TarotScript takes a position. The dominant strategy of this decade — bigger models, more tokens, emergent behavior — has a cost we've declined to pay. Hallucination is not a failure of scale. It is a failure of structure. Add structure back, and the failure mode changes.
The case rests on three claims. One: oracles need frames. Two: confidence is structural, not stochastic. Three: correspondence is older, and more robust, than probability. Each follows from the one before.
A language model is an oracle. Give it a question, receive a prediction. The prediction is drawn from a statistical surface — trained on a corpus, conditioned on a prompt. This is useful, and in many domains it is enough. But an oracle without a frame cannot be audited. You cannot ask it why — only what, and you cannot check the what against anything but itself.
A frame is a structure the oracle fills. A spread, a deck, a constraint — these are frames. The frame predates the answer. It says: this question has five positions, three guardrails, one action. When the oracle returns, you can inspect each position against the frame. The frame is the audit surface.
Every LLM emits a confidence, but the confidence is the same operation as the answer — a statistical trace. If the model is confident about a hallucination, it will tell you it is confident. The confidence does not index the world; it indexes the training distribution.
TarotScript emits confidence from dignity. Two positions agree elementally → confidence rises. Two positions oppose → confidence falls. The computation is pure, offline, inspectable; it depends only on the drawn cards and the element table. Zero LLM calls. When the agent says its confidence is low, it is because its structure flagged tension — not because its sampling did.
The tarot is a correspondence system. Four suits, four elements; major arcana mapped to planets and zodiac signs; numerology running through all of it. None of this is our invention — it is a Hermetic inheritance that predates the probability calculus by several centuries. We use it because it works: a correspondence table is a stable, shared, inspectable layer of meaning that everyone at the table can point at.
Probability is a lossy compression of correspondence. When you need to answer how likely, probability is irreplaceable. When you need to answer what kind, and why, correspondence is closer to the question. The stack we are building is a correspondence system with a probability layer bolted on, not the other way around.
Every card belongs to a correspondence table: element, numerology, planet, zodiac. Relationships between cards are first-class operations — sympathetic, opposed, harmonised, amplified. The runtime walks these correspondences instead of inferring over vectors. Meaning is inherited, not inferred.
Every cast emits a content-addressable receipt. Same source, same seed, same deck version: same hash. The receipt links to the exact cards drawn, the dignities computed, the facts emitted. When a decision is wrong, you can see which card carried it, and you can edit that card. No retraining, no prompt surgery. Just deck revision.
An agent is four deck files. A spread is a set of positions and a flow. A ritual is a sequence of statements. All three compose: rituals call rituals, spreads embed spreads, decks correspond with decks. The user calibrates by editing deck files, not by rewriting prompts. The calibration surface is structural.
Descend —