A single-card reading is simple math: one card, one topic, one written entry. Orim's decks together hold 78 (Tarot) + 90 (CRCT) + 72 (Tu Vi) cards, each with a hand-written, reviewed meaning per topic. Draw one, look it up, done.

Draw two or three cards in a spread, and the problem stops being simple. Position matters — the same card means something different as "where you stand" than as "what's opening up" — so this isn't combinations, it's permutations. Across three decks, two spread sizes, and the handful of topics (love, career, finance, self) that need their own written angle, that works out to roughly 6.15 million distinct card-position-topic combinations that would each need their own written synthesis to read as one coherent story instead of three separate statements.

Why that number rules out writing it by hand

Six million entries isn't a big writing project — it's an impossible one. At a conservative token estimate per entry, pre-writing synthesis for every possible combination would run into the billions of tokens, before a single one gets reviewed for quality. No editorial team, human or AI-assisted, produces and checks six million pieces of copy. The corpus approach that works perfectly for single cards — write it once, review it, ship it — cannot scale to combinations.

The actual math:

Deck	Cards	2-card perms (n × (n−1))	3-card perms (n × (n−1) × (n−2))
Tarot	78	6,006	456,456
CRCT	90	8,010	704,880
Tu Vi	72	5,112	357,840
Subtotal		19,128	1,519,176
× 4 topics (love, career, finance, self)		76,512	6,076,704

76,512 + 6,076,704 = 6,153,216 — roughly 6.15 million. Position matters (a card in slot 1 reads differently than the same card in slot 2), so this counts ordered permutations, not combinations, and this is only for 2- and 3-card spreads — a 5- or 10-card spread multiplies it further.

What actually happens without synthesis

Stack two or three independently-written, complete card readings next to each other with no bridging, and you get exactly what independently-written complete statements produce: sometimes they build on each other by luck, and sometimes they don't. Two cards can give opposing advice — one says move slowly, the next says don't hesitate — because neither was written knowing the other would ever sit beside it. That isn't a flaw unique to Orim's decks; it's what happens whenever a spread assembles pre-written, position-agnostic text without something to weave it into one reading.

Where the LLM's job starts — and where it stops

The fix isn't asking a model to invent a reading from a prompt. It's the same principle behind every other AI feature on Orim: the model never decides what a card means. That's already been written, reviewed, and locked down — the same corpus a single-card reading uses.

For a multi-card spread, the model's only job is synthesis: read the actual written meaning for each card that was drawn — its core meaning, its intro, and the specific topic's angle — and weave those already-true facts into one reading that acknowledges how the cards sit next to each other, instead of three disconnected verdicts. It can't introduce a star, a card, or a meaning that wasn't handed to it. It's doing the one thing a human reader does automatically when they lay out real cards on a table and read them as one spread: hold several true things at once and say what they add up to.

Why this is a different risk than a generic AI reading

An app that asks an LLM to compose your reading from scratch is gambling that the model's invented details happen to be right. This is the opposite: every fact going in was already written by a person and already true before the model ever saw it. The model's only freedom is in how it connects facts that are already fixed — not in deciding what they are. That's a narrow, checkable job, not an open-ended one.

Why single-card readings don't need this at all

A one-card draw already has a complete, reviewed answer sitting in the corpus — there's nothing to bridge. This only activates for 2-and-3-card spreads, where the problem it's solving actually exists.

How this compares to birth charts

Cards aren't the only place this scaling question comes up — Orim's Purple Star and Western natal charts face the same question. The actual numbers split in a way worth being precise about, not hand-waving.

Purple Star's chart-type space is smaller than the card deck's, not bigger. ~1,728 distinct structural configurations (12 months × 12 hour branches × 30 days × 5 element numbers × 2 genders = 43,200 raw inputs, collapsing to 12³ = 1,728 meaningfully distinct charts). Add the yearly horoscope layer — any age someone might ask about, roughly 100 realistic years — and that grows to about 172,800 possible full-reading contexts. Real, but still an order of magnitude smaller than the card deck's 6.15 million.

Western natal charts are the opposite case — they dwarf both. The interpretive unit isn't the whole chart, it's the aspect: the relationship between each pair of planets. With 10 tracked bodies (Sun through Pluto), that's 45 pairs, each capable of 5 aspect types or none — 6 states per pair: 6^45 ≈ 1.04 × 10^35. Even the coarser framing — which sign and house each of the 10 planets independently falls in, 144 states each — still comes out to 144^10 ≈ 3.83 × 10^21. Both dwarf the card deck's 6.15 million by twenty-plus orders of magnitude, because birth data is continuous — there's no fixed deck size to bound it.

System	What's counted	Approximate space
Card spreads (2-3 card)	Ordered card+topic permutations	~6.15 million
Purple Star (chart type)	Structural palace/star configurations	~1,728
Purple Star (+ horoscope)	Chart type × ~100 possible ages	~172,800
Western natal (aspects)	45 planet pairs × 6 aspect states	~1.04 × 10³⁵
Western natal (signs/houses)	10 planets × 144 states each	~3.83 × 10²¹

The lesson isn't "bigger number, more urgent LLM." It's that each domain needs its own math, not a shared assumption. Purple Star's chart-reading LLM call is justified by needing to weave 12 palaces into one narrative — not by an impossible-to-write combinatorial space, since 172,800 is large but not fundamentally unwritable the way millions are. Natal chart's case is the strongest of the three: genuinely uncountable, not just large, because the underlying birth data is continuous rather than drawn from a fixed deck.

Diagram list:

Hero: single card (written by hand) vs. 3-card spread (6M+ combinations)
Log-scale chart: single-card corpus size vs. multi-card combination count
Three card corpus entries feeding into one synthesis pass, one coherent output