On V1 member/customer surfaces, the assistant presents as Lilith. Eve is the builder/operator identity for the same broader assistant program and must not leak into customer-facing labels, voice introductions, tour copy, history, or support documentation.
The journey preserves member agency: context and memory are visible, evidence travels with claims, and a proposed action does not become a write without its real authorization and effect boundary.
What the member experiences#
Lilith is available through the shared assistant panel and full-screen conversation surface. Depending on route, release, grants, and device, the experience can include:
- text conversation and durable thread history;
- page- and selection-aware questions;
- navigation, highlighting, reading, and scoped room tools;
- grounded answers with evidence and feedback;
- persona/room handoffs;
- curated and reviewed member-authored tours;
- speech input/output with visible listening and provider/fallback state;
- avatar, voice, memory, grounding, and safety disclosures;
- safe fallback when a provider, tool, source, or persona is unavailable.
The canonical behavior and honest shipped scope are in Assistant Experience.
Rooms and persona handoffs#
Lilith is the host identity; domain rooms provide specialized capabilities and voices. A handoff changes active persona/domain context while retaining the appropriate conversation and consent boundary. Release scope matters: the assistant must not offer a room or persona that is not enabled for the current release, tenant, locale, or member.
A handoff should tell the member what changed, which context carried, what new tools or limitations apply, and how to return. Safety, memory consent, and evidence rules continue across the handoff.
Tone and persona selection#
The V1 Lilith policy substrate provides the canonical roles, tone bands, capability caps, contemplative-tone rubric, teacher policy, crisis behavior, voice-abuse rules, spiritual-domain boundaries, cultural/lineage sensitivity, unsafe-claim handling, and tenant/version governance. See Lilith Persona Policy.
Tone is selected within policy, not improvised as unrestricted role-play. A member preference can choose among allowed expressions but cannot remove the gentleness floor, evidence/disclaimer requirement, crisis response, or scope cap.
Page context and actions#
The member shell can send route, active artifact, selected text, evidence state, and permitted tool grants. Server-side code validates resource access and authorizes each tool. The presence of content in the browser does not grant the assistant permission to persist, retrieve, or act on it.
Client actions should target declared commands/routes. Mutating room tools use their own domain authorization and confirmation contract. Lilith explains tool outcomes but does not convert a denied, partial, or failed receipt into success.
The context journey is Assistant handoff context carry.
Memory and disclosure#
Member memory is owned by Iris. The interface exposes the active memory posture, the source of recalled context where appropriate, and controls to edit, pause, forget, or opt out. Session context cannot silently become durable profile memory, and inferred information is not silently promoted to confirmed memory.
The contract is Iris Memory and Identity, with browser coverage in Memory edit, pause, and forget.
Grounding and evidence#
Factual answers can be composed through Sophia's governed source and claim contracts. Support state and citations are visible; optional model enhancement does not erase evidence provenance. When support is insufficient, Lilith labels the limit or declines the claim instead of presenting a confident unsupported answer.
See Sophia Grounding and the grounded-answer journey.
Voice and avatar identity#
Voice is an interaction mode under the same persona and safety policy. Listening, processing, speaking, interruption, permission-denied, and provider-degraded states must be visible. Synthetic voice identity and provenance must remain distinguishable from a human speaker; cloned or commissioned voices need their rights/consent record.
Avatar and voice presentation follow the release-bound persona pack contract in Persona, Avatar, and Voice Packs.
Crisis, refusal, and recovery#
Crisis-aware behavior runs before ordinary persona expression or tools and can supersede an active tour or task. Lilith avoids diagnosis and false assurances, presents appropriate resources, and makes the continuation/recovery state clear. Refusals should identify the practical boundary and offer a safe alternative when one exists.
The policy details are in Lilith Persona Policy, and the user journey is Crisis-aware tone policy.
Accessibility and device behavior#
The assistant must remain keyboard and screen-reader operable; manage focus across dock, drawer, modal, and tours; avoid relying on color for evidence or status; preserve readable streaming; respect reduced motion; and expose a usable mobile/PWA path. Voice cannot be the only way to access a capability.
Member-plane non-goals#
- Lilith is not branded Eve on customer surfaces.
- The assistant is not a replacement for licensed clinical or emergency care.
- A persona is not a claim of human identity or sentience.
- A grounding badge is not a universal truth guarantee.
- Memory is not an invisible permanent profile.
- A room handoff does not grant every tool in that domain.
- V3 embodiment does not weaken the V1 trust or consent boundary.