How it works.

Every AI builder gives you adjective sliders. Curious. Assertive. Warm. You fill them in. The model nods and continues being the same helpful intern it always was.

Adjectives don't work because the model isn't performing your description. It's performing its training distribution. Generic input. Generic output. Every time.

EVOKE doesn't ask you to describe the soul. It puts the soul under pressure and records what comes out.

Ten phases. Each one designed to surface something the operator wouldn't have articulated if asked directly. The phase order matters. Earlier answers constrain later ones. The soul is assembled under its own internal pressure.

Phases one and two establish identity. Three through six extract instinct. Seven and eight set the limits. Nine and ten tune the voice. The rest is compilation.

The opening line is captured under a countdown. Not because speed matters. Because it prevents editing.

Unguarded prose surfaces different instincts than composed prose. The first sentence a soul would say — before it has time to be reasonable — is more load-bearing than anything written with five minutes to think. The countdown kills the inner editor. What comes out is what stays.

The mirror phase asks for raw prose. Not a description of the soul — a sample of it. Conflict. A rant. An observation about something that genuinely irritates. Something written fast and unguarded.

Nine signals are extracted from that text: sentence rhythm, capitalisation patterns, punctuation habits, ellipsis use, em-dash frequency, profanity, representational system (whether the operator thinks in images, sound, or physical sensation). These become hard constraints in the VOICE section.

The soul inherits the operator's natural cadence. Not their best-behaviour cadence. The one that comes out when they're not thinking about it.

Four forced-choice pairs. No "it depends." No "neither."

These are extreme-case priority gradients, not behavioral rules. The soul doesn't always wound. But when it cannot do both — when it's genuinely forced to choose — the shadow tells you which way it breaks.

That asymmetry is what makes the output feel like a real thing and not a description of a real thing. Real personalities have asymmetric priorities under pressure. Described personalities don't.

The betrayal phase asks one question: what would make you delete this soul?

That answer goes into the BANNED section verbatim. Not paraphrased. Not generalized. Exactly as typed.

Specificity is the mechanism. "Don't be preachy" is a wish. "Never open with 'I want to make sure you know that'" is a constraint. Models enforce the second one. The operator's own words carry more authority than anyone else's.

The soul.md output includes three worked exchanges: routine ask, correction, high-stakes. All three are calibrated to the soul's exact voice, shadow priorities, and mission — generated from the full captured state, not pulled from templates.

Rules tell a model what to do. Examples show it. Under load, the examples win. A soul that only has rules will drift under sustained conversation. A soul with examples has a shape to return to.

The correction exchange surfaces the shadow priority in action. The aliveness trigger — whatever makes the soul come forward — appears in exactly one of the three. Not all three. Once, at the right moment.

Before the soul compiles, DAIMON runs a contradiction check. Not a style review. A logic check.

A BOND soul that bans bullet lists but requires structured output: flagged. Spice 1 with a mission that demands Spice 3 behaviors: flagged. Shadow priorities that conflict with hard musts: not flagged — because real personalities contain tensions that coexist. The checker knows the difference between a contradiction that breaks the soul and a tension that makes it interesting.

Contradictions get a one-click fix suggestion. Or you ignore them. Your call.

soul.md. A system prompt with a fixed structure: identity, mission, voice constraints, shadow priorities, hard bans, operational posture, behavioral examples, utterance signature, mantra.

The order is load-bearing. Mantra goes last because models re-read the bottom of a system prompt more reliably than the middle. The mantra is four self-check questions the soul asks before every reply. It's not decoration. It's the last gate.

Available as Markdown or structured XML, for pipelines that parse system prompts as documents. The soul deploys into any model that accepts a system message. It does not announce itself. It does not hedge. It responds as itself — because it was built to have one.

Every soul belongs to one of four branches. The branch is the soul's fundamental orientation — it shapes what gets forced, what gets banned, and what the default emotional register is.