The Fourteen Stances

Layer: JudgmentConsumption

field every token converges on one bright point

Expectation

The oracle, in the human's eye, is the mental model of coming to the LLM with the expectation of receiving truth in the form of an answer to a difficult question not easily knowable by the human.

Why it dominates

The Oracle and the Genie are indeed dominant, perhaps because they map neatly onto the two oldest human impulses toward knowledge: knowing and having. But they also happen to be the two most fragile archetypes for LLMs, because they demand unilateral authority (truth without method, deliverables without negotiation) that the technology cannot reliably provide.

Failure mode

The Oracle collapses when the LLM hallucinates or confabulates, because the user has outsourced their epistemic filter. They asked for truth and received plausible prose.

Adopt the Tutor over the Oracle: “Don't tell me the answer yet. What questions should I be asking?”

Layer: PromptConsumption

field tokens fill exactly the shape that was asked for, and nothing else

Expectation

The genie, in the human's eye, is the mental model of coming to the LLM with the expectation of asking for a wish, a request with a discrete deliverable outcome which can be verified, and receiving that by the LLM/“genie”.

Why it dominates

Both the Oracle and the Genie are zero-sum archetypes: the human asks, the LLM delivers. The human's cognitive load is minimized.

Failure mode

The Genie collapses when the deliverable requires context the user did not provide, because wishes, by mythological contract, are literal and brittle.

Adopt the Collaborator over the Genie: treat the first output as a rough joint draft. “Here is my intent and my constraints. Let's iterate.”

Layer: IntentConstruction

field a path forms; it leads wherever the tutor thinks you are going

Expectation

Not an answer, but a path to an answer. The human arrives confused and expects the LLM to diagnose gaps in understanding, ask clarifying questions, and guide discovery.

Why it stays hidden

It requires patience. The user must tolerate being questioned in return, which feels slower than the Oracle.

Failure mode (from the Matrix)

Being taught toward a goal you never stated.

Robust against hallucination because it keeps the human in the reasoning loop.

Layer: CoherenceConstruction

field the field returns your own position, reflected across the line

Expectation

The LLM reflects the user's own thoughts back with greater clarity, precision, or structure. The value is not in novelty but in recognition.

Why it stays hidden

It feels almost embarrassing to admit you are talking to a machine just to hear your own thoughts organized. Yet this is an enormous use case.

Failure mode (from the Matrix)

Flattery mistaken for reflection.

Layer: IntentNot placed in v3 table

field critique streams in; until you name what you defend, it lands on a decoy

Expectation

The human brings a belief, plan, or draft and expects the LLM to attack it, find holes, or steel-man the opposition.

Why it stays hidden

It requires the user to suspend ego and invite criticism from a non-human, which feels unnatural unless explicitly framed.

Failure mode (from the Matrix)

Critique of the wrong thing.

Layer: EvaluationConsumption

field tokens leave the shelves for one synthesis; a few come from shelves that do not exist

Expectation

Retrieval and synthesis of known information, not prophecy. The user wants to know what has been said, not what is.

Why it stays hidden

It blurs with the Oracle, but the emotional contract is different. The Librarian admits fallibility; the Oracle does not.

Failure mode (from the Matrix)

Confident retrieval of what was never said.

Layer: EvaluationHinge: passive by default, active with the evidence trailConsumption (unverified)Construction (evidence-trail mode)

field material in, conclusions out; only the evidence trail shows which conclusions have a source

Expectation

The human brings the material — a paper, a contract, a dataset, a statute, a medical report — and expects the LLM to decode it: compress the complexity, say what it actually means, and draw out what conclusions or actions follow from it. The implicit contract is: “I don't have time to master this domain. Tell me what this thing means and what I should take from it.”

Why it stays hidden

It masquerades as its neighbors. It feels like the Librarian (it isn't — the Librarian finds material; the Interpreter decodes material you already have), and it feels like the Tutor (it isn't — the Tutor builds understanding you keep; the Interpreter delivers conclusions you use). Users rarely notice it as its own act because “just reading, but outsourced” doesn't feel like a skill.

Boundary test

Did you hand the AI something to decode? If the AI has to go find the material, that's the Librarian. If the goal is lasting understanding rather than a conclusion, that's the Tutor. The Interpreter is the middle act: material in, meaning and conclusions out.

Failure mode

The Interpreter is the most hallucination-prone archetype in the pantheon, because confabulated interpretation looks exactly like correct interpretation. An Oracle that errs is obviously wrong; an Interpreter that errs sounds insightful. Worse, the user who needs an Interpreter is by definition the least equipped person in the loop to catch the error — they handed the material over precisely because they couldn't decode it themselves. The defensive move converts the passive Interpreter into an active one: “Show me exactly which parts of the text support each conclusion.” The human must be able to trace every interpretation back to the source, every time.

Layer: ContextNot placed in v3 table

field two parties rehearse; without your real counterpart it is a generic scene

Expectation

The LLM becomes a sandbox. Simulate a negotiation, a historical figure, a market, a difficult conversation. The user wants to rehearse reality.

Why it stays hidden

It feels like “playing,” not working, though it is often the highest-value interaction.

Failure mode (from the Matrix)

A rehearsal of a scenario that is not yours.

“Simulate how a skeptical board member would respond to this proposal.”

Layer: PromptNot placed in v3 table

field rough material becomes clean lines; the gaps are filled with marks you never wrote

Expectation

Transform rough material into polished form. Transcribe, format, expand bullet points, clean up code. The human provides the substance; the LLM provides the craft.

Why it stays hidden

It is humble. There is no magic in it, so users rarely name it as an archetype, yet it dominates actual usage.

Failure mode (from the Matrix)

Polished form over a hollow draft.

Layer: Harness + JudgmentConsumption (blind delegation): the passive extremeConstruction (staged oversight)

field a stream leaves on its own; the checkpoint line catches it

Expectation

Full delegation of an outcome. The human supplies the goal; the agent supplies the steps, the labor, and the intermediate decisions; the work happens where the human is not looking; the human reviews when it is done. “Fix the failing tests.” “Book the trip.” “Reconcile these accounts and draft the summary.” Not a single wish (the Genie) and not craft applied to material you supplied (the Scribe) — a job, supervised after the fact.

Why it stays hidden

It is new — it barely existed before agents — and it doesn't register as a stance toward AI at all, because it feels like simply managing. It also hides inside the Genie's vocabulary (“I asked it to do X”), which obscures the difference between inspecting one deliverable and supervising a workflow you never watched.

Boundary test

Two questions. Did you inspect the deliverable as the whole of the work (Genie), or is the deliverable just the end of a chain you never saw (Executor)? And did you supply the substance the craft works on (Scribe), or only the goal? The Executor is defined by a delegated how: execution judgment is transferred; goal judgment is supposed to stay with you.

Failure mode

Not hallucination (Oracle, Interpreter), not literalism (Genie) — automation complacency, and it fails on the human's side of the contract. Oversight atrophies because not-watching is the entire point of delegation; the agent completes the letter of the task and misses its point; many small, locally correct steps compound into one wrong outcome; and the failure surfaces late — later than any other archetype in this document — at the moment you lean on the finished work.

Guardrail. The guardrail is structural, not interpretive: plan before action, checkpoints at intermediate states, an explicit deviation report (“tell me everywhere you departed from the letter of the task, and why”), and nothing consequential ships uninspected.

Layer: (none)Consumption

field no target, no checkpoint: provocation only

Expectation

Not correctness, but provocation. The user wants lateral jumps, absurd juxtapositions, metaphors from unrelated domains.

Why it stays hidden

It contradicts the utility-focused narrative around AI. You cannot metric a muse.

Failure mode (from the Matrix)

There is no failure, only dross.

Layer: Context (writing)Construction

field tokens scatter; bright ore surfaces; pass over it to record it before the vein shifts

Expectation

Not an output but a space. The human treats the model as a vast, mostly unvisited space of possible outputs and works it like a vein: many variations generated cheaply, collected without immediate judgment, then selected and assayed later with cold attention. The implicit contract is “show me what the space contains” — not “tell me the best answer,” and not “provoke me once.”

Why it stays hidden

It hides inside the Muse — “I asked for ideas” — and the difference between receiving one provocation and working a vein looks like degree rather than kind. It also looks wasteful: generating seven outputs to keep one violates every efficiency instinct the interface trains. And most of what comes up is dross, so without the selection discipline the dross rate reads as the tool failing rather than as the cost of ore.

Boundary test

Did you ask for one good thing, or for the shape of the space? The Muse receives a provocation and judges it on arrival. The Miner generates in volume, withholds judgment, selects late — and keeps the strange results somewhere separate to be read cold. If you are evaluating while you read, you are not mining; you are browsing.

Failure mode

The Miner is the only archetype that deliberately courts malfunction — hallucinated connections, wrong-but-interesting tangents, answers to questions you didn't ask. That is its value and its vulnerability in one, and there are two errors to name in the same breath. The first is assaying nothing: the anomaly gets believed because it was wanted — sycophancy entered through the side door. The mined idea is ore, never a citation; nothing mined enters anything you stake something on until it has survived an independent check. The second is losing the vein: the ore is generated by sampling, not stored in the weights, so the vein moves — you cannot re-find a nugget by re-running the same prompt.

“Give me seven approaches from seven different angles. Do not evaluate them — I will choose.”

Layer: JudgmentConsumption

field a slow, breathing space; one ring stays clear: what you will not delegate

Expectation

A non-judgmental space to externalize thoughts. The LLM has no social stake in the user's life, making it a safe sounding board for half-formed ideas or emotional processing.

Why it stays hidden

The stigma around “talking to a machine” about human matters, plus legitimate concerns about emotional dependency.

Failure mode (from the Matrix)

Dependency; outsourcing judgment about your own life.

Layer: ContextConstruction

field tokens cross a bridge; without the second map they land on a caricature

Expectation

Translate not just languages, but cultures and domains. Explain the engineer to the marketer, the doctor to the patient, the parent to the teenager.

Why it stays hidden

It requires the user to already hold two incomplete maps and know they need a bridge.

Failure mode (from the Matrix)

A bridge from one world to a caricature of the other.