your finished deck actually belongs at.

Where does card data come from?

All card data is sourced from Scryfall — names, oracle text, color identity, banlist legality, and prices. Card images are served from Scryfall's CDN. BinderBrew is independent and not affiliated with Wizards of the Coast or Scryfall.

How are missing-card buy lists ranked?

Missing cards are scored by role impact (ramp, removal, draw, lands) and unit price. The highest-impact, best-value cards land at the top so a partial purchase still meaningfully improves the deck. Marketplace links carry BinderBrew utm parameters; once affiliate programs approve us, those clicks contribute back to the project at no cost to you.

What should I read next?

Start a deck in the builder, compare the budget workflow, or read the confidence-report breakdown.