The knowledge layer

The gap it fills

The AI reasons well over what a card says and poorly over what a card is for. It can tell you Cultivate finds two lands. It has no grounded basis for the judgement that ramp costing the same as your commander competes with the turn you want to cast them, so a deck with a five-drop commander wants its ramp at two and three rather than four.

That kind of statement is not in a card's oracle text and not in the Comprehensive Rules. It is accumulated practice: written by content creators, argued over in primers, and mostly agreed on. Today the model produces it from training or not at all, which is why its strategic advice is fluent, unsourced, and occasionally confidently out of date.

What is already built

The chat pipeline has a slot for this and uses it every turn it thinks is relevant:

• The classifier sets a knowledge flag for questions about deck-building theory and evaluation frameworks.
• Layer 3 embeds your message and asks for the closest passages, filtered by the concepts resolved for that turn.
• A prompt block formats what comes back under "Deck Building Principles", with each passage carrying its source file and section heading.

All of it runs. The lookup finds no table, returns an empty list, and the block is omitted. Adding content is what turns it on.

What has to exist

A corpus, chunked by heading so each passage stays a complete thought. An embedding for each chunk, in the same 256 dimensions the card embeddings use. Topic tags so retrieval can filter by the concepts a question is about rather than by similarity alone. And a re-embedding job, so editing a passage is an edit to a Markdown file rather than a database operation.

The repository already carries the beginnings: notes on commander deck building, deck styles, power ratings and card types. They are used today only as hand-written summaries inside prompts, not retrieved.

Attribution is the hard part

Card text and rules text have one obvious owner. Deck-building wisdom does not. A passage saying two-drop ramp beats four-drop ramp under a five-drop commander is somebody's argument, and repeating it without saying whose turns an opinion into an apparent fact.

So the shape this has to take is sourced passages rather than absorbed ones. Every chunk carries its source file and section, the prompt block passes both to the model, and the intent is that advice drawn from the corpus can say where it came from. A reader who disagrees with a principle should be able to see whose principle it is and discount it.

The alternative, folding community consensus into the system prompt as house opinion, is easier and worse. It makes contested judgement calls look like settled fact, and gives you no way to tell strategic opinion apart from the grounded card and rules text described on the limits page.

What it would not fix

The failures on the limits page are mostly about rules and budget. This layer touches neither. It would make strategic advice better sourced and more consistent. A wrong answer about the stack stays wrong.