Grounding and limits
What the model is given, which answers to trust, and where the system falls down.
Four parts of the app call a language model, and they are not equally reliable. This page says which is which. The pages under it describe each one in enough detail that you can predict where it will struggle.
Where to trust it
| You are asking about | How it does | Why |
|---|---|---|
| Deck strategy, themes, what a list is trying to do | Good | The whole deck is in context, and judgement is what the model is for |
| Whether a card exists and what it says | Good | Card text is read from the database, not recalled |
| What to add or cut, and why | Reasonable | Candidates are real and filtered. The ranking is a judgement you should check |
| Curve, color balance, counts by category | Reasonable | Arithmetic over a list it can see. Errors are visible to you |
| How two cards interact on the stack | Weak | See below. Verify anything that matters |
| Layers, replacement effects, timing corner cases | Weak | Same |
| Whether a specific card is the best in its slot | Weak | It sees at most 40 candidates out of thousands |
The short version: it is a competent deck-building conversation partner and a rules adviser you have to check. Treat a rules answer as a starting point and look for the citation.
What grounding does and does not buy you
Grounding is a claim about inputs. In this system it means:
Card text comes from the database. Every card in your deck arrives with its full oracle text, and the model has tools to look up any other card. That text is ours, imported from Scryfall, so an errata'd card reads as it currently reads.
Worth being precise about the limit here: the model is instructed to name only cards it has been given, and it is not prevented from doing otherwise. What you get is a way to tell. Card names in a reply are matched against the database afterwards, so a name with no hover card is a name that matched nothing.
Rules text is stored verbatim. 3,273 Comprehensive Rules entries, 77,999 official rulings and 11,665 worked examples of card interactions, all quotable with a citation.
The judgement is the model's. Which card suits your deck, what to cut, why a curve looks wrong. Those are opinions, and grounding says nothing about whether they are good ones.
So grounding makes one failure easy to catch, inventing a card. It doesn't fully solve the model misreading a card that it can see, and that is the failure you might still hit.
What it gets wrong, and why
Each of these is a property of how the system is built, so you can predict when it will bite.
It can misquote a card that is sitting in its context. Full oracle text for every card in your deck, mainboard, sideboard and maybeboard, goes into the prompt uncapped on every turn. This will cause performance to be better than a model using recall alone, but it can still falter.
It runs out of budget on hard rules questions. The reasoner gets a fixed number of tool rounds. When they run out it loses its tools and answers from what it gathered, and those answers are the least reliable the system produces.
The cause is not reluctance to look things up. It searches the rules heavily, often several times over for one question. Interactions needing a chain of lookups are the ones that run past the limit, so multi-step stack and timing questions are where errors cluster.
Rules search matches on wording. Rules lookup works on text, so a question phrased the way players talk ("if they sac it in response, does it fizzle") retrieves well only when those words appear in the rule itself. The model compensates by searching again with different wording, and that spends the budget above.
Rules change, and training data lags. Where the game's rules have been revised, an answer can describe the older version with full confidence, on questions where it did search first. Recalled rules read exactly like retrieved ones from the outside, so the citation tells you more than the tone does.
It sees a slice of the card pool. A search returns 20 cards. Recommendations show the model 40 candidates. A better card that ranked 41st was never considered, and the answer will not tell you that.
Some cards cannot be ranked. 116 of 38,344 cards have no embedding, usually because they were added since the last embedding run. They still appear in results, appended after everything that could be ranked.
The four surfaces
They differ mostly in how much grounding they get, which is the best predictor of how much to trust each one.
| Surface | What it does | Grounding |
|---|---|---|
| Chat | The deck sidebar and general chat | Six tools, deck context, rules lookup |
| Recommendations | Cards to add, with reasons | Candidates from the database, no tools |
| Smart search | Plain English into query syntax | None. The output is validated instead |
| Deck analysis | The nine analysis panels | None. It reasons from a deck summary |
How the AI sees your deck covers what is loaded from a deck and when.
How to check it
Every claim above is checkable, which is the point of publishing them.
- Hover a card it named. Names in a reply are resolved against the database. No hover card means no match, which is the signal to check that one.
- Check whether a rules answer cites a rule number. An answer with no citation is one where the model did not reach a verified rule, whether it ran out of budget or never looked.
- Re-run a recommendation as a search. The AI uses the same query language you have.
- Read the price timestamp. Prices are imported on their own schedule. See where a price comes from.
How we know
Each surface has an evaluation suite: questions paired with written correct answers, run against the live pipeline, and scored automatically. They run on demand, because each run costs API calls.
The suites are where the limits above come from. We treat their scores as a floor: they currently lean toward single-step questions and so under-represent the multi-step interactions this page warns you about. Widening them is ongoing.
Which model does what
| Stage | Model |
|---|---|
| Intent classification | Haiku 4.5 |
| Concept resolution (one of three paths) | Haiku 4.5 |
| Card and concept ranking | Voyage voyage-3, truncated to 256 dimensions |
| Query translation, reasoning, analysis | Sonnet 4.6 |
Classification and concept resolution are narrow problems with short inputs, so they run on the cheap model. Everything a user reads is written by Sonnet.