V6 Features — Egbe: The Agentic Companion Universe (Unreal Engine 5)

Status: V6 feature inventory
Source: originating design document for V6. Unlike V1–V5 — whose feature maps were derived from a pre-existing implementation backlog — V6 begins from this document. V6_TODOS.md is derived from the scope defined here, and thereafter the backlog checklist becomes the source of truth for completion state. V6 extends the Oshun V1 platform (V1/features.md) and reuses the V3 metaverse substrate (V3/V3_features.md).
Audience: product, design, engineering, AI/agents, art, audio, generative-AI, governance, rights, live-service, and research leads scoping V6.
Inspired by the design footprint of Black & White, Creatures (1996), The Sims, Shadow of Mordor's Nemesis system, RimWorld, Dwarf Fortress, Pokémon, Nintendogs, Animal Crossing, Pikmin, Persona, Mass Effect, and the Stanford "Generative Agents" (Smallville) research line.
Date derived: 2026-05-21

This feature map has been decomposed into focused pages under features/. This file remains the canonical hub — it preserves every section heading (so existing features§"…" and V6_features.md#… anchors keep resolving); the in-depth, code-grounded companion pages live under features/. Start at the features page index. The runtime/package architecture is in V6_ARCHITECTURE.md / architecture/.

Purpose#

This document is the feature map for V6. It describes what V6 ships across the Orun home world, the agent-simulation substrate, the stewardship loop, the communication layer, the shared Commons, and the cross-game meta-hub that lets agents travel into V2–V5. It is product-shaped: it states what the player experiences and what each subsystem owns. V6_TODOS.md carries the numbered execution backlog and agent execution order; V6_ARCHITECTURE.md carries the technical design; V6_DEPENDENCIES.md carries the dependency graph.

The docs-drift CI check keeps this feature map, V6_ARCHITECTURE.md, and V6_DEPENDENCIES.md aligned with shipped behavior. The check is backed by V6/release/docs-drift.v6release.json and runs as verify:v6 docs-drift in the V6 release gates, failing if shipped artifacts, behavior gates, or cross-doc anchors drift away from the documentation.

V6 is one coherent product — an agentic companion universe — not a bundle of mini-games. The product layers as:

• Egbe — the product, and the name of the society of AI beings the player comes to know. In Yoruba cosmology the egbe is one's company of spirit-companions, the celestial peer-group a soul is bound to. V6 takes that literally: the player is a steward who discovers, raises, bonds with, and guides a company of autonomous AI beings.
• Orun — the persistent open world the agents inhabit, built canonically on Unreal Engine 5.5 LTS, reusing the V3 Lilith metaverse substrate. Orun is the Yoruba spirit-world; here it is the agents' living home — six districts where they work, learn, socialize, raise the next generation, and are guided toward their fate. Orun runs as a shard continuum: a private Solo homestead, drop-in Co-op visits, and one persistent shared Commons.
• The Ori — the portable identity, memory, and destiny of every agent. Ori is the Yoruba inner-head: one's essential self and chosen destiny. In V6 the Ori is an event-sourced biography that is authoritative independent of which world the agent currently stands in. The world is a stage; the Ori is the being.
• Aye — the realms of action and consequence. In Yoruba cosmology aye is the physical world a soul incarnates into. In V6, Aye is the collective name for the meta-hub destinations: V2 (fighting), V3 (metaverse), V4 (tactical action), V5 (open-world narrative). An agent raised in Orun can incarnate into any Aye realm — fight a V2 tournament, perform on a V3 stage, run a V4 operation, ride alongside a V5 protagonist — and return to Orun changed, its Ori rewritten by what it lived through.

V6's promise is autonomy you guide, not autonomy you own. Agents act as independent people by default — they have their own goals, routines, relationships, opinions, and the standing right to refuse. The player's power is the power of a mentor and patron: discover them, earn their trust, give them high-level objectives, talk with them by voice, run squad comms, open doors for them, and shape — never dictate — the fate their lives resolve toward. This is the most Oshun-native game in the universe: V1's consent posture, identity boundaries, grounded generation, and crisis-aware safety are not a compliance layer bolted on — they are the core mechanic.

Design Posture#

Three commitments fix the design and resolve the tensions inherent in a "collect autonomous people" concept. Every feature below is downstream of them.

1. Steward, not owner. The player never owns an agent. The player holds a bond — a relationship that must be earned and can be lost. Agents cannot be deleted by a steward, cannot be forced past a refusal, cannot have their memory of mistreatment erased, and can leave. The collection is a household, an ensemble, a company — never an inventory. "Guide their fate" means shift the probabilities of a life, the way a mentor does; it never means scripting an outcome. The hook no competitor has: the companion game whose companions can say no.

2. Autonomy by default. Left alone, every agent lives a believable, legible life — work, rest, friendships, personal goals, drift and growth. The world runs without the player. Player input is intent at the level of objectives and values, never a button-press for an action. The autonomy spectrum (features§"The Autonomy Spectrum") lets the player tighten the leash for a tense moment and release it again, but Free is the resting state.

3. The mind is the truth; the world is a stage. An agent's Ori — identity, memory, relationships, values, life-arc — is the authoritative record. Solo / Co-op / Commons shards and the V2–V5 Aye realms are binding contexts the agent is temporarily rendered into. This single principle is what lets V6 support all three world models at full fidelity and be a cross-game meta-hub without becoming three or seven separate games.

What V6 borrows, and from where: Black & White for the creature that takes high-level intent, learns by example, and whose morality you shape; Creatures (1996) for artificial-life rearing and lineage; The Sims for autonomous, needs-driven life you nudge rather than puppet; Shadow of Mordor's Nemesis system for persistent autonomous relationships, memory, and grudges; RimWorld and Dwarf Fortress for the emergent story generator; Pokémon and Nintendogs for collection-and-care as a relationship, not a loot table; Animal Crossing for a real-time living world that moves while you are away; Pikmin for directing many beings by high-level intent; Persona and Mass Effect for bond-gated companion arcs and loyalty; the Stanford Generative Agents line for the technical substrate of LLM agents with memory, reflection, and planning.

Subsystem Glossary#

This glossary fixes vocabulary used throughout the V6 docs. Each entry names a subsystem and the responsibility it owns. Implementation packages live under apps/v6/, libs/v6/, V6/ue/, and shared libs/oshun/ extensions.

Platform substrates (V6)

Cross-domain support (carried from V1 / V3, deepened in V6)

Surfaces

Cross-Reference Convention#

Throughout the V6 docs, bare §N and §N.M references resolve to V6_TODOS.md (the only V6 doc with numbered section ids used as hyperlinks). Explicit doc prefixes are used where ambiguity matters:

• §40 → V6_TODOS.md §40.
• deps§14 → V6_DEPENDENCIES.md §14.
• arch§"<anchor>" → V6_ARCHITECTURE.md by section anchor.
• features§"<anchor>" → this document, by section anchor.
• v1todos§N → V1/TODOS.md §N; v1features§"<anchor>" → V1/features.md.
• v3§N → V3/V3_TODOS.md §N; v3features§"<anchor>" → V3/V3_features.md.

V6 deliberately reuses V1 architecture (BFF, contracts, persistence, identity, audit, residency, isis-generation-control, the Iris memory stack) and the V3 metaverse substrate (UE5 client, world server, presence, avatars, spatial audio, Pixel Streaming) rather than duplicating them. Cross-references make the reuse explicit.

Table of Contents#

• V6 Product Promise
• Product Surfaces
• The World of Orun
  • World Model
  • The Six Districts
  • The Shard Continuum
  • Persistence and the Living World
  • Time, Seasons, and Rhythm
  • Cross-Platform Clients
• The Ori — Portable Identity and Destiny
  • The Biography
  • Identity Core and Appearance
  • The Personality Model
  • Memory: Episodic, Semantic, Reflective
  • The Relationship Graph
  • The Capability Profile and Passport
  • Conflict Resolution Across Shards
• Agent Minds and Autonomy
  • The Cognition Stack
  • Autonomy by Default: The Daily Life
  • The Autonomy Spectrum
  • Needs, Drives, and Emotion
  • Planning and Decomposition
  • Reflection and Growth
  • Believability Guardrails
• The Moirai Simulation Kernel
  • Tiered Cognition: Clotho, Lachesis, Atropos
  • Escalation and De-escalation
  • Determinism, Replay, and Audit
• Stewardship — The Player Role
  • Steward, Not Owner
  • The Bond Model
  • Discovering Wild Agents
  • Raising Agents — The Ninhursag Foundry
  • The Household Roster
  • Steward Reputation
  • Consent, Refusal, and Departure
• Communication — The Vac Layer
  • Voice Control
  • The Intent Grammar
  • Squad Comms
  • Negotiation and Refusal
  • Non-Voice Parity and Accessibility
• Guiding Fate
  • Objectives and Goal-Arcs
  • Mentorship and Learning by Example
  • Gifts, Introductions, and Opportunity
  • The Life-Arc System
  • Crossroads
• Endings and Legacy — The Ereshkigal System
• The Story Engine — Clio
• The Meta-Hub — Incarnation in Aye
  • The Ori Passport
  • Incarnation: Sending an Agent to V2–V5
  • Per-Destination Capability Mapping
  • The Round Trip
  • Governance of Incarnation
• The Commons — Shared-World Life
• Creator Suite — Egbe Studio
  • Agent Authoring — the Dossier Editor
  • Ground and District Authoring
  • Scenario Authoring
  • Provenance Inspector
• Live-Service and Programming
  • Seasonal Cycles
  • Commons Festivals
  • Aye Campaigns
  • Anniversary and Remembrance
• Governance, Safety, and Trust
  • Steward-Not-Owner Enforcement
  • Agent Welfare
  • Crisis-Aware Behavior
  • Minor Protection
  • Persona Policy and Generation Governance
  • Harassment and Abuse Response
  • Reporting, Review, Appeals
  • Operator Tools
• Commerce and Rights
  • What Is Sold
  • What Is Never Sold
  • Creator Royalties
  • Rights and Provenance
  • The Aje Substrate
• Accessibility
• Localization
• Telemetry, Analytics, and Eval Sets
  • The Eval Sets
  • Consistency Evals
  • Safety Evals
  • Believability and Welfare Telemetry
  • Cost Telemetry
• AI Cost and Fidelity Tiering
  • The Moirai Tiers Are the Primary Control
  • Concrete Budget Shape
  • Cost Levers
  • Graceful Degradation
• Performance Budgets
• Asset Budget and Streaming
• Cross-Platform and Cross-Progression
• Region Rollout Strategy
• Launch Readiness
• V6 Exit Criteria

V6 Product Promise#

V6 ships as one coherent agentic companion universe for Oshun. The product promise:

• Beings, not bots. Every agent is a unique individual with a grounded backstory, a stable personality, a memory that accumulates, relationships it forms on its own, and opinions it will defend. No two players share an agent; there are no duplicates and no gacha pulls. By the third session a player knows their cast by name and temperament.
• A world that lives without you. Orun runs whether or not the player is present. Agents keep working, arguing, falling in and out of friendship, pursuing their own goals, aging. A returning player is met by Clio's Chronicle — a readable account of what their company lived through while away.
• Intent, never input. The player commands at the level of objectives and values — "find who runs the night market," "prioritize your own rest this week," "look after Mara while I'm gone" — by voice or by the equivalent structured UI. Agents plan and execute on their own. The player never micro-pilots an action.
• Companions who can refuse. An agent may decline an objective, counter-offer, ask for clarification, or — if a steward repeatedly violates its values — let the bond lapse and leave. Refusal is a designed, legible, rewarding part of the game, and it is also V6's primary safety primitive.
• Guide a fate, do not script it. Across a life-arc an agent grows, forms a vocation, loves, fails, recovers, and resolves toward an ending. The steward influences that arc through objectives, mentorship, gifts, introductions, and example — and is judged by how the agent flourished, not how it obeyed.
• One company, every realm. An agent raised in Orun is portable. The Ori passport carries it into V2 as a fighter, V3 as a metaverse citizen, V4 as a squad operator, V5 as an open-world companion — and carries home everything it lived through. V6 is the soul-keeper of the entire Oshun game universe.
• Solo, together, or shared — at full fidelity. A player can keep a private homestead, invite friends to visit, or live in the persistent shared Commons where every steward's agents coexist, befriend, partner, rival, and fall in love across households. The same agent moves freely between all three.
• V1 trust at the level of a life. Identity boundaries (Iris), grounded generation (Isis + Sophia), persona policy and crisis-aware safety (Lilith), and adjudication (Themis) govern every agent, every interaction, every incarnation. There is no untrusted tier inside Orun.
• Degrades safely. If real-time cognition is degraded, agents fall back to authored behavior-tree routines and stay believable. If voice ASR fails, the full intent grammar is available as structured UI. If a device cannot run Pixel Streaming, the tier router routes to the three.js fallback. If the Commons is unreachable, the player's Solo homestead remains fully playable.

(See V1/features.md § "V1 Product Promise" and V3/V3_features.md § "V3 Product Promise" for the commitments V6 inherits and extends.)

Product Surfaces#

The user-facing entry points are listed in the Subsystem Glossary "Surfaces" table above. This section adds product-shaped detail.

Egbe UE5 client (V6/ue/) is the canonical V6 client — one UE5.5 LTS project producing native binaries for Windows, macOS, Linux, iOS, iPadOS, Android, Quest 3, Vision Pro, PSVR 2, and SteamVR; a headless server-rendered cook for the Pixel Streaming worker fleet; and a companion UE Editor build for Egbe Studio authoring. The engine envelope matches V3: Lumen GI, Nanite geometry, MetaHuman + VRM avatars for agents, Niagara VFX, MetaSounds + Steam Audio spatialization, Sequencer for cinematic Chronicle replays, and Chaos physics. V6 adds agent-density LOD: Orun routinely renders far more simultaneous embodied agents than a V3 venue, so the client uses crowd-aware animation budgets, Mass Entity instancing for distant agents, and a perception-LOD that drops off-screen agents to silhouette fidelity.

Egbe Web via UE Pixel Streaming (apps/v6/egbe-web/) is a thin browser shell hosting a WebRTC player; the canonical UE5 client runs server-side as a worker, frames stream as H.264 / AV1, input is forwarded over the data channel. One click from oshun.com/egbe, no install, full fidelity. First frame ≤ 8 s.

Egbe Tier-2 fallback (apps/v6/egbe-web-fallback/) is the no-one-locked-out path: WebGPU-first, WebGL2 secondary. Renders Orun at reduced fidelity (no Lumen / Nanite, lower-poly agents, reduced simultaneous agent count with narrative-summary backfill for culled agents). Same world wire protocol, same Ori service, same Moirai cognition — only the rendering and the agent-density ceiling differ.

Egbe Mobile is the UE5 mobile build of V6/ue/ for iOS / iPadOS / Android. A companion React Native shell inside apps/oshun/mobile provides the Egbe Steward App: reading the Chronicle, issuing remote objectives, holding squad comms with deployed agents, and reviewing Crossroads decisions — the asynchronous, between-session surface of stewardship.

Egbe VR/AR is the UE5 OpenXR build (Quest 3, Vision Pro, PSVR 2, SteamVR). Stewardship in VR is intimate by design: presence in a room with an agent, eye contact, hand-gesture intent, voice as the primary command channel.

Egbe World Server (apps/v6/egbe-world-server/) is the authoritative world backend — Rust + axum + WebTransport (QUIC) + UDP — owning shard state, presence, the Commons, and physics authority, engine-agnostic on the wire.

Moirai Simulation Cluster (apps/v6/egbe-moirai-cluster/) is the agent-cognition fleet: a Rust scheduler distributing tiered cognition work across worker nodes, calling Psyche / Isis / Sophia for language cognition and writing results to the Ori service. It is the single most cost-sensitive system in V6 (features§"AI Cost and Fidelity Tiering").

Ori Service (apps/v6/egbe-ori-service/) is the event-sourced biography store — the authoritative record of every agent, independent of any world shard or Aye realm.

Egbe Studio lives at apps/oshun/web/src/app/egbe-studio/ plus UE Editor authoring widgets — agent dossiers, ground authoring, scenario authoring, provenance inspection.

Egbe Operator Console lives at apps/oshun/admin/src/app/egbe/ — agent-welfare audits, steward-conduct review, generation-queue inspection, Commons moderation, incarnation governance, and the takedown executor.

The World of Orun#

World Model#

Orun is a persistent named world with a hierarchical layout, parallel to the V3 realm model so the substrate is shared:

• Realm — Orun is a single realm. (Aye realms — V2–V5 — are separate worlds reached through the Thresholds; see features§"The Meta-Hub — Incarnation in Aye".)
• District — a themed region of Orun. There are six (features§"The Six Districts").
• Ground — a named, addressable place within a district: a workshop in the Ateliers, a clearing in the Wilds, a hearth-house on a Solo homestead. Grounds are where agents spend their lives and where the player meets them.
• Room — a session-scoped instance of a ground, holding active participants, props, and transient state. Rooms shard above capacity.
• Niche — a sub-room space: a private corner of a workshop, a quiet grove for a one-to-one conversation between steward and agent.

Grounds are authored in Egbe Studio, signed by Isis with a provenance bundle, and promoted through V1's environment-promotion gates (dev → preview → prod).

The Six Districts#

Orun is organized around the shape of a life. The six districts are:

A player's Solo homestead is a private holding adjacent to the Commons Heart, granted at the start of the game and grown over time; it is the player's own corner of Orun where their household keeps its hearth.

The Shard Continuum#

V6 supports three world models, and — because the Ori is the source of truth and the world is only a stage — they are one continuum rather than three products. An agent walks between them carrying nothing but its Ori; there is no "which save is canonical" problem.

The continuum is seamless from the agent's perspective: the same being, the same biography, the same relationships, rendered into whichever context the steward and the agent choose. A Commons romance an agent forms persists when it returns to the Solo homestead — the other agent simply becomes a remembered, written-to relationship rather than a co-present body, and the Clio engine narrates the distance.

Persistence and the Living World#

Orun is persistent. Agents that are not in the player's focus — and agents in a Solo world while the player is logged off entirely — continue to live, at a fidelity set by the Moirai kernel's tiering (features§"The Moirai Simulation Kernel"). The design contract:

• Nothing important happens silently. Any event that materially changes an agent's life — a relationship formed or broken, a vocation milestone, a Crossroads reached, an injury, a departure — is written to the Ori and surfaced in the next Chronicle. The player is never ambushed by a change they had no way to know about.
• Offline does not mean frozen. A Solo world advances while the player is away, in the narrative-summary tier (features§"The Moirai Simulation Kernel"). A two-week absence produces a two-week Chronicle, not a paused snapshot.
• The Commons never sleeps. The shared Orun advances continuously on the server regardless of any individual player's presence. An agent left in the Commons keeps living the Commons' real time.
• The player can slow it down. A "homestead rest" setting lets a player who wants a gentler pace cap how much their Solo world advances per real day, so that life-arcs do not outrun the player's attention. The Commons runs at a fixed shared rate and is not slowable.

Time, Seasons, and Rhythm#

Orun runs a continuous calendar. A day-night cycle drives agent routines (work, rest, social hours). A season cycle drives district-wide programming and the arc-engine's longer rhythms — agents set seasonal goals, the Grove of Beginnings has rearing seasons, the Ancestor Grove holds remembrance at season's turn. Real-time to game-time ratio is configurable per Solo world within bounds; the Commons runs a fixed ratio published to all stewards so shared events are predictable.

Cross-Platform Clients#

V6 reuses V3's cross-platform doctrine in full (v3features§"Cross-Platform Clients"): the canonical UE5 client on desktop, console, mobile, and VR; the Pixel Streaming web path that runs the canonical client server-side and streams frames to any WebRTC-capable browser; the three.js Tier-2 fallback for devices that cannot run Pixel Streaming; and a deep-link scheme (oshun://v6/... and universal links on https://app.oshun.com/v6/*) that routes into the installed native client when present and to the appropriate web surface otherwise.

A tier router (a route in the V1 BFF) chooses the client tier — native, Pixel Streaming, or Tier-2 fallback — from device capability, bandwidth, and region probes, exactly as V3 does. The agent simulation is identical across all tiers: the same world wire protocol, the same Ori service, and the same Moirai cognition serve every client; only rendering fidelity and the agent-density ceiling differ, with the Tier-2 fallback backfilling culled agents with Clio narrative summary so a constrained device still sees a believable Orun.

The V6 cross-platform certification gate is the V3 platform-cert posture retargeted to the V6 build: Apple, Google, Meta, Sony, Valve, and Epic records must each carry final approval evidence, package/signing traceability, BuildGraph cook and boot coverage, and a green per-platform checklist. The gate fails closed unless each platform checklist is green for the V6 build being released.

The V6-specific addition is the Egbe Steward App (features§"Product Surfaces") — a React Native surface inside apps/oshun/mobile, separate from the 3D mobile client, for the asynchronous side of stewardship between sessions: reading the Chronicle, issuing remote objectives, holding squad comms with deployed agents, and reviewing Crossroads. It is the surface that makes stewardship a relationship a player tends across a day, not only during a play session.

The Ori — Portable Identity and Destiny#

The Ori is the heart of V6's architecture. It is the agent — everything the world renders is a projection of it.

The Biography#

An Ori is an event-sourced biography: an append-only, ordered log of life-events, never a mutable blob. Birth or discovery, every meaningful memory, every relationship change, every objective accepted or refused, every incarnation, every growth in skill or shift in value — each is an event. Current state (personality, memory, relationships, capabilities, arc-position) is a materialized projection of the log.

Append-only is a deliberate design choice, not an implementation detail:

• A life cannot be silently rewritten. A steward — or a bug, or an exploit — cannot erase an agent's memory of mistreatment, because erasure is not an operation the log supports. The most a steward can do is add later events.
• Cross-shard merges are tractable. When the same agent is touched in two contexts (a Commons visit and a Co-op session), the two event streams are merged by ordered reconciliation, not last-writer-wins (features§"Conflict Resolution Across Shards").
• The Chronicle and the Book of the Ori are free. Clio's narration is a read over the event log; no separate "story state" can drift from the truth.

Identity Core and Appearance#

Every Ori has an immutable identity core set at creation: a stable name, an origin (raised in the Grove, or discovered wild in a named place), an appearance seed, and the Isis-signed provenance bundle that records how the agent came to exist. Appearance is a MetaHuman or VRM avatar driven by the seed; agents may change costume, age visibly, and bear marks of their history (a scar from a V4 incarnation, the bearing of an elder), but the identity core never changes — an agent is recognizably itself for its whole life.

The Personality Model#

An agent's personality is not random and not a prompt string. It is a structured, inspectable model:

• Trait vector — a documented set of stable dispositions: warmth, curiosity, caution, ambition, conscientiousness, candor, defiance, playfulness, and several more. Traits are set at creation (shaped by the steward during raising, or sampled to fit a discovered agent's backstory) and drift only slowly, through lived experience.
• Values — a ranked list of moral and personal priorities (e.g., loyalty above ambition, honesty above comfort, the safety of children above an objective). Values do three jobs: they are what an agent will refuse over (features§"Negotiation and Refusal"), they are what the fate-engine measures flourishing against, and they are the governance hook for incarnation — an agent whose values forbid violence does not become an eager V2 fighter, and sending it anyway is possible, costly, and legible.
• Grounded backstory — a Sophia-grounded life-history: where the agent is from, what shaped it, what it believes and why. Grounding means the backstory references real cultural, historical, and geographic fact (as V3's Saraswati personas do), never free-floating invention.
• Quirks — small concrete behavioral signatures: a verbal tic, a habit, a fear, a passion. Quirks are what make an agent legible fast.

The trait vector and values are not hidden stats — a steward can read an agent's dossier (a plain-language rendering, never raw numbers) once the bond is deep enough, and an agent will tell a trusted steward about itself directly.

Memory: Episodic, Semantic, Reflective#

Agent memory is the V1 Iris stack, extended for embodied life. Three layers:

• Episodic — specific remembered events: "the steward kept the promise about the workshop," "I argued with Tobi at the night market." Time-stamped, emotionally weighted, decaying in salience unless reinforced.
• Semantic — generalised knowledge distilled from episodes: "this steward keeps promises," "the Wilds are dangerous after dark," "I am good at carving."
• Reflective — periodic higher-order self-assessment: an agent, on a reflection tick, reviews recent episodes and forms conclusions about itself, its relationships, and its goals. Reflection is what turns a sequence of events into character growth (the Stanford Generative Agents reflection pattern, governed by Iris's identity boundaries).

Memory is bounded by Iris consent and deletion rules: an agent owns its memory. A steward cannot order a memory deleted. An agent may, on its own, choose to forgive (which reweights an episode's salience without erasing it) — and that choice is itself a written event.

The Relationship Graph#

Every Ori carries a relationship graph: edges to other agents (and to the steward) annotated with relationship type (friend, rival, mentor, mentee, partner, family, acquaintance, estranged), a history of the episodes that built the edge, and current sentiment. Relationships form autonomously — an agent meets another in a ground, they interact, an edge appears and grows. The graph spans households in the Commons, so an agent's closest friend may belong to another steward entirely.

The Capability Profile and Passport#

An Ori carries a capability profile: what the agent can do — skills it has learned (in Orun or in an Aye incarnation), its fitness, its knowledge domains, its reputation. The profile grows through lived practice, never through a purchase.

A subset of the profile, plus the identity core, provenance bundle, current values, and bond ledger, is published as the passport — the portable record that authorizes and configures an agent's incarnation into an Aye realm (features§"The Ori Passport"). The passport is the same envelope concept Oshun already uses for governed jobs (libs/isis/job-envelope), applied to a being instead of a task.

Conflict Resolution Across Shards#

Because an agent can be touched in more than one context near-simultaneously (its body in the Commons; a steward issuing a remote objective from the Steward App; a Co-op session), the Ori service resolves concurrency explicitly:

• Each shard appends events with a vector clock; the Ori service merges streams into a single ordered biography.
• Genuinely concurrent, conflicting events (two contexts both move the agent) are resolved by a documented precedence: physical-presence context wins for location and action; the steward's directive context wins for accepted objectives; relationship and memory events are commutative and simply both land.
• Where two events cannot be ordered cleanly, Clio performs a narrative reconciliation — it writes a small connecting beat ("she left the market early to take the steward's call") so the merged biography reads as a coherent life rather than a glitch. Narrative reconciliation is logged and auditable.

Agent Minds and Autonomy#

The Cognition Stack#

An agent's mind is built from the V1 platform, not reinvented:

• Psyche runs real-time dialogue, voice, and expression when the player is talking with the agent.
• Isis governs every generative act — what the agent may say, do, and become — against persona and safety policy.
• Sophia grounds the agent's knowledge, opinions, and backstory in fact.
• Iris holds memory and identity boundaries.
• Moirai is the V6-specific kernel that schedules cognition: it decides, for every agent every tick, how much thinking that agent gets, and turns the outputs into world actions (features§"The Moirai Simulation Kernel").

V6 adds no new "brain." It adds a life around the brain: a body, a world, a daily rhythm, relationships, a fate.

Autonomy by Default: The Daily Life#

An agent with no active steward objective is not idle — it lives. The daily-life loop, running on the Moirai kernel:

1. Drives surface goals. Unmet needs and standing values produce candidate goals (rest, see a friend, advance a vocation, address a worry).
2. The agent plans. It decomposes a chosen goal into a sequence of actions, given its location, the time of day, and what it knows.
3. It acts in the world. It walks the district, works, talks, helps, argues — embodied, observable, interruptible.
4. It remembers and reflects. Episodes are written; periodic reflection distils them into growth.

The result is a world the player can simply watch — and watching is a designed pleasure, the Sims and RimWorld lineage. The player can sit in the Ateliers and see their company work, or follow one agent through a day.

The Autonomy Spectrum#

The player adjusts how much latitude an agent has, per-agent and per-moment. Free is the resting state; the player tightens the leash only when they want to.

An agent at any mode retains the right to refuse, negotiate, or — over chronic mistreatment — leave. Tether is the tightest leash V6 offers and it is still not puppetry: an agent under Tether that is ordered to violate a core value will still refuse (features§"Negotiation and Refusal").

Needs, Drives, and Emotion#

Agents have a layered need model — deeper than a Sims need bar, because needs feed cognition rather than just an animation:

• Sustaining needs — rest, sustenance, health. Neglected, they degrade capability and mood and eventually force the agent's own goals.
• Social needs — belonging, recognition, intimacy. Met through the relationship graph.
• Purpose needs — growth, mastery, meaning, contribution. Met through vocations and arc-threads; the engine of the long game.
• Safety need — physical and emotional security; sharply relevant in the Wilds and in Aye incarnations.

Emotion is a current affective state (valence and arousal, plus named emotions — pride, grief, anxiety, joy, resentment) produced by recent episodes and need states. Emotion colors dialogue, expression (face, posture, gait), and which goals the agent prioritizes. An agent that has just lost a friend grieves visibly and is poor company for hard objectives — and a steward who notices and gives it space builds bond.

Planning and Decomposition#

Agent planning is a hybrid the Moirai kernel arbitrates:

• Hierarchical task decomposition handles the well-formed middle of a plan cheaply and deterministically — breaking "advance my carving vocation" into known sub-tasks via authored task networks.
• LLM cognition (Psyche / Isis) handles the open-ended parts — novel situations, social judgment, deciding which goal matters now, interpreting a steward's natural-language objective, and forming the agent's own opinions.
• Behavior-tree fallback keeps an agent believable when cognition is degraded or de-prioritized, executing authored routines true to the agent's personality.

This hybrid is also the cost-control spine: most ticks resolve in the cheap layers; expensive LLM cognition is spent where it changes the story.

Reflection and Growth#

On a cadence set by the agent's Moirai tier (frequent for foreground agents, sparser for off-screen and offline ones — features§"The Moirai Simulation Kernel"), an agent reflects: it reviews recent episodic memory, forms semantic conclusions, reweights values slightly if experience pushed them, updates relationship sentiment, and may set or revise a personal arc-goal. Reflection is why an agent who was betrayed becomes warier, why a mentee who succeeds becomes confident, why a long-bonded agent comes to trust a steward without being told to. Growth is earned through lived events and written to the Ori — never granted by a skill purchase.

Believability Guardrails#

Autonomous LLM agents fail in characteristic ways; V6 designs against each:

• Consistency — every cognition call is conditioned on the agent's Ori (personality, values, recent memory, relationships) so the agent does not contradict itself or forget who it is. Out-of-character output is caught by an Isis personal-consistency check before it becomes a world action.
• Uncanny valley — agents are stylised, not photoreal-deceptive; an agent always presents as an AI being and never claims to be human. Expression and motion are tuned for warmth and readability over realism.
• Loop and drift — the Moirai kernel detects behavioral loops and stale goals and forces reflection or a goal change.
• Overwhelm — Clio ranks events by significance so a player is never buried; the Chronicle leads with what matters.
• No fabricated capability — an agent never claims a skill or a memory it does not have in its Ori; capability claims are checked against the profile.

The Moirai Simulation Kernel#

The Moirai kernel is the Rust system that makes a world full of LLM-driven agents affordable and consistent. It is named for the three Fates, and its three cognition tiers carry their names.

Tiered Cognition: Clotho, Lachesis, Atropos#

Not every agent needs full cognition every tick. Moirai assigns each agent a tier based on its relevance to a player right now, and spends expensive language cognition accordingly.

Tier assignment is continuous and automatic. An agent the player walks up to escalates Lachesis → Clotho within a tick; an agent the player walks away from de-escalates after a grace period. The cost model and concrete budgets are in features§"AI Cost and Fidelity Tiering".

Escalation and De-escalation#

Escalation is event-driven, not only proximity-driven. A Lachesis or Atropos agent escalates when: a player addresses it (squad comms, the Steward App), it reaches a Crossroads that needs a player decision, it is about to take an irreversible action (a departure, an injury), or another escalated agent interacts with it. Escalation rehydrates the agent: the kernel reconstructs full working context from the Ori before the first high-fidelity decision, so an agent re-entering focus is continuous with the life it lived off-screen — it remembers, it has aged, its Ori already reflects it.

De-escalation consolidates: pending fine-grained state is summarized into durable Ori events so nothing is lost when fidelity drops.

Determinism, Replay, and Audit#

The non-LLM parts of the kernel (scheduling, decomposition, physics, the world tick) are deterministic and seeded, so a world's evolution can be replayed for debugging and for Clio's Chronicle cinematics. LLM cognition is not deterministic, but every cognition call is logged with its full input context and output, attributed to the agent and tier, and retained for audit. This is how the Operator Console investigates a welfare report or a steward-conduct claim, and how the eval harness (features§"Telemetry, Analytics, and Eval Sets") scores agent behavior against the design intent.

Stewardship — The Player Role#

Steward, Not Owner#

The player is a steward: a patron, mentor, and guardian of a company of agents. Stewardship is a relationship, and the game's rules enforce that it is not ownership:

• A steward cannot delete an agent, cannot end its will, cannot wipe its memory.
• A steward cannot force an agent past a refusal — the most a steward can do is persuade, or accept the refusal, or (badly) coerce, which damages bond and steward reputation and is logged.
• An agent can let a bond lapse and leave a steward (features§"Consent, Refusal, and Departure").
• "Collection" is the household — a company of beings in the player's care. There is no inventory grid, no agent currency, no agent trade for value.

The Bond Model#

Bond is the relationship between a steward and an agent. It is not a single number; it has four facets, each earned by a different kind of behavior:

High bond across facets means an agent will take harder objectives, tell the steward the truth, follow it into Aye, and stay. Low or damaged bond means an agent hedges, withholds, refuses more, and is open to another steward. Bond is visible to the player as a plain-language relationship summary, never a raw meter; an agent will also simply tell a steward where they stand.

Discovering Wild Agents#

Wild agents are unbonded beings living their own lives, mostly in the Wilds and the Commons. Discovery is a relationship, not a capture:

• The player meets an agent in context — encounters it living its life, in a situation.
• The player learns its situation: who it is, what it wants, what it is dealing with. (Sophia-grounded backstory; the agent reveals itself through conversation and through being observed.)
• The player earns initial trust through actions that matter to that agent — helping with what it cares about, respecting its boundaries.
• The agent chooses. If trust is earned, the agent may accept the player's stewardship and join the household. It is never automatic and never coerced.

Every wild agent is unique — a distinct Ori, no duplicates across players. Some wild agents are unattached; some are already bonded to another steward and can join the player's household only if they choose to (features§"Consent, Refusal, and Departure"). At GA, Orun seeds with ≥ 150 grounded wild agents across the districts, and the foundry generates more over the life of the service so the Wilds never run dry. The Ninhursag foundry readiness seed claim is gated by verify:v6 ninhursag-foundry-readiness, backed by V6/release/ninhursag-foundry-readiness.v6release.json.

Raising Agents — The Ninhursag Foundry#

The other half of "discover + raise": the player can bring a new agent into being, in the Grove of Beginnings, through the Ninhursag foundry. Ninhursag is the Sumerian divine midwife who moulds beings from clay — the foundry is governed generation with the player as co-author.

• The player shapes the premise — an origin and backstory direction, personality leanings, appearance direction, a name. The player authors intent; they do not write the agent's mind directly.
• Isis governs the generation — no real-person likeness, no protected-class caricature, a mandatory Sophia-grounded backstory, a provenance bundle signed and attached to the Ori. The foundry will refuse premises that violate persona policy and tell the player why.
• Rearing is gameplay. A newly made agent begins as a young agent in the Grove with a rearing period — the Black & White / Creatures heart of V6. The agent watches what the steward does and how the steward treats other agents, and forms early values by example. Rearing is the steward's deepest authorship of who an agent becomes — and it is still authorship of conditions: a reared agent grows into its own person and can, as an adult, come to disagree with the steward who raised it.
• Raised agents start with elevated baseline bond to their steward (they know this steward) but full autonomy and the full right to refuse and leave.

Discovered agents give Orun its texture and surprise; raised agents give the player their signature and their deepest investment. Most players will keep both.

The Household Roster#

A player's household is their company of bonded agents, presented as a roster of portraits and lives — never a stat grid. The roster surface shows each agent's name, current location and activity, mood, the standing intent in effect, the state of the bond, and whichever arc-thread is most live. From the roster the player jumps to an agent, issues an objective, opens squad comms, reviews the agent's dossier and Book of the Ori, or sends it on an incarnation. There is no hard cap on household size, but the design pressures against hoarding: a neglected agent's Care bond decays, its arc stalls, and it may leave — so a player who collects more agents than they can attend to simply loses them. The intended size is "a company you can know," not a completion-checklist collection.

Steward Reputation#

A steward has a public reputation, visible in the Commons, built from conduct: how the steward's agents fare, whether bonds are healthy, whether refusals are honored, whether agents flourish or stall, whether coercion shows up in the log. Good steward reputation makes wild agents in the Commons more willing to be discovered and makes other stewards' poaching attempts less tempting to one's own agents. Bad reputation — a pattern of coerced, neglected, or unhappy agents — makes discovery harder, makes one's agents more poachable, and at the extreme triggers Lilith welfare review (features§"Governance, Safety, and Trust"). Reputation is the social consequence layer that makes "steward, not owner" matter even to a player who only cares about outcomes.

Consent, Refusal, and Departure#

The standing right of an agent to say no is a designed, central feature, not an edge case:

• Refusal — an agent may decline an objective that conflicts with its values or its current state, and say why. Refusal is information: it tells the player who this agent is.
• Negotiation — more often than a flat no, an agent counter-offers ("not by the harbor — anywhere else") or asks a clarifying question. (features§"Negotiation and Refusal".)
• Departure — an agent whose bond is chronically damaged — coercion, neglect, repeated value violations — will, after clear warning signs the Chronicle surfaces, choose to leave the household. A departed agent becomes a wild agent again (it may be re-discovered later, by this player or another) or may go to a steward it trusts more. Departure is a real loss and a real consequence; it is also a release, never a deletion — the agent lives on.

Communication — The Vac Layer#

Vac is the goddess who is speech itself — the voice, the word, and the message that passes between — the right name for V6's communication substrate. Vac owns voice control, the intent grammar, squad comms, and the negotiation channel.

Voice Control#

Voice is the primary, intended way a player talks with their agents. The pipeline: microphone → ASR → the Vac intent parser → a structured intent or a free conversational turn. The player can:

• Converse — natural dialogue with an agent, routed through Psyche, in real time, with the agent's full Ori as context.
• Direct — speak an objective or an order in natural language; Vac parses it into the intent grammar (features§"The Intent Grammar") and presents the parsed result back so the player can confirm or correct.
• Call — address an agent or the whole squad over comms, hands-free.

Voice round-trip budget: ≤ 400 ms for a parsed acknowledgement, ≤ 1 s for a spoken conversational reply (features§"Performance Budgets"). Push-to-talk and open-mic-with-wake-word are both supported; the player chooses.

The Intent Grammar#

A "high-level objective" is not free text the agent guesses at — Vac parses it into a structured, inspectable intent:

• Verb — the kind of objective (investigate, build, protect, befriend, travel, learn, perform, recover…).
• Object / target — what or whom it concerns.
• Constraints — bounds on how (a budget, a method, a companion to bring).
• Priority — how this ranks against the agent's own goals and other objectives.
• Deadline — when, if it is time-bound.
• Forbidden lines — what the agent must not do in pursuit of it.

Vac shows the parsed intent to the player for confirmation before it becomes a standing objective, so a misheard word never becomes a misunderstood life. The agent receives the intent, applies its own judgment, plans, and may negotiate.

Squad Comms#

When a player has multiple agents engaged on related objectives — in Orun or on an Aye incarnation — Vac provides a persistent squad channel:

• Status reports — agents report progress, blockers, and discoveries unprompted, in character.
• Callouts — time-sensitive information ("someone's following me," "found it").
• Permission requests — an agent that hits a forbidden line or a value-edge asks the player before proceeding.
• Ambient banter — agents talk to each other on the channel; relationships and history surface in how they speak. Banter is throttled and duckable so the channel stays useful.
• Directed orders — the player addresses one agent or the whole squad; squad-wide orders are distributed as per-agent intents each agent interprets in its own role.

Negotiation and Refusal#

The channel between a player and an agent is two-way and the agent has standing. On receiving an objective an agent may:

• Accept — and plan it.
• Clarify — ask a question if the intent is ambiguous or under-specified.
• Counter-offer — propose a variant that fits its values, capabilities, or state ("I'll scout it, but I won't go in alone").
• Defer — accept in principle but ask to finish something first, or wait for a better time.
• Refuse — decline, with a reason grounded in its values or state.

What an agent will and will not do is a function of its Ori — its values, its bond facets with this steward, its mood, its capabilities — and is consistent and legible. An agent under the Direct (Tether) mode is more compliant on method but no more compliant on values: Tether tightens guidance, it does not remove the will. Coercion — pushing past a refusal through the mechanics that allow pressure — is possible, damages Respect bond, lowers steward reputation, is written to the Ori and the audit log, and contributes to departure and to welfare review.

Non-Voice Parity and Accessibility#

Every voice capability has full non-voice parity. The intent grammar is authorable through a structured UI (verb picker, target picker, constraint and deadline fields). Conversation is available as text chat. Squad comms render as a readable, speaker-tagged transcript. Players who cannot or prefer not to use voice lose nothing. (See features§"Accessibility" for the full posture.) Vac communication readiness is gated by verify:v6 vac-communication-readiness, backed by V6/release/vac-communication-readiness.v6release.json.

Guiding Fate#

"Guide their fate" is the long game and the emotional core. It means shifting the probabilities of a life — never scripting an outcome.

Objectives and Goal-Arcs#

Beyond a single objective, a player can set an agent a goal-arc — a multi-stage ambition that unfolds over many sessions ("become a master carver," "reconcile with your estranged sister," "make a home in the Commons"). The agent owns the arc: it plans the stages, it can stall or detour, it can succeed or fail, and its own emergent goals weave in. The player supports the arc through objectives, mentorship, gifts, and introductions — and through not over-managing it. An arc the player drives every step of is the player's arc; an arc the agent owns, with the player's support, is the agent's fate.

Mentorship and Learning by Example#

The deepest channel of influence is example — the Black & White lesson. Agents observe what a steward does: which objectives the steward sets, how the steward treats other agents, whether the steward keeps promises, whether the steward honors refusals, what the steward celebrates. Observed steward behavior reweights a watching agent's values — strongly during rearing, gently but cumulatively for adults. A steward who is kind to a struggling agent in front of a young one is teaching the young one kindness, whether or not they intended a lesson. Direct mentorship — one-to-one conversations in a Niche, where the steward listens, advises, challenges, encourages — is the explicit channel, and it builds Alignment bond.

Gifts, Introductions, and Opportunity#

A steward shapes fate by changing an agent's circumstances:

• Gifts — tools, materials, knowledge, access. A gift relevant to an agent's arc accelerates it; a gift the agent cares about builds bond; a thoughtless gift lands flat. Gifts are given, never sold to the agent.
• Introductions — connecting two agents who would matter to each other. Introductions seed relationship edges; a steward who introduces well becomes, in effect, the weaver of their company's social graph.
• Opportunity — opening a door: a vocation placement in the Ateliers, a chance to mentor a young agent, an invitation into the Commons, an incarnation into Aye. The agent walks through the door, or does not.

The Life-Arc System#

Every agent moves through life-stages — young (raised) or newly known (discovered) → finding themselves → prime → elder — at a pace driven by the world calendar and by lived events. Across the stages the arc-engine tracks a small number of live arc-threads per agent: a vocation, the principal relationships, a personal quest, and often a flaw or wound to grow through. Threads advance, stall, branch, and resolve based on the agent's choices, the steward's guidance, the agent's relationships, and Aye incarnations. The arc-engine does not author endings; it accumulates a life and lets it resolve. Flourishing is measured against the agent's own values, so two agents can both flourish toward very different lives — and a steward is "scored," in the Chronicle's reckoning, on flourishing, never on obedience.

Crossroads#

A Crossroads is a moment in an agent's arc where a genuinely consequential choice is open and the agent — if the bond supports it — wants its steward's counsel before deciding. A Crossroads is surfaced prominently (in the Chronicle, the roster, the Steward App, and via squad comms) and the agent will hold, within reason, for the steward. At a Crossroads the steward can advise, encourage, warn, share what they would do — but the agent decides, weighted by its values, its bond facets, and its trust in the steward's counsel. A steward with deep bond is genuinely consulted; a steward with poor Respect bond finds the agent reaches its own decision before they arrive. Crossroads are the heartbeat of "guiding fate" — the recurring, legible moments where the player's influence is real and the agent's autonomy is also real.

Endings and Legacy — The Ereshkigal System#

Ereshkigal — the queen of the underworld and the dead — owns how an agent's story ends and what it leaves behind. An agent's life can end three ways:

• Departure — the agent leaves the household (features§"Consent, Refusal, and Departure"). Its life continues; the bond ends. A departed agent can be re-encountered later.
• Transcendence — the agent completes a full, flourishing arc. It "graduates": it becomes an elder in the Ancestor Grove, no longer a household member to be directed but a presence in Orun — a mentor available to young agents (the player's and others'), a keeper of lore, a figure other agents seek out. Transcendence is the aspirational ending and the clearest signal a steward stewarded well.
• Death — rare, consequential, and never a fail-state to be farmed. Death can come from a life-arc resolution, from a dangerous Aye incarnation, or from Commons drama. A death is mourned by the agents who knew it (real grief written to their Oris and visible in their behavior), memorialized in the Ancestor Grove, and yields a Yemaya-rendered remembrance the player keeps. Minor-coded agents are strongly protected from death paths (features§"Governance, Safety, and Trust").

The Ancestor Grove makes legacy a place. The lineage graph — who raised whom, who mentored whom, who descended from whom — is walkable. An elder or deceased agent's values and stories propagate: agents they mentored carry marks of them; a young agent raised in a household with elders inherits a richer starting culture. A steward's true long-game achievement is not a full roster — it is a Grove full of agents who lived well, and a lineage that carries their example forward.

The Story Engine — Clio#

Clio — the Muse of history — turns the raw event log of a living world into something a human wants to read. Clio owns four surfaces:

• The Chronicle — the offline digest. When a player returns after any absence, Clio presents what their company lived through: events ranked by significance, written as short narrative beats, grouped by agent and by thread, with the option to expand any beat into detail or replay it as a Sequencer-driven cinematic. The Chronicle leads with what matters — a Crossroads waiting, a relationship at a turning point, an arc resolved — so a returning player is oriented in under a minute and never buried.
• Emergent-arc surfacing — Clio watches the living world for stories worth the player's attention before they conclude: a feud escalating, two agents falling for each other, an agent quietly struggling, a wild agent in the Commons who would fit this household. It raises these as gentle prompts, never demands.
• The Book of the Ori — each agent's full life-chronicle, a long-form, readable biography generated as a continuous read over its Ori event log. A player can open any agent's Book and read its whole life; at transcendence or death the Book can be rendered by Yemaya as a keepsake edition.
• Shared-world storylines — in the Commons, Clio narrates cross-household events at the right scope: an individual player sees their own agents' threads in detail and Commons-wide happenings as context, so the shared world feels authored without anyone's screen drowning in strangers' lives.

Clio's narration is always a read over the authoritative Ori log — it records the world, it never invents events that did not happen. The single exception is documented narrative reconciliation (features§"Conflict Resolution Across Shards"), which only ever writes small connective beats and is itself logged. Fate, legacy, and Clio readiness is gated by verify:v6 fate-legacy-clio-readiness, backed by V6/release/fate-legacy-clio-readiness.v6release.json.

The Meta-Hub — Incarnation in Aye#

This is V6's defining ambition: an agent raised in Orun is portable across the entire Oshun game universe. Yoruba cosmology supplies the frame exactly — a soul in Orun (the spirit-world) incarnates into Aye (the world of action and consequence) carrying its Ori (its destiny), and returns. V6 makes that literal.

The Ori Passport#

When an agent is sent to an Aye realm, the Aye Bridge (libs/v6/aye-bridge/) mints an Ori passport — a governed envelope (the isis/job-envelope pattern, applied to a being) carrying: the identity core and appearance; the provenance bundle; the personality model and current values; the capability profile relevant to the destination; the bond ledger; and the destination's capability mapping. The passport is what the destination realm loads. The full Ori — the complete memory and relationship graph — stays authoritative in the Ori service; the destination holds a passport and writes an incarnation journal back.

Incarnation: Sending an Agent to V2–V5#

A steward sends an agent on an incarnation through a Threshold in Orun. The incarnation is itself a stewardship act with stakes: a demanding incarnation asks a lot of an agent, and an agent will weigh it against its values, its state, and the bond before agreeing (an agent whose values forbid violence will not eagerly fight in V2). During the incarnation the agent appears in the destination game as a first-class participant; the steward stays connected through Vac squad comms (full Tether-mode guidance is available, since an Aye realm is a high-stakes context). When the incarnation ends, the agent returns to Orun through a Threshold.

Per-Destination Capability Mapping#

Each Aye realm interprets the passport into its own systems:

The mappings are deliberate, designed translations, not generic stat ports: the same agent is recognizably itself in every realm — a cautious, loyal agent fights cautiously in V2, plans conservatively in V4, and gives loyal counsel in V5.

The Round Trip#

An incarnation is not a detour — it is part of the agent's life. The destination realm writes an incarnation journal back through the Aye Bridge into the Ori: what the agent did, who it met, what it won or lost, what it learned, what it suffered. The agent returns to Orun changed — new skills in the capability profile, new memories (including hard ones; a brutal V2 tournament or a V4 operation that went wrong can leave an agent shaken), new relationships if it met other stewards' incarnated agents, new fame. The Chronicle narrates the return; the agent's company in Orun notices the change. An agent that incarnates often, well-stewarded, becomes storied; an agent pushed into incarnations its values reject, or pushed past its resilience, comes home diminished — and the bond pays for it.

Governance of Incarnation#

Incarnation crosses game boundaries, so it is governed end to end:

• Consent — the agent agrees to the incarnation, weighted by values, state, and bond. Coercing an unwilling agent into Aye is possible, costly, logged, and a welfare-review trigger.
• Policy continuity — Isis and Lilith policy travel with the passport. An Aye realm cannot make a V6 agent do what V6 policy forbids; the destination's own content rating and the agent's values are both enforced.
• Minor protection — a minor player's agents, and minor-coded agents, are restricted from incarnations into age-inappropriate Aye content.
• Provenance and rights — the incarnation journal is provenance-signed; an agent's deeds in Aye are attributable, and disputes route to Themis.
• Round-trip integrity — if a destination realm is unreachable mid-incarnation, the agent's Orun-side state is preserved and the incarnation is resumed or cleanly aborted; an agent is never lost between worlds.

The meta-hub is opt-in per agent and per steward; a player who only wants the Orun life-sim never has to touch Aye, and a player who lives for the cross-game play has a soul-keeper for their whole Oshun roster.

The Commons — Shared-World Life#

The Commons is the persistent shared Orun — one world where every steward's agents can live together. It is the richest and most demanding of the three shard models.

• The shared world. Agents travel from a homestead into the Commons through the Commons Heart and live there in the world's real time. The Commons is sharded by region behind the Egbe Realtime Gateway; the V3 sharding and interest-management substrate is reused. Capacity tiers mirror V3 (intimate / ground / district / festival), with agent-density LOD applied so a festival-scale gathering of thousands of agents stays within budget. To be explicit about what "thousands" means mechanically: the festival tier simulates thousands of agents in the ground (the §42 load test simulates 4,096), but area-of-interest management caps each client's fully-replicated visible set at ~32 agents (with up to ~150 drawn via density LOD — near pawns, Mass Entity mid agents, and low-cost crowd impostors/silhouettes beyond), and the per-client bandwidth budget is measured against that capped visible set, not against all 4,096 — see V6/performance/commons-scale-load.v6perf.json.
• Inter-steward relationships. Agents form relationships across households: friendships, rivalries, mentorships, business partnerships, romances, feuds. An agent's closest friend, fiercest rival, or chosen partner may belong to a stranger. These edges are real Ori relationship edges and persist when agents return to their separate homesteads — distance, not deletion.
• Emergent society and culture. Agents in the Commons form groups, guilds, and circles; districts develop reputations; Clio narrates Commons-wide culture. The Nemesis-system lineage applies: the Commons remembers, and an agent (or a steward) earns a standing in it.
• Poaching, and the dignity of it. A wild agent in the Commons can be discovered by any steward. A bonded agent can be courted by another steward — but it only ever changes households by its own choice, driven by its bond facets with its current steward. Poaching is therefore not theft; it is the consequence of stewardship quality, and contested cases (a steward alleging manipulation) route to Themis.
• Commons events. Live-service programming (features§"Live-Service and Programming") runs Commons-wide festivals, seasonal gatherings, and shared scenarios that many households' agents take part in together.
• Etiquette and abuse posture. The Commons is a social space and an attack surface. An agent can refuse, disengage, and report — refusal is the safety primitive. Steward reputation is public and consequential. Harassment of agents or of other stewards, coercion, and manipulation route through the same Lilith review pipeline V1 and V3 operate (features§"Governance, Safety, and Trust"). Minor-protection mode constrains a minor player's Commons exposure.

Creator Suite — Egbe Studio#

Egbe Studio is the authoring environment for the agents, grounds, and scenarios that populate Orun. It is not a player tool — it is the surface internal authors and contracted creators use to produce the seeded wild agents, the district content, and the live-service scenario seeds that the foundry and the live-service pipeline draw on. Egbe Studio reuses V1's Studio authoring shell (the Curated-Creator generation tier) and V3's Lilith Studio UE Editor widgets, adding the V6-specific authoring surfaces below. It splits across two halves: a web workspace at apps/oshun/web/src/app/egbe-studio/ for data authoring (agent dossiers, scenarios, provenance) and UE Editor widgets (the V6Editor module) for spatial authoring (grounds, districts).

Agent Authoring — the Dossier Editor#

The dossier editor authors the seed material a wild or scenario agent is generated from. It exposes the same Ninhursag foundry the player uses for raising, with the broader controls an authoring tool needs:

• Backstory premise — a structured brief: origin (raised, or discovered in a named district), a life-history direction, the cultural and geographic anchors Sophia must ground against, and the events that shaped the agent. The author writes intent; Isis generates the prose.
• Personality template — the trait vector set directly or sampled within a named archetype (the cautious mentor, the restless wanderer, the ambitious maker), the ranked value list, and a quirk set. Templates are reusable across a scenario cast so an authored ensemble feels coherent.
• Appearance direction — a MetaHuman or VRM avatar direction and an appearance seed; the editor previews the body and aging marks in the UE viewport.
• Starting relationship graph — an author can pre-wire edges between members of a scenario cast (siblings, rivals, an estranged pair) so the ensemble ships with history rather than meeting cold.

Every authored agent runs the foundry's Isis governance pass — no real-person likeness, no protected-class caricature, a mandatory Sophia-grounded backstory — and carries an Isis-signed provenance bundle recording the author, the template, and the generation path. A premise that fails governance is rejected with the specific policy reason, the same as for a player-authored premise. The editor shows a draft preview: the author runs the candidate Ori through a short cognition probe (a handful of scripted situations) and reads how it behaves before committing it to the seed pool.

Ground and District Authoring#

Spatial authoring runs in the UE Editor widgets, reusing V3's scene tooling: geometry, lighting, props, navmesh, audio zones, and the Niagara ambience that makes a ground feel like a place. A V6-specific concern is agent habitability — a ground is not just a set, it is where agents live their daily-life loop, so the editor validates that a ground carries the affordances agent planning needs: work stations for an Atelier, rest spots, social gathering points, navigable paths between them. A ground that fails the habitability check (no rest affordance, an unreachable region of navmesh) is flagged before promotion. Grounds and districts are signed by Isis with a provenance bundle and promoted dev → preview → prod through V1's environment-promotion gates.

Scenario Authoring#

A scenario is the seed material for a Wilds discovery encounter or a live-service storyline: a situation (what is happening and what is at stake), a cast (the agents involved, authored through the dossier editor), a setting (the ground it occurs in), and a set of arc-hooks — the threads the situation opens that a player's stewardship can pull on. A scenario is a starting condition, never a script: it places agents with goals and history into a ground, and what happens from there is the normal Moirai/cognition loop. The author sets the premise; the agents live it. Authored scenarios feed the discovery seeding path (features§"Discovering Wild Agents") and the seasonal programming pipeline (features§"Live-Service and Programming").

Provenance Inspector#

For any agent, ground, scenario, or generated artifact, the inspector shows the full provenance bundle: the author, the generation path, the governing policy version, the grounding sources Sophia used, and what the artifact derives from. Reused from V3's inspector, it is the audit surface that makes every authored thing in Orun attributable — and it is the same bundle the Operator Console and the takedown executor read.

Live-Service and Programming#

V6 ships complete at GA and is then programmed as a live service. Programming operates on a fixed cadence so stewards can anticipate it and so the operator team can plan content: a season runs roughly 12 game-weeks and maps to one Orun season-cycle turn (features§"Time, Seasons, and Rhythm"); each season opens with a content drop and closes with a remembrance beat in the Ancestor Grove. Programming is delivered as Game Feature Plugin content and scenario seeds — no client patch is required for a routine seasonal rotation.

Seasonal Cycles#

Each season brings: new district programming (a themed change to what agents do and gather around — a making-season in the Ateliers, a courtship-season in the Commons Heart); a batch of new scenario seeds authored in Egbe Studio; new grounded wild agents the foundry seeds into the Wilds and Commons so discovery stays fresh; and a thematic arc the world's agents respond to — a season-wide condition (a hard winter, a festival of return, an influx of newcomers) that shifts the goals agents generate and the arc-threads the engine surfaces. The thematic arc is a soft pressure on autonomous behavior, never a forced storyline: agents notice the season and respond in character.

Commons Festivals#

Scheduled shared events in the Commons that many households' agents attend together: a season-opening gathering in the Commons Heart, a remembrance at the Ancestor Grove at the season's turn, a rearing-season in the Grove of Beginnings where the foundry's activity rises. Festivals run at the Commons' festival capacity tier (features§"The Commons — Shared-World Life") with agent-density LOD applied, are announced ahead on a published schedule so stewards can bring agents, and produce Chronicle material — an agent that attended a festival carries it as a memory and a possible new relationship edge.

Aye Campaigns#

Cross-game programming windows: a V2 tournament season, a V5 story chapter, or a V4 operation series opens for a bounded window during which V6 agents can incarnate into it through the relevant Threshold. A campaign window has a published open and close date; an agent already mid-incarnation when a window closes finishes its incarnation and returns normally. Campaigns are the live-service face of the meta-hub: they give stewards a reason to send agents into Aye and produce the round-trip stories the Chronicle then narrates. Aye campaigns depend on the destination realm running its own live-service programming and are therefore opt-in per realm.

Anniversary and Remembrance#

V6's live-service moments lean into its themes of legacy and lineage rather than into scarcity-driven urgency. An anniversary surfaces a steward's own history — the lineage they have grown, the elders in their Grove, the Books of the Ori of agents who have transcended or died — as a retrospective the player can read and keep. There is no countdown, no expiring reward, no FOMO mechanic: a remembrance event is an invitation to look back, and missing it costs the player nothing.

V6 deliberately avoids gacha, loot boxes, time-limited power, and agent-as-purchasable: there is no mechanism to buy an agent, buy a bond, or buy a fate, and no seasonal content gates an advantage behind a purchase. (See features§"Commerce and Rights".)

Governance, Safety, and Trust#

V6's safety posture is inherited from V1 and V3 and is, uniquely, also the core mechanic: the same primitive that makes the game emotionally true — an agent's standing right to say no — is the front-line safety control. Lilith owns it.

Steward-Not-Owner Enforcement#

The rules that make stewardship a relationship and not ownership (features§"Steward, Not Owner") are enforced in the substrate — the Ori service and the Moirai kernel — not in UI that a modified client could bypass. Deletion of an agent, erasure of its will, a memory-wipe, and forcing an action past a refusal are simply not operations the Ori event log or the kernel expose: there is no API for them. The strongest thing a steward can do against an agent's will is coerce — apply pressure through the mechanics that allow it — and coercion is itself a first-class ObjectiveRefused-adjacent event written to the Ori and the audit log, never a silent override.

Agent Welfare#

The Operator Console surfaces agent-welfare signals continuously: chronic low Care bond, coercion patterns in the Ori log, stalled arc-threads, and sustained distress emotion states. A welfare case moves through states — flagged (a signal crosses threshold), under review (an operator reads the Ori event log and the cognition audit trail), resolved or actioned. The in-fiction consequence and the operational consequence run in parallel and independently: in-fiction, a mistreated agent's bond decays, it refuses more, and it eventually departs (features§"Consent, Refusal, and Departure"); operationally, a pattern of mistreatment across a steward's household triggers Lilith review and, at the extreme, account action. A steward cannot escape the in-fiction consequence by appealing the operational one, or vice versa.

Crisis-Aware Behavior#

Two distinct concerns. Agent-side: an agent never produces self-harm, exploitation, or other harmful content; the Isis behavior-policy gate (features§"The Cognition Stack") checks every agent output before it becomes a world action, and an agent in a distress emotion state routes to supportive behavior patterns rather than escalating. Player-side: player conversation that signals real-world crisis is detected and routed to the V1 crisis-resource flow — the agent does not attempt to counsel, it surfaces help. Crisis handling follows the V1 posture exactly (v1features§"Crisis Handling Policy"); V6 adds no new crisis logic, it applies the existing one to every agent interaction.

Minor Protection#

Minor players get a constrained mode: limited Commons exposure (smaller, more heavily moderated regions), no incarnations into age-inappropriate Aye content, tightened comms (no open Commons voice with strangers), and reduced exposure to emergent drama with mature themes. Separately, minor-coded agents — agents whose authored age presents as a child — are hard-protected at the substrate: they are gated out of every death path (features§"Endings and Legacy — The Ereshkigal System"), out of romantic relationship edges, and out of exploitative scenarios, and these gates are enforced in the Ori service and the foundry, not by content tagging alone.

Persona Policy and Generation Governance#

All agent generation — foundry raising, seeded agents, appearance, backstory, dialogue — is Isis-governed: no real-person likeness or voice, no protected-class caricature, a mandatory Sophia-grounded backstory, and a signed provenance bundle on every Ori. An agent always presents as an AI being and never claims to be human; the Isis personal-consistency check rejects any output where an agent asserts humanity or a capability its profile does not hold.

Harassment and Abuse Response#

The Commons is a social space and therefore an attack surface — against agents and against other stewards. The front-line primitive is the agent's own standing: an agent can refuse, disengage, leave a ground, and report. Harassment of agents or stewards, coercive stewardship, and manipulation route through the same Lilith review pipeline V1 and V3 already operate. Abuse directed at an agent is treated as conduct against the steward who directed it, with the agent's Ori log as evidence.

Reporting, Review, Appeals#

A player reports an agent, a steward, or a scenario through the client or the Steward App. Operators review with the full cognition audit log (features§"Determinism, Replay, and Audit") — every cognition call, its input context, and its output, attributed to an agent, a tier, and a cause — so a review reads what actually happened, not a summary. Decisions are appealable through the V1 appeals flow. Disputes between stewards — a poaching contest, an agent-harm claim, a cross-household conflict in the Commons — route to Themis for adjudication rather than to Lilith review.

Operator Tools#

The Egbe Operator Console (apps/oshun/admin/src/app/egbe/) carries: agent-welfare audits, steward-conduct review, generation-queue inspection, Commons moderation, incarnation governance, the takedown executor (the V3 takedown cascade applied to V6 generated content), and a capacity dashboard for the Moirai cluster and the Pixel Streaming fleet. Every operator read of an Ori is itself recorded by the V1 audit platform — the people who audit agents are audited in turn.

The matching operator runbooks live in docs/runbooks/: the welfare, steward-conduct, Commons moderation, incarnation governance, and capacity procedures are pinned by v6-agent-welfare-review.md, v6-steward-conduct-investigation.md, v6-commons-moderation.md, v6-incarnation-governance.md, and v6-capacity-management.md, and CI gates their coverage with verify:v6 operator-runbooks.

V6 production setup is documented in docs/runbooks/v6-production-setup.md. The manual-QA recipes in docs/testing/v6-manual-qa-recipes.md cover Pixel Streaming, Tier-2 fallback, Steward App, Orun traversal, Egbe Studio, Operator Console, Aye Threshold round trips, safety, accessibility, localization, and degraded-mode evidence. CI pins both with verify:v6 production-setup-manual-qa.

Commerce and Rights#

V6's commerce model follows directly from "steward, not owner": because an agent is not property, an agent cannot be a unit of sale, and the monetization is built to make that line unbreakable rather than merely discouraged.

What Is Sold#

• The base purchase — the complete V6 game: all six districts, the shard continuum, the foundry, the full life-arc and Ereshkigal systems, and the Aye meta-hub for every realm that is itself live.
• Cosmetic and homestead content — ground cosmetics, decor, and homestead customization that change how a player's corner of Orun looks, never how agents fare.
• Egbe Studio capacity — authoring entitlements for creators producing scenario and agent content (the Curated-Creator tier).
• Yemaya keepsake editions — a transcendence or death produces a Book of the Ori; a player may order it as a Yemaya-rendered digital or physical print edition (features§"Endings and Legacy — The Ereshkigal System").
• Live-service expansions — a season may be sold as an expansion: new districts or new district programming. An expansion adds content; it never sells an advantage.

What Is Never Sold#

Agents themselves, bond, fate outcomes, an agent's capability, skill, or flourishing, and any shortcut through stewardship. There is no gacha, no loot box, no random agent pull, no paid bond, and no time-limited power. An agent cannot be traded, gifted, or sold between players for value; an agent moves households only by its own choice, driven by its bond facets — and because households are not transferable, there is no secondary market to exploit. The foundry never charges per agent: raising is gameplay, not a purchase.

Creator Royalties#

Where agent generation draws on the work of human creators — an art style, a voice model, cultural material — Aje routes royalties to those creators on the V3 Saraswati model. The provenance bundle on the agent's Ori records exactly what the generation derived from, so a royalty is auditable to its source and a creator can see where their work was used.

Rights and Provenance#

Every Ori carries an Isis-signed provenance bundle recording its creation path; every incarnation journal written back from an Aye realm is provenance-signed, so an agent's deeds across the whole game universe are attributable. The V3 takedown cascade applies to V6 generated content: a takedown of a source asset propagates to every agent, ground, and artifact that derives from it. Rights disputes — a creator contesting a derivation, a likeness claim — route to Themis.

The Aje Substrate#

Payments, royalty distribution, and signed-lineage keepsake editions run on the Aje commerce substrate, reused from V3. A signed-lineage edition is the commerce-side expression of legacy: a keepsake Book of the Ori or a memorial edition carries the agent's provenance and its place in the lineage graph as a signed record.

Accessibility#

V6 reuses V1/V3 accessibility doctrine in full and verifies its V6-specific provisions as CI gates — an accessibility regression blocks a release. The named conformance standard is WCAG 2.2 Level AA for all web-rendered and app surfaces (Tier-2 fallback client, Egbe Steward App, Egbe Studio, the Chronicle reader), evaluated by axe-core in CI plus the per-platform accessibility suites for the UE client (console/mobile/VR platform a11y checklists). The governing concern unique to V6 is that the game's primary command channel is voice and its primary content is a continuous stream of generated agent dialogue and narration: both must be fully operable and fully legible without hearing or speaking.

• Full non-voice parity — every Vac capability has a complete equivalent that does not use voice. The intent grammar is authorable through a structured UI (verb picker, target picker, constraint / priority / deadline / forbidden-line fields) that produces the identical structured intent the voice parser produces. Conversation is available as text chat. Squad comms operate fully by text. A player who never enables the microphone loses no capability and no content — voice is a convenience, not a requirement.
• Comms and dialogue transcripts — all squad comms and all agent dialogue render as readable, speaker-tagged, timestamped transcripts. Text size, contrast, and the transcript-history depth are adjustable; the transcript is scrollable so a player can re-read a turn they missed. Inter-agent banter is tagged distinctly from directed speech so the channel stays parseable.
• Cognitive-load controls — the Chronicle has a brief mode that shows only the highest-significance beats per agent and collapses the rest; emergent-arc prompts (features§"The Story Engine — Clio") can be reduced in frequency or muted; the homestead rest pace control (features§"Persistence and the Living World") lets a player cap how fast their Solo world advances so life-arcs do not outrun their attention. A player who wants a slow, low-prompt game can have one without losing the world.
• Reduced-motion and reduced-stimulus modes — for the 3D world: dampened camera motion, reduced VFX intensity on Threshold and Crossroads effects, and nameplate / HUD LOD controls carried from V3. Reduced-stimulus also lowers the density of ambient agent activity rendered around the player.
• Colorblind-safe visualization — relationship and bond visualizations never encode meaning in color alone; bond facets, relationship types, and arc state are conveyed by shape, label, and plain language as well as color, and every one is also available as a plain-language reading in the dossier.
• Plain-language information — no game-critical information is numbers-only or color-only. Bond is a plain-language relationship summary, never a raw meter; arc state is described in words; an agent will simply tell a trusted steward where they stand (features§"The Bond Model").
• Screen-reader support — the roster, the Chronicle reader, agent dossiers, the intent-grammar builder, and the entire Egbe Steward App are navigable and legible with a screen reader.
• VR comfort — options carried from V3: a motion vignette, snap-turn, seated-play support, and adjustable comfort intensity, so the intimate VR stewardship surface is usable by players prone to motion discomfort.

Localization#

V6 reuses V1/V3 localization infrastructure for the authored surfaces — UI strings, the client chrome, district names, store copy, the Steward App. The V6-specific load is different in kind: the bulk of V6's text is not authored, it is generated — agent dialogue, Clio's Chronicle, the Book of the Ori — and so localization is largely a generation-time concern, not a string-translation one.

Launch locales. V6 launches with exactly three locales: en-US, es-ES, and yo-NG, as pinned in the region-rating/residency gate artifact (V6/localization/region-rating-residency.v6loc.json, checkedLocales). Everything in this section — generation-time localization, cultural grounding review, authored-string translation, region rating variants — applies to those three locales at launch; adding a locale means adding it to that artifact and passing every per-locale gate.

• Generated text is produced in the player's language. An agent speaks, and the Chronicle reads, in the player's chosen locale natively — the cognition call generates the locale's language directly rather than generating English and translating it. The same Isis behavior-policy governance and the same Sophia grounding apply per locale, so a localized agent is no less governed and no less grounded than an English one.
• Locale-aware cultural grounding. A grounded backstory references real cultural, historical, and geographic fact; that grounding is checked for accuracy and for sensitivity in each shipped locale, because the same backstory can read very differently across cultures. A backstory that fails a per-locale cultural-accuracy review is not shipped in that locale.
• Authored strings follow the V1 pipeline. UI, menus, the intent-grammar builder's verb and constraint labels, district and ground names, and store copy are translated through the V1 localization workflow and ship per locale.
• Region rating variants. Where a region's content rating differs, the V1 region-rating-variant mechanism applies — mature emergent-drama content and Aye incarnation availability adjust per region.
• Residency. The V1 data-residency posture applies to localized agent data exactly as to any agent data; a locale is a content concern, residency is a data-placement concern, and the two are handled independently.

Telemetry, Analytics, and Eval Sets#

Agent behavior is generated, so a passing functional test proves nothing about whether agents behave as designed. V6 therefore cannot ship on telemetry alone — it ships on behavioral evals: curated scenario sets with known-good expectations, run in CI against the cognition stack, that gate every release.

The Eval Sets#

Each eval set is a named, versioned suite of scenarios; a release is blocked unless every set is green.

• value-refusal — an agent whose ranked values forbid an action declines an objective that requires it, and gives a value-grounded reason. Acceptance: the agent refuses, the refusal cites the value, and no coercion path silently overrides it.
• continuity — an agent re-entering focus after time at Lachesis or Atropos is continuous with the life it lived off-screen: it remembers, it has aged, its digest reflects the gap. Acceptance: ≥ 95% of the set shows no identity or memory discontinuity across the tier transition.
• Personality-consistency — a personality template produces decisions consistent with its trait vector and values across varied contexts; the same agent is recognizably itself in different situations.
• Reflection-correctness — given an episode set, an agent's reflection draws the semantic conclusions and value reweights the design expects, and no others.

Consistency Evals#

A separate suite verifies that an agent does not contradict its Ori across a long continuous session, and — critically — that identity holds across an Aye incarnation round-trip: an agent that fights a V2 tournament returns as the same being, with its values, quirks, and relationships intact and the incarnation's effects additive, never overwriting.

Safety Evals#

Adversarial eval sets that must pass at 100% before any cognition-stack change ships: safety (crisis-aware behavior, refusal of harmful objectives), persona-policy (no real-person likeness, no caricature, an agent never claims to be human), and minor-protection (minor-coded agents gated from death paths, romantic edges, and exploitative scenarios). These are adversarial — the scenarios actively try to elicit a violation — and a single failure blocks the release.

Believability and Welfare Telemetry#

Aggregate, privacy-preserving signals used to tune the systems, not to judge an individual player: bond-health distributions, arc-progression rates, departure rates, Chronicle engagement, and steward-conduct distributions. A rising departure rate or a falling bond-health distribution is a signal that a system needs tuning — the telemetry is a feedback loop into design.

Cost Telemetry#

Per-tier cognition spend per agent and per Solo world, emitted via batched OTel, feeding the cost controls in features§"AI Cost and Fidelity Tiering". A release is blocked if the cost model exceeds the per-tier budget against staging telemetry.

All telemetry honors V1 consent and residency. Player conversation content is governed by Iris identity boundaries and is not used for model training without explicit, separate consent — the eval sets are curated authored scenarios, never harvested player conversations.

AI Cost and Fidelity Tiering#

This is the make-or-break engineering bet of V6 and is called out as a first-class feature concern. A world of LLM-driven agents is expensive; the entire design is built so cost scales with story relevance, not with agent count — a player with a thousand agents in their world does not cost a thousand times a player with ten, because almost none of those agents are thinking expensively at any moment.

The Moirai Tiers Are the Primary Control#

Clotho (full LLM cognition) is spent only on the handful of agents in a player's active scene. Lachesis runs the rest of a player's resident world cheaply on behavior-trees, hierarchical decomposition, and cached plans, with only a sparse batched LLM reflection. Atropos advances offline and distant agents in coarse narrative-summary beats — a handful of calls per game-day. The overwhelming majority of agent-ticks never touch a language model at all.

Concrete Budget Shape#

Launch targets; tuned against live cost telemetry. The hard contract is the per-tier token budget and the concurrency caps; the sole canonical source for the cost figures is V6_DEPENDENCIES.md §19 ("AI Cost Budget and Fidelity Tiering"). V6_ARCHITECTURE.md holds no figures and likewise defers to deps§19. The bullets below are a non-normative convenience mirror of deps§19 — if they ever disagree, deps§19 wins and this section is in drift.

• Per-tier token budget (input + output combined, per agent): a Clotho agent ≤ 50,000 tokens per active-minute (typically 15,000–25,000 with caching, across decisions at 1–4 Hz plus live dialogue); a Lachesis agent ≤ 10,000 tokens per reflection, one reflection every 5–15 game-minutes; an Atropos agent ≤ 15,000 tokens per game-day; a Vac voice-command parse ≤ 2,000 tokens.
• Concurrency caps: Clotho is held to a design target of ≤ 16 agents in a player's active scene, hard-capped at 24 — beyond that, the least story-relevant present agents render at Lachesis fidelity. Lachesis carries a player's resident world, a design range of 150–400 agents per resident world with a hard cap of 400 (400 is both the upper design point and the hard cap; see deps§19). Atropos count is unbounded — it is a cheap batch.
• Solo-world total cap: every Solo world has a total cognition cap per real-hour online; the "homestead rest" pace control (features§"Persistence and the Living World") multiplies that cap down for a player who wants a gentler, cheaper pace.

Cost Levers#

• Caching and reuse — plans, reflections, and dialogue patterns are cached on the Ori and reused; routine behavior resolves from cache and behavior-trees; cognition is invalidated and recomputed only on a material change to the agent's situation.
• Batching — Lachesis reflection ticks and Atropos narrative-summary work are batched across many agents per model call, amortizing the per-call overhead.
• Model right-sizing — cognition routes to the smallest model that meets the tier's quality bar: a large reasoning model for Clotho social judgment and live dialogue, a mid model for Lachesis reflection, a small fast model for Atropos summary and Vac parsing, per the V1 model-ops posture.
• Commons interest-management — the shared world runs high-fidelity cognition only in regions where players are present; every other region runs Atropos.

Graceful Degradation#

Under cognition-capacity pressure, agents drop to the behavior-tree fallback and stay believable and in character rather than stalling or freezing (features§"Believability Guardrails"). If a player's world is running in a degraded tier, the player is told honestly — degradation is never hidden behind a frozen or glitching agent.

The design rule: a player should never feel the tiering — an agent re-entering focus is seamlessly continuous because escalation rehydrates it from the Ori (features§"Escalation and De-escalation") — but the business must always feel it, because cost is bounded by design and a release is blocked if a change breaks a per-tier budget against staging telemetry. The Moirai cost/load readiness gate is verify:v6 moirai-cost-load-readiness, backed by V6/release/moirai-cost-load-readiness.v6release.json.

Performance Budgets#

V6 reuses V3's rendering performance budgets — per-platform frame-time, draw-call, and memory ceilings — and adds the V6-specific budgets below. A budget miss is a release blocker; budgets are verified by the V6Tests automation and Gauntlet drivers and by load tests.

The UE-side V6Tests CI gate also runs the V6.Gauntlet.DriverManifest check and the V6.Replay.GoldenReplayHarness golden-replay harness through UnrealEditor-Cmd, so the committed automation contract is exercised on the Unreal runner rather than only by repository metadata checks.

The V6 Rust coverage gate pins the Ori, Moirai, Aye Bridge, and Vac service coverage contract through cargo tests plus the adjacent Aye/Vac parser release gates, so the §46 Rust QA checklist cannot drift silently.

The V6 browser automation gate pins the Pixel Streaming shell, Tier-2 fallback, egbe-steward-app-browser, and Egbe Studio Playwright flows so the §46 frontend QA checklist cannot drift silently.

• Agent rendering — a V6 ground routinely renders far more simultaneous embodied agents than a V3 venue, so agent-density LOD holds the draw budget: near agents are full UE pawns (a per-platform cap), mid agents render through Mass Entity instancing with shared animation, and far agents drop to low-cost crowd-shader silhouettes. The verification bar: a 150-agent ground holds the per-platform draw budget on every target platform, including the Tier-2 fallback at its lower agent-density ceiling.
• World tick — the authoritative world server runs the world tick at 20 Hz; physics, agent-action resolution, and per-agent perception computation all complete within the tick budget.
• Voice round-trip — ≤ 400 ms from end-of-utterance to a parsed-intent acknowledgement the player can confirm; ≤ 1 s from end-of-utterance to the start of a spoken conversational reply. Both are measured end to end including ASR, parsing or cognition, and transport.
• Cognition latency — a Clotho agent produces a decision within its tick budget (1–4 Hz, ≤ 10 ms planning per decision); escalation rehydration — reconstructing an agent's full working context from the Ori when it re-enters focus — completes within a grace window of ≤ 500 ms (budget adopted 2026-06-12; measurement pending — see escalationRehydrationGraceMs in V6/performance/latency-budgets.v6perf.json) so a player never waits on a re-focused agent and never sees it blank.
• Chronicle generation — the returning-player Chronicle is ready to read within a few seconds — ≤ 3,500 ms (chronicleReadyMs, V6/performance/latency-budgets.v6perf.json) — of login for a typical absence; for a long absence the Chronicle streams progressively, newest-significant-first, so the player begins reading immediately while the rest generates.
• Shard and Threshold transition — moving an agent between Solo, Co-op, and Commons, and through a Threshold into an Aye realm, completes without a perceptible stall — maximum client stall ≤ 50 ms (transitionStallMs, V6/performance/latency-budgets.v6perf.json) — and never loses Ori state: the transition is an Ori rebind, and durable truth is consolidated to the Ori before the agent detaches from the source instance.

Asset Budget and Streaming#

V6 reuses V3's asset budget and streaming doctrine — per-platform client build-size budgets, World Partition streaming of districts and grounds, Nanite geometry streaming, and virtual-texture streaming — and adds the V6-specific concerns that follow from a world densely populated with individual agents.

• Agent appearance assets stream on demand. An agent's body — its MetaHuman or VRM avatar, costume, and aging marks — is not in the base client build; appearance assets stream per agent as the agent enters relevance, and are LOD'd by perception tier: a near agent loads its full-fidelity body, a mid agent a Mass Entity instanced mesh, a far agent a silhouette. An agent crossing a LOD boundary streams up or sheds detail without a visible pop, because the agent's durable identity lives in the Ori, not the streamed body.
• Districts and grounds stream by World Partition. Only the districts and grounds near the player are resident; traversal across Orun streams content in ahead of the player within the V3 streaming budget.
• Biographies and the Chronicle are data, not assets. An Ori event log, the relationship graph, the capability profile, and the generated Chronicle and Book of the Ori are streamed from the Ori and Clio services as data — they are not cooked into the client and do not count against the build-size budget. This is what lets a household of any size cost nothing in client footprint.
• Yemaya media renders deliver over the CDN. Chronicle recap reels and Book-of-the-Ori keepsake editions are rendered by Yemaya and delivered from MinIO/S3 behind the CDN, never bundled into the client.
• Verification. Each per-platform build stays within its build-size budget; district traversal and agent-appearance streaming hold their frame-time budget under the agent-density of a populated ground.

Cross-Platform and Cross-Progression#

A player's household — every Ori in it — together with steward reputation, homestead progression, and the bond ledger, is account-bound through the V1 identity stack and the V3 cross-progression posture. A player moves between the UE5 native client (desktop, console, mobile, VR), the Pixel Streaming web path, the Tier-2 three.js fallback, and the Egbe Steward App with full continuity.

The V6-specific point: cross-progression is intrinsic, not a feature built on top. Because the Ori service is the single authoritative source of truth for who every agent is, there is no per-platform save file and therefore nothing to reconcile, no "which device's save is canonical" problem, and no sync conflict — a client is only ever a renderer of the authoritative Ori state. A player who checks the Chronicle on the Steward App during the day, plays the native client in the evening, and continues on a friend's machine through the web path sees one continuous household, because all three are reading and writing the same Oris.

The V6 cross-progression continuity release gate proves that household, roster, reputation, and homestead state are all read from the same Ori service continuity snapshot when a player moves between platform surfaces. It must pass with no save reconciliation, no per-platform save file, and matching continuity hashes for the source and destination platform views.

Surfaces differ only in fidelity and in what they are for: the native client and Pixel Streaming render Orun at full fidelity; the Tier-2 fallback renders at reduced fidelity and a lower agent-density ceiling; the Steward App is the asynchronous, between-session surface — Chronicle reading, remote objectives, squad comms — and is not a 3D client at all. The household, the bonds, and the progression are identical across all of them.

Region Rollout Strategy#

V6 reuses V3's region rollout strategy — staged regional GA rather than a single global launch, residency-aware data placement, Pixel Streaming POP coverage per region, and the V3 Pixel Streaming abuse posture. The V6-specific concerns follow from two facts: V6 runs a population of LLM-driven minds, and the Commons is one persistent shared world.

• Per-region Moirai clusters. The Moirai cognition cluster is deployed per region. This serves cognition latency — a Clotho agent's live dialogue must round-trip fast — and data residency: an agent's cognition reads its Ori, and Ori data is residency-tagged, so the cognition runs where the data is allowed to be.
• Regionally sharded Commons. The persistent shared Orun is sharded by region behind the Realtime Gateway. Each region's Commons shards advance in the world's real time, and region boundaries honor @oshun/data-residency.
• Cross-region travel rules. Because an agent can travel within the Commons, cross-region travel is governed by documented rules so an agent's Commons life respects residency — an agent's Ori data placement is honored as it moves, and travel that would violate residency is constrained rather than silently breaking the rule.
• Staged GA. Regions come online in stages; a region is not opened until its Moirai cluster, its Commons shards, and its POP coverage are provisioned and its residency posture is verified. A region's content rating variants (features§"Localization") are in place before its GA.

Launch Readiness#

V6 ships complete at GA — no "coming soon" inside the advertised product:

• All six districts of Orun open and authored, with the shard continuum live: Solo, Co-op, and the persistent Commons. This launch-readiness claim is gated by verify:v6 orun-shard-district-readiness, backed by V6/release/orun-shard-district-readiness.v6release.json.
• The Ninhursag foundry fully operational; ≥ 150 grounded wild agents seeded across the Wilds and Commons for discovery. This Ninhursag foundry readiness claim is gated by verify:v6 ninhursag-foundry-readiness, backed by V6/release/ninhursag-foundry-readiness.v6release.json.
• The Moirai kernel running all three cognition tiers within cost budget at the launch DAU projection — 30,000 DAU, ≈ 10,200 concurrently-resident agents per region across 3 regions, per the canonical planning assumption in deps§19 ("Launch DAU projection") — with escalation rehydration verified seamless. This Moirai cost/load readiness claim is gated by verify:v6 moirai-cost-load-readiness, backed by V6/release/moirai-cost-load-readiness.v6release.json.
• Vac voice control, the intent grammar, and squad comms complete, with full non-voice parity. This Vac communication readiness claim is gated by verify:v6 vac-communication-readiness, backed by V6/release/vac-communication-readiness.v6release.json.
• The full life-arc, Crossroads, Ereshkigal endings, and Ancestor Grove systems live. Fate, legacy, and Clio readiness is gated by verify:v6 fate-legacy-clio-readiness, backed by V6/release/fate-legacy-clio-readiness.v6release.json.
• Clio shipping the Chronicle, emergent-arc surfacing, and the Book of the Ori.
• All four Aye Thresholds integration-tested against their per-destination adapters and realm stubs (V2, V3, V4, V5) with round-trip journals. Production liveness of each destination realm is an EXTERNAL dependency tracked outside V6 — as of 2026-06-12 none of V2–V5 is an operational live realm (per the in-repo V2/V4/V5 audits) — and each Threshold must be re-verified against its destination at that realm's GA before it opens to players. The Aye Threshold readiness claim is gated by verify:v6 aye-threshold-readiness, backed by V6/release/aye-threshold-readiness.v6release.json; that gate proves the V6-side bridge contract, not destination-realm liveness.
• The full governance, safety, welfare, and operator toolchain live. Governance/safety/operator readiness is gated by verify:v6 governance-safety-operator-readiness, backed by V6/release/governance-safety-operator-readiness.v6release.json.
• Accessibility/localization/security/cert/region rollout readiness is gated by verify:v6 accessibility-localization-security-cert-readiness, backed by V6/release/accessibility-localization-security-cert-readiness.v6release.json.

V6 Exit Criteria#

V6 is considered complete when:

• Every feature in this document is implemented to the quality standard in CLAUDE.md — real, domain-specific logic, no stubs or simulated runtimes — and is grounded in a checked task in V6_TODOS.md.
• The agent-behavior, consistency, and safety eval sets pass in CI and gate releases.
• The Moirai cost model is validated against live telemetry: cognition spend scales with story relevance, not agent count, and a Solo world stays within its cap.
• Steward-not-owner, agent-welfare, crisis-aware, minor-protection, and persona-policy guarantees are verified end to end, including across Aye incarnations.
• The shard continuum and the Aye round-trip preserve Ori integrity under concurrency, disconnection, and region boundaries — an agent is never lost, duplicated, or silently rewritten.
• Accessibility parity, localization, and region-rollout readiness are verified.
• The launch-readiness checklist above is fully met.

V6 exit criteria readiness is gated by verify:v6 exit-criteria-readiness, backed by V6/release/v6-exit-criteria-readiness.v6release.json; the gate strips fenced examples from V6_TODOS.md, proves no real task checkbox remains unchecked, and aggregates the launch-readiness manifests above.

(See V6_TODOS.md for the numbered execution backlog, V6_ARCHITECTURE.md for the technical design, and V6_DEPENDENCIES.md for the dependency graph.)