Glossary

Moonborn-specific vocabulary lives here. General LLM terminology (prompt, embedding, token) is not redefined.

A

audit — The post-generation LLM-as-judge pass that scores a persona on coherence, depth, cultural fidelity, voice distinctiveness, and realism. Default model: Claude Opus 4.7. Min default score: 3.5/5.

B

Big Five — The OCEAN personality model (openness, conscientiousness, extraversion, agreeableness, neuroticism). Moonborn writes Big Five trait values into a persona's Self layer.

C

Cascade — One of four edit modes. Edit Soul; regenerate Self, Mask, Surface in order, each constrained by the new Soul.

D

DNA — A six-axis fingerprint derived from a persona's Big Five traits + Soul archetype intensity. Used for visual identity (the radar chart) and similarity queries.

drift — The cosine distance between a chat reply's embedding and the persona's voice fingerprint. High drift means the persona has slipped out of voice.

drift score — The per-reply numeric value of drift, in [0, 1].

distinctiveness — The cosine distance between a persona's voice fingerprint and a baseline (default: chatgpt-default). High distinctiveness means the persona reads as itself, not as generic.

E

ensemble — A graph of typed relationships between personas. Injected into multi-persona chat session prompts.

F

fingerprint — Short for "voice fingerprint" (see V).

fork — Create a new persona by copying an existing one + writing a parent edge into the lineage tree.

L

lineage — The directed graph of persona forks. Every persona has a root, an ancestry, and zero or more children.

Lock — One of four edit modes. Preserve specific layers while regenerating the rest.

M

Mask — The third layer of a persona. Voice, tone, signature phrases, social role.

memory — Session-scoped state (short-term context window + long-term pgvector retrieval + cold tier). Not cross-session by default.

moderation — The three-stage runtime safety pipeline: input intent screen, output content screen, impersonation + PII checks.

P

persona — A four-layer character document (Soul, Self, Mask, Surface) plus runtime metadata (fingerprint, DNA, audit verdict). The unit of work in Moonborn.

persona-as-code — Mirroring personas to a Git repository as YAML + markdown.

pipeline — The six-step generation flow: Intent parse → Soul draft → Self enrich → Mask build → Surface ground → Audit.

provocation tests — A 33-test suite that probes persona stability under role-breaking, emotional load, jailbreak attempts, etc.

R

Refine — One of four edit modes. Apply a named axis transform (warmer, darker, more-authoritative, etc.) to a layer.

RBAC — Role-based access control. Seven built-in roles; custom roles on Enterprise.

S

SAML — Security Assertion Markup Language; one of the SSO protocols Moonborn supports (Enterprise).

SCIM — System for Cross-domain Identity Management; the RFC 7644 provisioning protocol (Enterprise).

Self — The second layer of a persona. Big Five, archetype, values, attachment style.

Soul — The first layer of a persona. Core desire, fear, wound, growth arc.

Surface — The fourth (outermost) layer of a persona. Name, age, location, occupation, appearance.

V

voice fingerprint — A 50-scenario embedding baseline of a persona's Mask. Used to score drift on every chat reply.

W

workspace — A project container inside an org. RBAC, config, and personas are workspace-scoped.