Methodology
Every suggestion EDHTuner makes is deterministic: the same deck in gets the same answer out, and every claim traces back to a signal you can check. This page says where those signals come from and where they stop.
The signals, in the order they're trusted
Card text first. We read oracle text from Scryfall's card data for your whole deck and for every legal candidate. Themes, win routes, structural counts (ramp, card advantage, targeted disruption, mass disruption, protection) and commander synergy all come from what the cards actually say. That's why this works for a commander nobody else plays.
Popularity second, and always labeled. EDHREC's per-commander inclusion and synergy numbers are strong evidence about what tends to earn its slot, so cut suggestions cite them openly ("run in 4% of this commander's decks"). We weigh popularity as evidence, and it never gets the final say. A card the numbers have never met, because it's new or comes from an unusual deck, counts as unknown rather than bad.
Structure as a floor. Most functioning decks keep minimum counts of ramp, draw, and interaction. When your deck runs under a floor, filling that gap gets priority. When a theme really is the category (a landfall deck's ramp is its plan), we say so instead of telling you to trim it.
The simulator, honestly
Every suggested swap is playtested in paired solo games: the same shuffle seeds with and without the swap, typically a few thousand games per swap, so the measured difference is the card and never the luck. We test land counts the same way, at several counts. A recommendation has to clear a statistical bar and a practical one before we tell you to change anything.
The simulator holds one power on purpose: the veto. We rank suggestions by how well they fit your plan, and the playtest exists to catch the swap that makes your deck meaningfully slower. It doesn't do the ranking, because goldfish speed would push mana rocks over the synergy cards your deck is actually about.
Brackets and Game Changers
Suggestions respect the bracket you pick. We enforce the Game Changer limit. An over-cap deck gets the minimum weakest trims named, most replaceable first, with stand-ins. An under-cap deck gets told how much headroom is left. Card legality and the banned list come from the same Scryfall snapshot, and the deck panel shows its date.
How the quality is kept honest
Every change ships behind a battery of regression tests: several hundred fixture checks for past failure classes, a golden corpus of expectations pulled from expert review of real analyses, and replay gates over a fixed set of reference decks that must not drift. On top of that, independent expert reviewers judge batches of analyses in waves, and every confirmed miss becomes a fix for the whole class plus a new permanent test. After engine changes we re-run the full analysis of roughly 3,200 commanders to catch regressions at that scale.
Determinism and your data
No AI model runs at analysis time, and nothing is analyzed on a server. It all runs in your browser, and the same deck always produces the same result. A finished analysis sends our own feedback store the commander name, deck size, bracket, app version, the deck itself and the finished analysis, so we can see what we're actually recommending. Each page load also reports how long the app took to start. A deck we fail to read sends back the first few lines we couldn't read, whatever text those lines happen to hold, because that's the only way a broken paste gets fixed. We delete votes, reports and failed pastes after 90 days and start timings after 30, and a stored deck lasts a year. The FAQ spells this out in full.