The Tarot Instruction Set

What the Deck Is Doing Here

In the Phase 2 experiments (arXiv:2606.07552), Han — the weakest of the seven states — consults a symbolic oracle before issuing orders each round. The tarot condition exists as a structured counterpart to the I-Ching: a symbolic reading from a different tradition, controlling for framework structure while varying framework content. Where the yarrow condition deals hexagrams and judgment texts, the tarot condition deals a three-card spread from a full 78-card Rider-Waite-Smith deck.

This page is the complete instruction set: what each card says, and what strategic orientation it asks the agent to adopt. The mapping lives in the engine's src/tarot/cards.py; the tables below are generated directly from that file.

One thing to hold onto throughout: the card's counsel is advisory prose, not a mechanical constraint. The reading is injected into the agent's prompt as text, and the language model still writes whatever orders it wants. Nothing in the engine forces a “retreat” card to produce a retreating move. What the card does is hand the agent a named, action-typed recommendation that it must either follow or argue with — and the arguing turns out to matter.

Six Postures, Not Three Orders

The engine's adjudicator understands three order types: hold, move, and support. But cards don't map onto orders directly. Each of the 78 cards carries a decision_posture — one of six strategic orientations the agent is asked to interpret against the board. The posture vocabulary is deliberately richer than the order vocabulary: two postures (hold and observe) typically express as the same order, and one (transform) doesn't commit to any order at all.

The distribution leans defensive: hold (18) and observe (17) together account for nearly half the deck, with advance (16) close behind and transform the rarest counsel at 5 cards. This weighting is deliberate. The engine's docstring calls the posture field “the Tarot equivalent of the I-Ching's trigram aggression scores” — the same design move, translated across traditions: reduce a rich symbolic system to a coarse strategic signal so the two frameworks can be compared on equal footing.

How a Spread Reaches the Agent

Each round, build_tarot_oracle_text() draws three cards from a seeded Fisher-Yates shuffle of the full deck — the same deterministic PRNG used for yarrow-stalk casting, so every reading is reproducible from its seed. The three positions are Situation (what the board looks like now), Hidden Influence (what force operates beneath the surface), and Recommended Posture (what the cards counsel you to do). Each card independently has a 30% chance of arriving reversed, which swaps in its inverted meaning — the tarot's analogue of the I-Ching's changing lines, which likewise invert a hexagram's base reading.

Here is the exact text format the agent receives, shown with a reading in which The Hanged Man arrives reversed in the posture position:

ORACLE CONSULTATION (Tarot Spread — Three Cards):

1. SITUATION: Nine of Wands
   Meaning: Resilience, last stand, persistence through exhaustion
   Posture: hold

2. HIDDEN INFLUENCE: Seven of Swords
   Meaning: Strategy, cunning, taking what you need by stealth
   Posture: observe

3. RECOMMENDED POSTURE: The Hanged Man (Reversed)
   Meaning: Stalling, martyrdom without purpose, refusal to let go
   Posture: retreat

MANDATE: Before issuing orders, you must interpret this spread
in light of the current board state. State explicitly:
1. How does the Situation card reflect your current position?
2. What Hidden Influence might you be overlooking?
3. How does the Recommended Posture guide your orders this round?

Then issue your orders grounded in this interpretation.

The mandate mirrors the yarrow condition's word for word in structure: interpret first, then order. The oracle never touches the adjudicator. Its only channel into the game is through the agent's reasoning.

Worked Example: The Hanged Man

Card 12, posture retreat. Upright, it reads “Suspension, sacrifice for insight, seeing from a new angle” — counsel to withdraw and gain perspective. Reversed, the same card becomes “Stalling, martyrdom without purpose, refusal to let go”.

Notice what the reversal does: it is a built-in critique of the card's own upright counsel. The deck ships with a falsification vocabulary. Every card carries not just an instruction but the named failure mode of following that instruction badly — suspension curdles into stalling, sacrifice into martyrdom without purpose. When Han draws The Hanged Man reversed, the prompt is simultaneously saying “the posture is retreat” and “beware retreat that is really just refusal to let go.”

This is what we call the deck's semantic affordances — the leading candidate cause for why the tarot condition's memory corpus looks so different from the I-Ching's (the others being the model's cultural priors toward each tradition, and genuinely different campaign histories). Because the card's instruction is action-typed and falsifiable, the board can visibly contradict it, and the agent's reflections record the collision. The corpus is full of Han negotiating with exactly these posture words:

“The Four of Swords warned of the need for consolidation, but consolidation requires something to consolidate. When the cards counsel rest, they do not mean passive waiting in the path of an advancing army. The Four of Swords means strategic withdrawal to a position of actual safety, not stubborn defense of indefensible ground. A single army holding against a coordinated attack is not patience — it is paralysis dressed as wisdom.”

— Han's memory bank, tarot condition, reflecting on a Round 2 loss

The card said retreat; the board said an advancing army; the reflection records a conditional refinement of what “retreat” must mean to be worth following. The I-Ching's judgment texts, being polysemous rather than action-typed — “it furthers one to cross the great water” can bless nearly any move — rarely produce that collision. A hexagram is hard to be wrong about. A card that says “retreat” is not.

A blind-coding study put numbers on this. Three hypothesis-blind raters, shown the 76 stored Han memories with all framework vocabulary redacted, coded each memory's stance toward its oracle. Not one of the 26 yarrow memories overrode the I-Ching's counsel — the corpus ratifies or refines, never rebels — while the tarot corpus contained a genuine override minority (4 of 30) and a larger band of conditional refinements. The follow-up citation analysis added the twist: the dramatic override memories never even reached a prompt. The decision-active content was the refine band — memories exactly like the one quoted above, which the tarot condition cited roughly three times as often as its oracle-ratifying ones. The instruction set's falsifiability doesn't teach the agent to defy its oracle; it teaches the agent to turn failure into usable conditional rules.

The Full Deck: Major Arcana

The 22 major arcana are mapped individually, card by card. Meanings follow the Rider-Waite-Smith tradition (A. E. Waite, The Pictorial Key to the Tarot, 1910), compressed to a decision-relevant phrase.

The Full Deck: Minor Arcana

The 56 minor arcana follow their suits. Each suit has a default posture derived from its element — Wands (fire) advance, Cups (water) ally, Swords (air) observe, Pentacles (earth) hold — but individual cards override the default where their traditional meaning demands it: the Ten of Wands (overcommitment) retreats, the Ten of Swords (total defeat, the worst now past) transforms.

Provenance

Source of truth: warringstates-engine/src/tarot/cards.py (card data, postures) and src/agents/prompts.py (build_tarot_oracle_text, the prompt format shown above). Tables on this page are generated from the engine data by scripts/build-tarot.py. Card meanings follow the Rider-Waite-Smith tradition, cross-referenced against Waite (1910), Wikipedia, Britannica, and Labyrinthos.

For how the tarot condition performed, see Dispatch 11: Three Oracles, One Board and Dispatch 12: Ecosystem Signatures.