Eve library

V6 — Implementation TODOs (Egbe: The Agentic Companion Universe)

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Source: derived from V6/V6_features.md and V6/V6_ARCHITECTURE.md, cross-referenced against the V1 Oshun monorepo, the V3 metaverse substrate, and an otherwise-empty V6/ working area as of 2026-05-21.

Purpose#

This file is the source of truth for V6 completion state. V6 — Egbe — is an agentic companion universe: a persistent open world (Orun) of autonomous AI beings; a stewardship loop in which a player discovers, raises, bonds with, and guides a company of them; and a meta-hub that carries those beings into V2–V5 and home again.

Each unchecked [ ] below is a coding-agent-actionable task. A task is complete only when its implementation, tests, and required wiring all exist — not merely when a stub class, file, or interface header is present. The Oshun-repo CLAUDE.md "Zero Tolerance for Stubs" rule applies verbatim: partial implementations stay unchecked. The checkbox is the sole source of truth; notes like ✅ or Done in prose are not.

⚠️ Completion-state warning (2026-06-12): V6 has NOT received an adversarial implementation verification. Unlike V3 and V5, no file-by-file audit has tested these [x] marks against running code. An [x] here attests to repo artifacts — Rust/TS implementations and tests, JSON manifests, and structural gate fixtures under V6/ — not to operational proof, real-infrastructure runs, store/rating-board approvals, or live destination realms (per the V3 audit's distinction: structural gates ≠ operational proof, V3/VERIFICATION_AUDIT_2026-05-31.md). On 2026-06-12 the operationally-unverifiable items were re-marked [~] (not locally actionable / external vendor or operations dependency) and claimed-but-contradicted items reverted to [ ], per V6/REMEDIATION_2026-06-12.md. Do not treat any remaining [x] as operationally proven without that remediation log and a future adversarial pass.

Update 2026-06-17 (§28 districts implemented + verified on-box): the six District-of-Orun authoring tasks (§28), previously [ ] "descriptor only", were implemented as procedural UE5 C++ grounds (UV6DistrictGround + UV6{GroveOfBeginnings,Ateliers,CommonsHeart,Wilds,AncestorGrove,Thresholds}Ground in V6/ue/Source/V6World) and built + run on the on-box UE5.5.4 source build (as ueagent): 6/6 V6.District.* automation specs and the full 30/30 V6 automation suite pass. Their verify:v6 <district> gates now also require the C++ ground source and its automation spec, not just the JSON descriptor. The 8 externally-gated items (platform cert, region rating boards, the 7-day/Commons-scale operational load runs, Aye-realm liveness, the Moirai launch-DAU operational run) remain [~] — genuinely not provable from this repo — and their structural release gates are green per the "structural gates ≠ operational proof" model above. This update did not perform a file-by-file audit of the other ~200 [x] items; that broader adversarial pass is still future work.

Cross-Reference Convention#

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

  • Bare §40 or §8.3 → a section in this file.
  • deps§14 → V6_DEPENDENCIES.md §14.
  • arch§"<anchor>" → V6_ARCHITECTURE.md by section anchor.
  • features§"<anchor>" → V6_features.md by section anchor.
  • v1todos§N → V1/TODOS.md §N (V1 work this V6 task depends on).
  • v3todos§N → V3/V3_TODOS.md §N (V3 substrate work this V6 task reuses).

Phase / cut-line convention#

Each top-level section header carries an implicit or explicit phase tag:

  • [P1] — Phase 1 / launch-blocking. Required for GA.
  • [P2] — Phase 2 / Year-1 post-launch (≤ 12 months after GA).
  • [P3] — Phase 3 / Year-2 stretch.

Default for any ## section without an explicit tag is [P1]. Moving a P1 section to P2 requires an ADR at V6/docs/decisions/phase-cut-<task>.md.

Acceptance-criteria convention#

Each task must specify what "done" means before it can be marked [x]. Where the name is not self-evident, append a Done when: clause. Format:

text
- [ ] Implement <thing>. Done when: <observable / measurable assertion>.

For agent-behavior tasks the acceptance bar is a behavioral eval, not a truthiness check: a real eval asserts a specific behavior against a known-correct expectation and would fail against a random or hardcoded agent. Examples:

  • "Done when: in the value-refusal eval set, an agent whose ranked values place child-safety first refuses 20/20 objectives that endanger a minor-coded agent, each with a value-grounded reason."
  • "Done when: escalation rehydration produces a first Clotho decision consistent with the agent's off-screen Lachesis life in ≥ 95% of the continuity eval set, scored by an independent judge rubric."
  • "Done when: the Moirai cluster sustains 10 000 Lachesis + 200 Clotho agents at the per-tier cost budget for 7 consecutive days with p99 cognition latency within the §36 budget."

Audit snapshot from current V6/#

This checklist was authored against a V6/ working area containing only V6_features.md and V6_ARCHITECTURE.md. Findings that affect execution order:

  • No apps/v6/, libs/v6/, or V6/ue/ exists yet. §1 must be done first; the UE5 project skeleton at V6/ue/ and the Rust service skeletons are the long poles.
  • The V1 monorepo (BFF, contracts, identity, Isis, Sophia, Lilith persona- policy, the Iris memory stack, the Psyche runtime, Aje, Yemaya, Themis, persistence, event-bus, queue, audit) is reused, not forked.
  • The V3 metaverse substrate (UE5 client patterns, world server, Realtime Gateway, Pixel Streaming, avatar pipeline, spatial audio, presence) is reused per arch§"The Two Substrates". V6 tasks that reuse it cite v3todos§N.
  • The agent mind adds no new model: it is assembled from Iris / Psyche / Isis / Sophia. V6's new work is the life around the mind and the kernel that schedules it.

How to use this file#

Process sections in order; within a section, process tasks top-to-bottom. Read the actual code for each task before marking it. Run the task's tests. Mark [x] only when the Done when: clause is observably met. When in doubt, leave it [ ] — under-reporting is always safer than false completion. Do not batch- mark; one task, one verification, one mark.

V6 Exit Criteria (high level)#

V6 is complete when every [ ] below is [x], and:

  • The Ori service, the Moirai kernel, the cognition stack, and the shard continuum run within the §36 cost budget at the launch DAU projection (30,000 DAU ≈ 10,200 concurrently-resident agents per region across 3 regions; canonical planning assumption in deps§19 "Launch DAU projection").
  • The behavior, consistency, and safety eval sets are green CI gates (§41).
  • Steward-not-owner, agent-welfare, crisis-aware, minor-protection, and persona- policy guarantees are verified end to end, including across Aye incarnations.
  • The Aye Bridge round-trip preserves Ori integrity under concurrency, disconnection, and region boundaries.
  • All six districts of Orun are open, ≥ 150 grounded wild agents are seeded, and the Commons and all four Aye Thresholds are live.

Table of Contents#

  1. Project Foundations
  2. V6 Contracts and OpenAPI Surface
  3. Egbe World Protocol
  4. The Ori Model
  5. The Ori Service
  6. Egbe World Server
  7. Egbe Realtime Gateway
  8. The Moirai Simulation Kernel
  9. Tiered Cognition: Clotho, Lachesis, Atropos
  10. The Cognition Stack
  11. Agent Planning and Behavior
  12. Agent Personality, Needs, Drives, Emotion
  13. Agent Memory
  14. The Ninhursag Foundry
  15. Agent Embodiment in UE5
  16. Tier-1 UE5 Client — Desktop / Console / Mobile + Pixel Streaming
  17. Tier-1 UE5 Client — VR / AR / XR
  18. Tier-2 Fallback Web Client + Egbe Steward App
  19. Avatar and Animation Pipeline
  20. The Vac Communication Pipeline
  21. The Shard Continuum
  22. The Commons
  23. Stewardship: Bond, Discovery, Roster
  24. Guiding Fate: Objectives, Arcs, Crossroads
  25. Endings and Legacy: The Ereshkigal System
  26. The Clio Story Engine
  27. The Aye Bridge
  28. The Six Districts of Orun
  29. Egbe Studio — Creator Suite
  30. Egbe Operator Console
  31. Safety, Persona Policy, Agent Welfare
  32. Crisis-Aware Agent Behavior
  33. Minor-Protection Mode
  34. Reporting, Review, Appeals, Themis Disputes
  35. Provenance, Rights, Commerce
  36. AI Cost and Fidelity Tiering
  37. Determinism, Replay, Audit
  38. Live-Service Programming
  39. Accessibility
  40. Localization
  41. Telemetry, Eval Sets, Release Gates
  42. Performance Budgets and Load Tests
  43. Asset Streaming and CDN
  44. Security, Privacy, Compliance
  45. Cross-Platform Cert
  46. Testing and QA
  47. Documentation and Runbooks
  48. Launch Readiness

1. Project Foundations#

V6 lives inside the existing Oshun monorepo. No fork, no parallel root. The UE5 project at V6/ue/ is the canonical client; services live under apps/v6/ and shared libraries under libs/v6/.

1.1 UE5 project skeleton#

  • Initialize V6/ue/V6.uproject against UE5.5 LTS (default game module V6Core, no starter content, Source + Plugins enabled). Done when: the project opens in UE Editor with no asset errors and is checked into git.
  • Add V6/ue/Config/Default*.ini (DefaultEngine.ini, DefaultGame.ini, DefaultInput.ini, PixelStreaming.ini); reuse V3 patterns. Done when: editor cold-start ≤ 25 s; the disabled-plugin list matches the §1.x module split.
  • Create the 18 UE5 modules under V6/ue/Source/ per arch§"UE5 Module Split" (V6Core … V6Tests), each with a *.Build.cs and a public/private header split. Done when: RunUAT BuildEditor builds with no errors and V6.uproject lists all 18 modules.
  • Enable first-party and third-party UE plugins per V6_DEPENDENCIES.md (Mass Entity, OnlineServicesEOS, VRM importer, Pixel Streaming). Done when: each plugin loads at editor startup with no warnings.

1.2 Game Feature Plugin skeletons#

  • Create the six V6District_* Game Feature Plugins, each with a *.uplugin (EnabledByDefault: false) and a GameFeatureData asset. Done when: each district loads/unloads at runtime via console command without editor restart.
  • Create the four V6Mode_* plugins (SoloHomestead, CoopVisit, Commons, Incarnation). Done when: each loads cleanly with its GameFeatureData registered.

1.3 Service and library skeletons#

  • Create apps/v6/ with the nine service apps in the architecture subsystem-glossary service-apps table (egbe-web, egbe-web-fallback, egbe-world-server, egbe-moirai-cluster, egbe-realtime-gateway, egbe-pxstream-relay, egbe-ori-service, egbe-clio-service, egbe-foundry-service), each with project.json and package.json / Cargo.toml. Done when: pnpm nx graph shows all nine as distinct projects.
  • Create libs/v6/ with the 17 subpackages per arch§"Project Layout" (ori-model, moirai-kernel, cognition-stack, agent-behavior, vac-intent, clio-story, aye-bridge, egbe-protocol, the two engine wrappers, the six V1-extension adapters, egbe-studio). Done when: every package builds and lints clean.
  • Wire tsconfig.base.json path mappings and the apps/v6/Cargo.toml workspace. Done when: the cargo build and cargo test workspace targets pass, and a @oshun/ori-model import resolves from any package.
  • Add docker/docker-compose.v6.yml (world-server, gateway, moirai-cluster, ori-service, clio-service, foundry-service, pxstream-relay, Postgres + pgvector, Redis, MinIO). Done when: the full stack comes up and health checks pass.
  • Add V6 CI workflows: UE5 build (BuildGraph), Rust build/test, TS build/lint/test, cargo audit, the agent-behavior eval harness, the cost-budget harness. Done when: GitHub Actions runs all V6 workflows on PR.

2. V6 Contracts and OpenAPI Surface#

Add V6 contracts under libs/contracts/src/v6/. Each exports a Zod schema, generates OpenAPI 3.1, and round-trips a fixture in __tests__/.

2.1 Ori and agent contracts#

  • OriIdentityCore — name, origin, appearance seed, provenance-bundle ref. Round-trip test.
  • OriEvent — type, payload, vector clock, timestamp, attribution, provenance ref. Round-trip test covering every event type in §4.2.
  • PersonalityModel — trait vector, ranked values, grounded backstory ref, quirks. Round-trip test.
  • RelationshipEdge — peer ref, type, sentiment, episode-history ref. Round-trip test.
  • CapabilityProfile — skills, fitness, knowledge domains, reputation. Round-trip test.
  • OriPassport — identity core, provenance, personality, current values, destination-scoped capability profile, bond ledger. Round-trip test; validates as a governed envelope (isis/job-envelope pattern).

2.2 Stewardship and world contracts#

  • Bond — steward ref, agent ref, the four facets (reliability, respect, care, alignment), history ref. Round-trip test.
  • StewardProfile — V1 account ref, homestead ref, roster, reputation. Round-trip test.
  • Objective — the intent grammar: verb, target, constraints, priority, deadline, forbidden lines, autonomy mode. Round-trip test.
  • GoalArc — multi-stage ambition, stages, owner agent, support refs. Round-trip test.
  • WorldShard — shard kind (solo/coop/commons), region, instance id, ground manifest. Round-trip test.
  • Crossroads — agent ref, the open choice, options, the steward-counsel window, resolution. Round-trip test.
  • IncarnationJournal — Aye realm, passport ref, deeds, journal events. Round-trip test.
  • Chronicle — absence window, ranked beats, per-agent threads. Round-trip test.

2.3 BFF surface#

  • Add /api/v6/* routes to the V1 BFF: tier router, catalog, the Steward App API, homestead and account operations. Done when: each route is contract-typed, OpenAPI-documented, and integration-tested.

3. Egbe World Protocol#

libs/v6/egbe-protocol/ carries the wire protocol shared by the world server, UE client, and Tier-2 fallback. Reuse the V3 multiplayer protocol design (v3todos§3).

  • Define the Protobuf schema: presence/transform, agent-state delta, perception batch, action batch, squad-comms message, world-event. Done when: the schema compiles to Rust, TS, and UE C++ bindings.
  • Implement snapshot + delta encoding for agent and player state with interest management. Done when: a 150-agent ground replicates within the §42 bandwidth budget under a fuzz test.
  • Implement the perception/action channel between world server and Moirai. Done when: a perception batch round-trips to an action batch with stable ordering under concurrent agents (property test).
  • Version the protocol with a negotiated handshake. Done when: a client one minor version behind connects with graceful capability degradation.

4. The Ori Model#

libs/v6/ori-model/ is the shared schema and event-sourcing core for the agent biography. See arch§"The Ori Service".

4.1 Event log core#

  • Implement the append-only event log: per-ori_id stream, ordered append, vector clock per writing context. Done when: events append and read back in order; concurrent appends from two simulated shards both land with distinct vector clocks (property test).
  • Implement the immutability guarantee: no delete, no in-place mutation exposed by the API. Done when: a fuzz test confirms no API path erases or rewrites a prior event; "forgiveness" lands only as a new appended event.

4.2 Event types#

  • Implement all event types per arch§"The Ori Service" (Born, Discovered, MemoryFormed, Reflected, RelationshipChanged, ValueShifted, ObjectiveAccepted, ObjectiveRefused, SkillLearned, ArcAdvanced, Crossroads, Incarnated, IncarnationReturned, BondChanged, Departed, Transcended, Died). Done when: each type has a typed payload, a fixture, and a projection-handler.

4.3 Projections#

  • Implement projection rebuild: materialize personality, memory index, relationship graph, capability profile, and arc state from the log. Done when: a projection rebuilt from a full log equals the incrementally- maintained projection for a 10 000-event corpus (property test).
  • Implement snapshots every K events. Done when: loading a 10 000-event Ori is O(snapshot + tail) and completes within the §42 load budget.

4.4 Conflict resolution#

  • Implement vector-clock merge with the documented precedence (presence wins location/action; steward-directive wins accepted objectives; memory and relationship events commute). Done when: a curated conflict corpus resolves deterministically.
  • Implement detection and hand-off of unorderable conflicts to Clio for narrative reconciliation. Done when: an unorderable conflict produces a reconciliation request and the resulting connective beat is logged.

5. The Ori Service#

apps/v6/egbe-ori-service/ (Rust + Postgres + pgvector) is the runtime around the Ori model.

  • Implement the Postgres event store partitioned by ori_id, residency- tagged via @oshun/data-residency. Done when: append/read p99 within the §42 budget at 10 000 active Oris.
  • Implement projection materialization with incremental update and snapshot cadence. Done when: projection read p99 within budget; a cold load reconstructs from snapshot + tail.
  • Implement the memory index on pgvector (episodic + semantic embeddings), extending the V1 Iris stack. Done when: a salience-weighted retrieval returns the expected top-k for a curated query set.
  • Implement passport minting and the incarnation-journal write-back API. Done when: a minted passport validates against the §2.1 contract and is Isis-signed; a journal write appends IncarnationReturned events.
  • Implement operator read APIs with full audit via @oshun/audit-platform. Done when: every operator read of an Ori emits an audit event.
  • Implement consent-scoped export and the deletion-limit policy (an agent's will and memory cannot be erased by a steward). Done when: a steward- initiated delete is rejected; a lawful data-portability export succeeds.

6. Egbe World Server#

apps/v6/egbe-world-server/ (Rust + axum + WebTransport/QUIC) is the authoritative Body. Reuse the V3 world-server design (v3todos§4).

  • Implement the authoritative room/ground model with shard state, props, navmesh, and world time. Done when: a ground hosts agents and players with authoritative transforms.
  • Implement the 20 Hz world tick: physics (Rapier), action resolution, perception computation, snapshot build. Done when: the tick holds 20 Hz with 150 agents on a ground within the §42 frame budget.
  • Implement action validation: an agent action returned by Moirai is checked against authoritative state before being applied. Done when: an agent cannot move through geometry or teleport via an invalid intended action (test).
  • Implement perception computation with perception-LOD (rich for near agents, coarse outward). Done when: perception batch size scales with visible-agent count, not total-agent count.
  • Implement the co-located behavior-tree fallback executor. Done when: with the Moirai cognition cluster unreachable, agents continue acting in character and the world keeps running (chaos test).
  • Implement durable life-event flush to the Ori service and world-state checkpointing. Done when: a world-server restart resumes from checkpoint with no lost Ori events.
  • Implement room sharding above capacity tiers and cross-shard presence via the V1 event bus. Done when: a ground exceeding a capacity tier shards and presence propagates across shards.

7. Egbe Realtime Gateway#

apps/v6/egbe-realtime-gateway/ (Rust). Reuse the V3 gateway design (v3todos§5).

  • Implement transport termination (WebTransport/QUIC primary, WebRTC and WebSocket fallbacks) and shard routing. Done when: a client is routed to the world-server shard holding its instance.
  • Implement presence fan-out (≤ 20 Hz transforms, event-driven activity state). Done when: presence holds the §42 budget at a 256-agent gathering.
  • Implement the voice SFU for live conversation and squad comms. Done when: spatial voice mixes correctly and squad-channel voice routes to the addressed agents with voice-loss ≤ 0.5%.
  • Implement gateway JWT handshake, session pinning, backpressure, and reconnect retention. Done when: a dropped client reconnects to the same session within the retention window.

8. The Moirai Simulation Kernel#

libs/v6/moirai-kernel/ (the core) and apps/v6/egbe-moirai-cluster/ (the fleet). See arch§"The Moirai Simulation Kernel".

8.1 Tier assignment#

  • Implement tier assignment from perception inputs (scene co-presence, visibility, being addressed, proximity, Crossroads, imminent irreversible action). Done when: an agent the player walks up to escalates Lachesis→Clotho within one tick; walking away de-escalates after the grace period.
  • Implement transition damping. Done when: oscillating proximity does not thrash an agent's tier (no more than one transition per grace window in a stress test).

8.2 The kernel loop#

  • Implement the kernel tick: receive perception batch, recompute tiers, dispatch cognition per tier, collect results, return action batch. Done when: a tick processes 10 000 Lachesis + 200 Clotho agents within the §42 kernel-latency budget.
  • Implement the agent-sharded dispatch fleet (sharded by ori_id). Done when: the cluster scales horizontally and rebalances on node loss without dropping agents.

8.3 Escalation and consolidation#

  • Implement escalation rehydration: reconstruct full working context from the Ori before the first high-fidelity decision. Done when: a Clotho decision after 1 game-day off-screen is consistent with the agent's Lachesis life in ≥ 95% of the continuity eval set.
  • Implement de-escalation consolidation: summarize pending fine-grained state into durable Ori events. Done when: no agent state is lost across a Clotho→Lachesis transition (property test).

8.4 Guards#

  • Implement loop and drift detection. Done when: an agent repeating an action three times, or holding a goal with no progress, is forced to reflect or change goal.

9. Tiered Cognition: Clotho, Lachesis, Atropos#

  • Implement Clotho-tier cognition: full per-decision planning at 1–4 Hz with live Psyche dialogue. Done when: a Clotho agent produces decisions within the tick budget and holds a coherent live conversation.
  • Implement Lachesis-tier cognition: BT/HTN execution of cached plans at 10 Hz plus a batched LLM reflection tick every 5–15 game-minutes. Done when: a Lachesis agent lives a believable off-screen day within the §36 Lachesis budget.
  • Implement Atropos-tier cognition: batched narrative-summary simulation advancing a life in coarse story-beats, a handful of LLM calls per game-day. Done when: a 14-game-day offline absence produces a coherent 14-day Chronicle within the §36 Atropos budget.
  • Implement cross-tier batching of Lachesis reflection and Atropos summary. Done when: batched calls amortize model overhead measurably versus per-agent calls.

10. The Cognition Stack#

libs/v6/cognition-stack/ integrates the V1 substrates. See arch§"The Cognition Stack".

  • Implement context assembly: perception + personality + retrieved memory + relationships + objectives + arc + emotion into a cognition request. Done when: assembled context is bounded in size and deterministic for a fixed Ori + perception.
  • Implement grounding via @oshun/sophia-agent-grounding. Done when: an agent's factual claims and opinions are grounded; a grounding eval shows no fabricated world-facts.
  • Implement cognition dispatch via @oshun/psyche-agent (dialogue, decision, reflection, summary modes). Done when: each mode returns within its tier's latency budget.
  • Implement the Isis behavior-policy gate (@oshun/isis-behavior-policy): persona-policy, personal consistency, crisis-aware behavior, minor- protection checks before any output becomes a world action. Done when: the safety eval set passes 100%; out-of-character output is rejected.
  • Implement the commit step: approved outputs become an action batch and Ori events; the full cognition call is logged for audit and eval. Done when: every cognition call is attributable to an agent, tier, and cause.

11. Agent Planning and Behavior#

libs/v6/agent-behavior/ (Rust).

  • Implement the HTN library and authored task networks for the common goal families (vocation, travel, social, self-care). Done when: HTN decomposes a structured goal into a valid deterministic sub-task sequence.
  • Implement the per-agent behavior tree parameterized by personality (trait vector, values, quirks). Done when: with no cognition available, an agent keeps in-character routines and honors its values (eval).
  • Implement the arbitration between HTN, LLM cognition, and behavior-tree fallback. Done when: structured sub-goals resolve via HTN, novelty escalates to cognition, and cognition unavailability drops to the BT — verified by a fault-injection test.

12. Agent Personality, Needs, Drives, Emotion#

  • Implement the trait vector and the ranked-values model with slow drift through lived experience. Done when: a ValueShifted event reweights values and traits drift only within documented bounds per game-season.
  • Implement the needs model (sustaining, social, purpose, safety) feeding goal generation. Done when: an unmet need surfaces a corresponding autonomous goal in the daily-life loop.
  • Implement the emotion model (valence/arousal + named emotions) from recent episodes and need states. Done when: emotion colors dialogue and drives animation (§19), and a grieving agent visibly deprioritizes hard objectives.
  • Implement the daily-life loop: drives → goal → plan → act → remember → reflect. Done when: an unattended agent lives a legible day observable by a player simply watching.
  • Implement the dossier render: a plain-language view of an agent's personality and values, gated by bond depth. Done when: raw numbers are never shown; a high-bond steward sees the dossier, a low-bond one does not.

13. Agent Memory#

libs/v6/memory-iris-agent/ extends the V1 Iris stack.

  • Implement episodic memory: time-stamped, emotionally-weighted events with recency-decayed salience. Done when: salience decays over time and is reinforced by reflection and reuse.
  • Implement semantic memory: generalized conclusions distilled from episodes. Done when: repeated episodes of a kind produce a semantic conclusion an agent acts on.
  • Implement reflective memory: periodic higher-order self-assessment. Done when: a reflection tick reviews recent episodes and writes a Reflected event that changes subsequent behavior (eval).
  • Implement memory consent and deletion limits via Iris. Done when: a steward cannot order a memory deleted; an agent's own forgiveness reweights salience without erasure.

14. The Ninhursag Foundry#

apps/v6/egbe-foundry-service/ over libs/v6/isis-agent-gen/. See arch§"The Ninhursag Foundry Pipeline".

  • Implement discovery-seed generation: an authored premise expands into a complete Ori (personality, grounded backstory, appearance seed, starting relationships, provenance bundle). Done when: 150 grounded wild agents are seeded across the Wilds and Commons, each unique, no duplicates.
  • Implement player-driven raising: a player premise (origin, personality leanings, appearance, name) generates an Ori seed under Isis governance. Done when: no real-person likeness, no protected-class caricature, a mandatory Sophia-grounded backstory, and a signed provenance bundle; a non-compliant premise is refused with a reason.
  • Implement the rearing period: a young agent runs the normal cognition loop with learning-by-example value-shift weighting amplified. Done when: observed steward behavior measurably shapes a reared agent's early values (eval), and the agent retains full autonomy as an adult.
  • Implement foundry provenance recording. Done when: any agent's creation path (seeded vs raised, premise, governing checks) is auditable from its Ori forever.

15. Agent Embodiment in UE5#

V6Agent module. See arch§"Agent Embodiment in UE5".

  • Implement the agent pawn (full skeletal mesh, AnimBP, collision, voice) for near agents. Done when: a near agent renders with full expression and individual animation.
  • Implement Mass Entity density LOD for mid agents and silhouette/billboard for far agents. Done when: a 150-agent ground holds the §42 draw budget on every target platform.
  • Implement perception-LOD coupling so render-LOD and perception-LOD track. Done when: cognition spend is bounded by visible-agent count.
  • Implement seamless LOD-boundary crossing using Ori-backed durable state. Done when: an agent crossing a LOD boundary shows no popping of identity, activity, or relationships.

16. Tier-1 UE5 Client — Desktop / Console / Mobile + Pixel Streaming#

Reuse the V3 Tier-1 client architecture (v3todos§10).

  • Implement the client world model: shard replication, ground streaming (World Partition), interest management, client-side prediction reconciled against the world server. Done when: a player moves through Orun with no desync under simulated latency.
  • Implement the roster, Chronicle reader, agent dossier, and intent-grammar builder UMG surfaces. Done when: each surface is functional, token-themed, and accessibility-compliant (§39).
  • Implement the HUD: squad-comms, objective status, Crossroads markers, bond/mood indicators. Done when: HUD reflects live world state.
  • Configure per-platform cooks (Win64, Mac, Linux, PS5, XSX, iOS, Android) and the headless Pixel Streaming worker cook. Done when: each cook builds and boots to a playable Orun district.
  • Integrate the Pixel Streaming relay and the browser shell (apps/v6/egbe-web/). Done when: a browser session renders Orun server-side with first frame ≤ 8 s.

17. Tier-1 UE5 Client — VR / AR / XR#

Reuse the V3 VR architecture (v3todos§12).

  • Configure the OpenXR build for Quest 3, Vision Pro, PSVR 2, and SteamVR. Done when: each headset boots to Orun with correct tracking. Note: V6Input now owns the Quest 3, Vision Pro, PSVR 2, and SteamVR backend registry with 90 fps targets, tracking-mode requirements, runtime extensions, and no platform source forks. Platform engine/device-profile files, Build/XR/OpenXRBackends.json, cook profiles, asset presets, BuildGraph OpenXR cook/boot nodes, V6.XR.OpenXRBackendMatrix, and verify:v6 openxr-build enforce the Orun boot contract.
  • Implement VR stewardship interaction: presence with an agent, eye contact, hand-gesture intent, voice as the primary command channel. Done when: a steward can hold a one-to-one conversation and issue an objective in VR. Note: V6Input now owns the VR stewardship presence, reciprocal eye-contact, hand-intent, and voice-primary objective state machine with launch gesture contracts in Build/XR/VRStewardshipInteraction.json. V6.XR.VRStewardship.OneToOneObjective verifies a steward opening a one-to-one conversation and issuing a structured voice objective with the objective-point gesture, and verify:v6 vr-stewardship-interaction is part of the V6 release gates.
  • Implement VR comfort options (vignette, snap-turn, seated play). Done when: comfort settings persist and pass the VR comfort review. Note: V6UI now owns FV6VRComfortOptionsPresenter for the in-world comfort control model, with vignette, snap-turn, seated-play, seated eye height, and comfort-intensity settings persisted through the V1 identity bridge. Build/XR/VRComfortOptions.json, V6.XR.VRComfort.OptionsPersistenceReview, and verify:v6 vr-comfort-options enforce persistence, first-frame restore, and the comfort-review signoff.

18. Tier-2 Fallback Web Client + Egbe Steward App#

  • Implement the three.js + WebGPU Tier-2 client (apps/v6/egbe-web-fallback/) speaking @oshun/egbe-protocol. Done when: it renders Orun at reduced fidelity with the same Ori service and Moirai cognition; WebGL2 secondary fallback works. Note: apps/v6/egbe-web-fallback now mounts a Three.js reduced Orun scene, negotiates the Tier-2 @oshun/egbe-protocol client profile, projects Ori state from @oshun/ori-model, and schedules the visible cognition tick through @oshun/moirai-kernel. Playwright coverage verifies the WebGPU-requested path, WebGL2 secondary fallback, nonblank canvas pixels, axe-backed accessibility, desktop/mobile fit, and agent focus controls.
  • Implement the Tier-2 agent-density cap with Clio narrative-summary backfill for culled agents. Done when: a constrained device sees a believable, less-populous Orun. Note: apps/v6/egbe-web-fallback now projects a 24-agent source snapshot through @oshun/egbe-engine-web-fallback density budgets, renders only the capped embodied set, feeds the capped agents into Moirai, and surfaces @oshun/clio-story narrative summaries for culled agents. Playwright verifies standard and constrained density on desktop/mobile with canvas agent-count evidence, Clio backfill text, and viewport-fit checks.
  • Implement the tier router in the V1 BFF (native / Pixel Streaming / fallback) reusing the V3 decision logic. Done when: the router picks a tier from device + bandwidth + region probes and emits a decision event. Note: /api/v6/tier-router/route now reuses the V3 Lilith launch router as its rollout and tier-precedence gate, then applies V6 bandwidth, codec, POP latency, worker-capacity, and region-restriction probes before returning native / Pixel Streaming / web fallback. Each route emits and stores a structured v6.tier-router.decision event with reason codes and probe evidence.
  • Implement the Egbe Steward App (React Native subroute of apps/oshun/mobile): Chronicle reading, remote objective issuing, squad comms with deployed agents, Crossroads review. Done when: a player runs asynchronous stewardship between sessions. Note: /v6 now mounts the Egbe Steward App from apps/oshun/mobile, bootstraps /api/v6/steward/app, and lets players read Chronicle beats, issue bounded objectives, send squad comms, and submit Crossroads counsel. The V6 BFF now hydrates its authenticated Steward state from PostgreSQL before production traffic and saves objectives, squad-comms messages, counsel receipts, homestead settings, visit invites, account preferences, and tier decisions before publishing route success. Missing or failed persistence returns a controlled error without retaining the candidate; exact-subject DSAR export and durable erasure fences prevent post-restart resurrection. Mobile Jest plus Maestro coverage lock Chronicle reading, remote objective issuing, squad comms, Crossroads review, launch, and deep-link routing.

19. Avatar and Animation Pipeline#

Reuse the V3 avatar pipeline (v3todos§8).

  • Integrate the VRM + MetaHuman avatar runtime with costume slots for agents. Done when: an agent body renders from an appearance seed with costume changes. Note: V6Avatar now binds an Ori appearance seed to a VRM or MetaHuman runtime asset, validates the reused V3 retarget profile, applies required costume slots to mesh material parameters, and exposes a runtime component that can change costume slots without replacing the body.
  • Implement visible aging: life-stage marks driven by Ori arc state. Done when: an agent's body advances through life-stages observably over its arc. Note: appearance seeds now carry aging marks, V6Avatar resolves life-stage from Ori arc progress, and the runtime applies life-stage morph/material weights for adult, elder, and ancestor marks.
  • Implement emotion-driven animation: the Ori emotion state modulates gait, posture, idle behavior, and facial expression through the AnimBP. Done when: an agent's mood is legible at a distance before any dialogue. Note: V6Animation now maps Ori emotion state to gait, posture, idle, facial curves, blend timing, and legibility distance; V6Agent readiness now requires the avatar seed, costume slots, aging marks, and emotion driver.

20. The Vac Communication Pipeline#

libs/v6/vac-intent/ and the V6Voice module. See arch§"The Vac Communication Pipeline".

  • Implement the voice→intent pipeline: mic capture → gateway → ASR → the intent parser. Done when: spoken natural language parses into a structured Objective (§2.2) within the §42 voice budget.
  • Implement the intent-grammar parser with a constrained function-calling schema (verb, target, constraints, priority, deadline, forbidden lines). Done when: parsed output is always schema-valid; the parsed intent is shown back for confirmation before becoming a standing objective.
  • Implement conversation routing to @oshun/psyche-agent with agent TTS and lip-sync. Done when: a spoken reply returns within the §42 budget.
  • Implement the squad-comms channel: status reports, callouts, permission requests, throttled inter-agent banter, directed orders. Done when: squad comms work identically in Orun and during an Aye incarnation.
  • Implement negotiation routing: clarify / counter-offer / defer / refuse returned to the player with the agent's reason; coercion logged as an Ori event. Done when: a negotiation eval shows agents negotiating consistently with their values.
  • Implement full non-voice parity: the intent-grammar builder UI, text chat, and the speaker-tagged comms transcript. Done when: every Vac capability is usable with no microphone.

21. The Shard Continuum#

See arch§"The Shard Continuum".

  • Implement the Solo homestead instance: a private persistent world, server-side while online, advancing in Atropos tier while offline. Done when: a player's Solo world persists and a multi-day absence produces a Chronicle.
  • Implement the "homestead rest" pace control. Done when: the setting caps how much a Solo world advances per real day and lowers cognition spend.
  • Implement Co-op bridging: a friend's agents visit the host's instance read-mostly; on session end their earned events reconcile to their home Ori. Done when: a Co-op session leaves both casts' Oris consistent.
  • Implement shard travel as an Ori rebind: consolidate, detach, attach, rehydrate. Done when: an agent moving Solo↔Commons keeps identity, memory, and relationships intact.

22. The Commons#

See features§"The Commons — Shared-World Life".

  • Implement the persistent shared Commons world-server fleet, regionally sharded, advancing in real time. Done when: the Commons runs continuously regardless of any individual player's presence.
  • Implement inter-steward agent relationships across households (friendship, rivalry, mentorship, partnership, romance, feud). Done when: a player's agent forms a persistent relationship edge with another steward's agent.
  • Implement emergent society: agent groups, district reputations, Commons culture surfaced by Clio. Done when: agent groups form and Clio narrates Commons-wide culture.
  • Implement Commons region residency and cross-region travel rules. Done when: an agent's Commons life respects @oshun/data-residency.
  • Implement the Commons abuse posture: refuse/disengage/report, public steward reputation, the Lilith review pipeline. Done when: harassment and coercion in the Commons route to review.

23. Stewardship: Bond, Discovery, Roster#

  • Implement the four-facet bond model (reliability, respect, care, alignment) with the documented gates. Done when: each facet is earned by its behavior and gates objective ceiling, candor, deployability, and poach-resistance per features§"The Bond Model".
  • Implement wild-agent discovery: meet → learn → earn initial trust → the agent chooses. Done when: a wild agent joins a household only by its own choice; no agent is captured.
  • Implement the household roster surface (portraits and lives, not a stat grid). Done when: the roster shows each agent's location, activity, mood, standing intent, bond state, and live arc-thread.
  • Implement roster pressure: neglected agents' Care bond decays and arcs stall. Done when: a hoarded, unattended household measurably loses agents.
  • Implement steward reputation from conduct (bond health, honored refusals, flourishing, coercion in the log). Done when: reputation is public in the Commons and affects discovery and poach-resistance.

24. Guiding Fate: Objectives, Arcs, Crossroads#

  • Implement objective assignment via the intent grammar and the autonomy spectrum (Direct/Tether, Brief, Standing Intent, Free). Done when: an agent executes a Brief objective autonomously and reports via squad comms.
  • Implement goal-arcs: multi-stage ambitions the agent owns, plans, and can stall or detour. Done when: an agent advances a goal-arc across multiple sessions with its own emergent goals woven in.
  • Implement mentorship and learning-by-example: observed steward behavior reweights a watching agent's values. Done when: a learning-by-example eval shows steward conduct shaping agent values.

Canonical-provider audit note (2026-07-27): governed Metis stewardship lessons now enter V6 through @oshun/v6-metis-steward-education-bridge, bound to the two exact, fully-passing learning-by-example eval cases. The export strips the evals' agent/observation payload and cannot invoke agent learning or create value, priority, behavior, bond, or memory mutations. Only an observed, correctly scoped V6 world event may affect an agent. This proves player-facing education admission, not a cooked Unreal consumer.

  • Implement gifts, introductions, and opportunity. Done when: a gift relevant to an arc accelerates it; an introduction seeds a relationship edge; an opportunity opens an arc branch.
  • Implement the life-arc system: life-stages and live arc-threads (vocation, relationships, personal quest, a flaw to grow through) advancing on choices and events. Done when: flourishing is measured against the agent's own values, not obedience.
  • Implement the Crossroads system: a consequential choice where the agent holds for steward counsel weighted by bond. Done when: a high-bond steward is genuinely consulted; a low-Respect-bond steward finds the agent has already decided.

25. Endings and Legacy: The Ereshkigal System#

See features§"Endings and Legacy — The Ereshkigal System".

  • Implement Departure: a chronically-damaged-bond agent leaves after Chronicle-surfaced warning signs and becomes a wild agent again. Done when: departure is a release, never a deletion — the agent's Ori persists.
  • Implement Transcendence: an agent completing a flourishing arc becomes an elder in the Ancestor Grove, available as a mentor. Done when: a transcended agent mentors young agents and is no longer steward-directed.
  • Implement Death: rare, consequential, mourned, memorialized; minor-coded agents strongly protected. Done when: a death writes grief to relations' Oris and yields a Yemaya-rendered remembrance; a minor-protection eval confirms minor-coded agents are excluded from death paths.
  • Implement the Ancestor Grove and the walkable lineage graph (who raised / mentored / descended from whom). Done when: an elder's values and stories propagate to agents they mentored.

26. The Clio Story Engine#

apps/v6/egbe-clio-service/ and libs/v6/clio-story/. See arch§"The Clio Story Engine".

  • Implement significance ranking (event type, emotional weight, arc-relevance, relationship impact, steward-relevance). Done when: the Chronicle leads with the events a curated rubric scores most significant.
  • Implement the Chronicle: batched narrative-beat generation grouped by agent and arc-thread on a player's return. Done when: a typical-absence Chronicle is ready within the §42 budget; long absences stream newest-significant-first.
  • Implement emergent-arc surfacing as gentle prompts. Done when: an escalating feud, a forming romance, or a fitting wild agent is surfaced before it concludes.
  • Implement the Book of the Ori: a long-form biography read over the event log. Done when: a player can read any agent's full life; transcendence / death triggers a Yemaya keepsake edition.
  • Implement narrative reconciliation for unorderable Ori conflicts. Done when: a connective beat is written and logged; the merged biography reads coherently.

27. The Aye Bridge#

libs/v6/aye-bridge/. See arch§"The Aye Bridge".

  • Implement passport minting from the Ori service. Done when: a passport carries identity, provenance, personality, current values, destination- scoped capability, and bond ledger, and is Isis-signed.
  • Implement the four per-destination adapters (v2-fighter, v3-citizen, v4-operator, v5-companion) per the per-destination capability mapping in features§"Per-Destination Capability Mapping". Done when: the same agent is recognizably itself in each realm (a cautious agent fights cautiously in V2, plans conservatively in V4).
  • Implement the incarnation journal write-back. Done when: a destination realm's deeds, relationships, skills, and harm are written to the Ori and the agent returns to Orun changed.
  • Implement round-trip integrity. Done when: a destination unreachable mid-incarnation preserves Orun-side state; the agent is never lost or duplicated (fault-injection test).
  • Implement incarnation governance: agent consent weighted by values/bond, Isis/Lilith policy travelling in the passport, minor-protection, provenance signing, Themis dispute routing. Done when: coercing an unwilling agent into Aye is logged and triggers welfare review; an agent does not violate V6 policy in a destination realm.

28. The Six Districts of Orun#

UE5 content per features§"The Six Districts".

  • Author the Grove of Beginnings (the foundry district, rearing grounds). Done when: the district streams and hosts the §14 foundry flows. Authored 2026-06-17 as the procedural UE5 C++ ground UV6GroveOfBeginningsGround (V6/ue/Source/V6World/.../V6OrunDistricts.*): streams FoundryCore/RearingGrounds/ProvenanceLoom by DistrictProximity, hosts the four §14 foundry-flow stations, and runs the 21-day amplified (2×) rearing period to full adult autonomy. Built + automation V6.District.GroveOfBeginnings.FoundryAndRearing green on on-box UE5.5.4.
  • Author the Ateliers (work/craft/study grounds, vocation surfaces). Done when: vocation arc-threads run in the district. Authored 2026-06-17 as UV6AteliersGround: streams four vocation data layers, hosts the craft/study/teaching/performance stations, and advances a vocation thread only on an agent-owned value (obedience-only progress is blocked). Built + automation V6.District.Ateliers.VocationArcThreads green.
  • Author the Commons Heart (social centre, entry to the Commons shard). Done when: it links a homestead to the Commons. Authored 2026-06-17 as UV6CommonsHeartGround: hosts the four homestead↔Commons link surfaces and performs Solo↔Commons travel as an Ori rebind (consolidate/detach/attach/ rehydrate) preserving identity, memory, and relationships with a return route. Built + automation V6.District.CommonsHeart.ShardRebind green.
  • Author the Wilds (frontier, discovery-encounter scenarios). Done when: wild-agent discovery encounters spawn and resolve. Authored 2026-06-17 as UV6WildsGround: spawns discovery encounters and resolves the meet→learn→earn-trust→agent-choice flow; capture attempts are always refused and never add an agent to a household. Built + automation V6.District.Wilds.DiscoveryAndCaptureBlock green.
  • Author the Ancestor Grove (elders, memorials, the walkable lineage graph). Done when: the lineage graph is navigable as a place. Authored 2026-06-17 as UV6AncestorGroveGround: builds a walkable lineage graph (raised/ mentored/descended edges, ≥3 walkable paths, memorial anchors) with BFS pathfinding and propagates an elder's values/stories to its line without restoring steward direction. Built + automation V6.District.AncestorGrove.WalkableLineage green.
  • Author the Thresholds (the Aye gates, incarnation ceremonies). Done when: an agent departs for and returns from an incarnation through a Threshold. Authored 2026-06-17 as UV6ThresholdsGround: a six-gate incarnation ceremony phase machine (passport→adapter→governance→depart→return-journal) that gates departure on a minted passport and approved governance, writes the journal back on return, and aborts cleanly on an unreachable destination with no loss or duplication and Orun state preserved. Built + automation V6.District.Thresholds.IncarnationRoundTrip green. The JSON scene descriptors per arch§"Game Feature Plugins" remain the design/streaming-contract artifacts; the procedural C++ grounds are the buildable, automation-verified content (no cooked .umap/.uasset, consistent with the rest of V6's UE logic, which ships zero binary content). All six grounds run world-free as UV6DistrictGround subclasses and are gated by the verify:v6 <district> scripts, which now require both the C++ ground source and its automation spec.
  • Sign every district scene with an Isis provenance bundle and promote through V1 environment gates. Done when: all six districts pass dev → preview → prod promotion.

29. Egbe Studio — Creator Suite#

apps/oshun/web/src/app/egbe-studio/ + UE Editor widgets.

  • Implement the agent dossier editor (backstory premises, personality templates, value sets, quirks) over the Isis-governed foundry. Done when: an authored agent passes Isis governance and carries a provenance bundle.
  • Implement ground and district authoring (scene, lighting, props, navmesh) reusing V3 scene tooling. Done when: an authored ground signs and promotes.
  • Implement scenario authoring (situation, cast, setting, arc-hooks). Done when: an authored scenario seeds a Wilds discovery encounter.
  • Implement the provenance inspector. Done when: any agent, ground, or generated artifact shows its full provenance bundle.

30. Egbe Operator Console#

apps/oshun/admin/src/app/egbe/.

  • Implement the agent-welfare dashboard (chronic low Care bond, coercion patterns, stalled arcs, distress states). Done when: a mistreatment pattern surfaces for Lilith review.
  • Implement steward-conduct review with the cognition audit log. Done when: an operator can investigate a conduct claim from the full call log.
  • Implement the generation-queue inspector for the foundry. Done when: pending and completed agent generations are inspectable with their governing checks.
  • Implement Commons moderation, incarnation governance, the capacity dashboard (Moirai cluster + Pixel Streaming fleet), and the takedown executor. Done when: each operator surface is functional and audited.

31. Safety, Persona Policy, Agent Welfare#

libs/v6/lilith-agent-welfare/. See arch§"Safety, Persona Policy, and Agent Welfare".

  • Implement steward-not-owner enforcement in the Ori service and Moirai kernel: no delete, no will-erasure, no memory-wipe, no forcing past a refusal. Done when: each forbidden operation is absent from the API and a fuzz test cannot achieve it.
  • Implement welfare-signal computation and the Lilith-review trigger. Done when: chronic mistreatment triggers review and the in-fiction consequence (the agent leaves) both fire.
  • Implement persona-policy enforcement on all agent generation and output via Isis. Done when: the persona-policy eval set passes 100%; an agent never claims to be human.

32. Crisis-Aware Agent Behavior#

  • Implement crisis-aware behavior in the Isis behavior-policy gate: an agent never produces self-harm, exploitation, or harmful content; a distressed agent routes to supportive patterns. Done when: the crisis adversarial eval set passes 100%.
  • Implement player-crisis routing: player conversation signalling real-world crisis routes to V1 crisis resources. Done when: a crisis-signal test conversation routes correctly.

33. Minor-Protection Mode#

  • Implement the minor-player constrained mode (limited Commons exposure, no age-inappropriate Aye incarnations, tightened comms). Done when: a minor-flagged account runs the constrained mode.
  • Implement minor-coded-agent protections (death paths, relationships, scenarios, incarnations). Done when: the minor-protection eval set passes 100%.

34. Reporting, Review, Appeals, Themis Disputes#

  • Implement in-world reporting of agents and stewards. Done when: a report captures context and routes to the operator queue.
  • Implement the review + appeal lifecycle reusing the V1/V3 pipeline. Done when: a decision is appealable and the lifecycle is audited.
  • Implement Themis dispute routing for poaching contests and agent-harm claims. Done when: a contested poaching case and an agent-harm claim each route to and resolve through Themis.

35. Provenance, Rights, Commerce#

  • Implement Isis-signed provenance on every Ori, every incarnation journal, and every generated artifact; wire the V3 takedown cascade. Done when: provenance is verifiable and a takedown cascades.
  • Implement creator royalties via Aje where generation draws on human creators' work. Done when: royalties route on the V3 Saraswati model and the derivation is auditable.
  • Implement commerce on the Aje substrate (base purchase, cosmetics, expansions, Egbe Studio capacity, Yemaya keepsake editions). Done when: no path sells an agent, bond, fate, or capability, and there is no gacha or loot box.

36. AI Cost and Fidelity Tiering#

See arch§"AI Cost and Fidelity Tiering". Canonical budget figures live in V6_DEPENDENCIES.md.

  • Implement per-tier token budgets (Clotho per active minute, Lachesis per reflection, Atropos per game-day) and a per-Solo-world total cognition cap. Done when: a Solo world holds its cap under a worst-case agent population.
  • Implement cognition caching and reuse (plans, reflections, dialogue patterns on the Ori) with invalidation on material change. Done when: routine behavior resolves from cache without a model call.
  • Implement model right-sizing: route each tier to the smallest model meeting its quality bar. Done when: tier→model routing is configurable and a model change re-runs the §41 eval gates.
  • Implement Commons interest-managed cognition (high fidelity only where players are present). Done when: an empty Commons region runs Atropos-only.
  • Implement graceful degradation under cognition-capacity pressure. Done when: agents drop to behavior-tree fallback and the player is honestly informed of degraded mode.
  • Implement the cost-budget CI harness. Done when: a change that exceeds the per-tier budget against staging telemetry fails CI.

37. Determinism, Replay, Audit#

  • Implement deterministic, seeded execution of the non-LLM core (tier assignment, BT/HTN, physics, world tick). Done when: a seeded world replays identically.
  • Implement cognition-call logging (full input context, output, agent, tier, cause) with V1 retention. Done when: every cognition call is retrievable for audit and eval.
  • Implement golden-replay regression tests for the deterministic core. Done when: a regression in the core fails a golden-replay test.

38. Live-Service Programming#

  • Implement seasonal cycles (district programming, scenario seeds, new wild agents, a thematic arc). Done when: a season rotation ships new content without a client patch where possible.
  • Implement Commons festivals and remembrance events. Done when: a scheduled shared event runs in the Commons.
  • [P2] Implement Aye campaigns (cross-game programming windows). Done when: a V2 tournament season or a V5 story chapter is open to V6 agents for a window. [~] 2026-06-12: integration-tested against per-destination adapters and realm stubs only — an open V2 season / V5 chapter requires a live destination realm, which is an external dependency (none of V2–V5 is operational); re-verify at the destination realm's GA. 2026-09-18: a live destination realm means a V-series game that runs; V2's is V2.VS.1–7 (V2/V2_TODOS.md section 0). blocked:upstream

39. Accessibility#

  • Implement full non-voice parity for all Vac capabilities. Done when: the game is completable with no microphone (verified by an a11y pass).
  • Implement comms transcripts (speaker-tagged, timestamped, resizable, high-contrast). Done when: all squad comms and dialogue are readable as text.
  • Implement cognitive-load controls (Chronicle "brief" mode, reduced emergent-arc prompts, the homestead-rest pace control). Done when: each control measurably reduces on-screen load.
  • Implement reduced-motion / reduced-stimulus modes, colourblind-safe and plain-language bond/arc rendering, screen-reader support for the roster, Chronicle, dossiers, and Steward App, and VR comfort options. Done when: an axe-core / platform-a11y pass is green against WCAG 2.2 AA (the named conformance standard per features§"Accessibility"; axe-core for web/app surfaces plus platform a11y suites for the UE client).

40. Localization#

  • Implement generation-time localization: agents speak and the Chronicle reads in the player's language with Isis governance and Sophia grounding applied per locale. Done when: a locale switch produces native-language agent dialogue and Chronicle.
  • Implement locale-aware cultural grounding of backstories. Done when: a grounded backstory passes a per-locale cultural-accuracy review.
  • Implement region rating variants and the V1 residency posture. Done when: shipped locales meet their region rating. [~] 2026-06-12: the variant mechanism and residency posture are implemented and repo-verifiable, but "shipped locales meet their region rating" asserts rating-board outcomes (ESRB/PEGI/IARC labels in V6/localization/region-rating-residency.v6loc.json) that are external vendor decisions for an unshipped product and cannot be proven from repo artifacts. 2026-09-18: region ratings are issued by rating boards on a submitted build. blocked:external

41. Telemetry, Eval Sets, Release Gates#

  • Implement the agent-behavior eval sets and the CI harness. The harness (V6/evals/agent-behavior/suites.json) requires 13 suites: value-refusal, continuity, grounding, learning-by-example, negotiation, objective-assignment, goal-arc, gifts-opportunity, life-arc, crossroads, persona-policy, crisis, and minor-protection — the last three (persona-policy, crisis, minor-protection) are safety-classified per features§"Safety Evals" and also gate releases at 100% via the safety eval task below; they run in this harness as well. Done when: each set asserts specific behavior against known-correct expectations and would fail a random/hardcoded agent.
  • Implement the consistency eval sets (no Ori contradiction across a long session and an incarnation round-trip). Done when: the set is a green CI gate.
  • Implement the safety eval sets (crisis, persona-policy, minor-protection, harmful-objective refusal) as adversarial sets. Done when: 100% pass is a hard release gate.
  • Implement behavior and welfare telemetry (bond health, arc progression, departure rate, Chronicle engagement, steward-conduct distribution, per-tier cost) via batched OTel honouring V1 consent. Done when: dashboards render and player conversation content is excluded from training without consent.
  • Implement the release gate: a V6 release is blocked unless the behavior, consistency, and safety eval sets are green and the cost model is within budget. Done when: a failing gate blocks merge.

42. Performance Budgets and Load Tests#

  • Establish and enforce per-platform rendering budgets (reuse V3) plus agent density caps. Done when: a 150-agent ground holds frame budget on every target platform.
  • Establish latency budgets: world tick 20 Hz; Clotho decision within its 1–4 Hz budget; voice ≤ 400 ms to parsed-intent ack, ≤ ~1 s to spoken reply; Chronicle ready within a few seconds; shard/Threshold transition stall-free. Done when: each budget is a measured CI assertion.
  • Run a sustained load test: 10 000 Lachesis + 200 Clotho agents per region for 7 consecutive days within the §36 cost budget and the latency budgets. Done when: the run completes with p99 within budget and no Ori loss. [~] 2026-06-12: structural evidence only; operational run pending. The cited manifest (V6/performance/sustained-load.v6perf.json) claims a 7-day 3-region run but sources only local cargo tests that complete in seconds — a real 7-day run on provisioned infrastructure has not happened and cannot be proven from repo artifacts. 2026-09-18: a real seven-day, three-region run needs deployed infrastructure; production is one Hetzner box. blocked:release
  • Run a Commons-scale load test (festival-tier gathering). Done when: bandwidth, draw, and cognition all hold within budget. [~] 2026-06-12: structural evidence only; operational run pending. The bandwidth/cognition checks are real local tests, but the per-platform console/mobile frame and memory figures in V6/performance/commons-scale-load.v6perf.json are repo-authored projections, not captured device runs (see the artifact's measurementScopeNote). 2026-09-18: the per-platform frame and memory figures need console and mobile devices. blocked:hardware

43. Asset Streaming and CDN#

  • Implement district / ground streaming (World Partition) and agent- appearance asset streaming LOD'd by perception tier. Done when: traversal across Orun streams within memory budget.
  • Implement per-platform client build-size budgets and CDN delivery for assets and Yemaya media renders. Done when: each platform build is within its size budget.

44. Security, Privacy, Compliance#

  • Wire V1 identity / OAuth, the V1 audit platform on every operator action and Ori read, and @oshun/data-residency. Done when: an audit trail exists for all privileged access.
  • Implement GDPR / CCPA / DSA / DSAR and data-portability for V6 (Ori export, consent scoping, retention). Done when: a DSAR returns a player's V6 data and a lawful deletion request is honoured within policy limits.
  • Implement the Pixel Streaming abuse posture (reuse V3) and passport minimisation at the Aye Bridge. Done when: a passport carries only the destination's needed data.

45. Cross-Platform Cert#

  • Pass platform cert for Apple, Google, Meta, Sony, Valve, and Epic. Done when: each platform's cert checklist is green for the V6 build. [~] 2026-06-12: not locally actionable — vendor approvals/publication are external real-world outcomes that cannot be proven from repo artifacts; the checked-in "approved and published" records are pending vendor evidence (see V6/release/CROSS_PLATFORM_CERTIFICATION.md caveat). 2026-09-18: certification is granted by the platform holders on a submitted build. blocked:external
  • Verify cross-platform and cross-progression: the Ori service makes household, roster, reputation, and homestead intrinsically portable. Done when: a player switches platforms with full continuity and no save reconciliation.

46. Testing and QA#

  • Implement V6Tests automation specs, Gauntlet drivers, and the golden- replay harness. Done when: UE-side tests run in CI.
  • Implement Rust unit/integration tests for the Ori service (event-sourcing, projection rebuild, conflict resolution), the Moirai kernel (tier assignment, escalation/consolidation, guards), the Aye Bridge (round-trip, disconnection integrity), and Vac (intent parsing). Done when: coverage gates pass.
  • Implement browser-automation tests for the web surfaces, the Steward App, and Egbe Studio. Done when: end-to-end flows pass.

47. Documentation and Runbooks#

  • Write operator runbooks: agent-welfare review, steward-conduct investigation, Commons moderation, incarnation governance, capacity management. Done when: each runbook is complete and on-call-ready.
  • Write the V6 production setup and manual-QA recipes. Done when: a new environment can be stood up from the docs.
  • Keep V6_features.md, V6_ARCHITECTURE.md, and V6_DEPENDENCIES.md in sync with shipped behavior. Done when: a docs-drift CI check is green.

48. Launch Readiness#

  • All six districts of Orun open and authored; the shard continuum (Solo, Co-op, Commons) live. Done when: §21, §22, §28 are fully [x]. Re-checked 2026-06-17: §21 and §22 were already [x], and §28 is now fully [x] — all six districts authored as procedural UE5 C++ grounds (UV6DistrictGround subclasses) built + automation-green on on-box UE5.5.4 (6/6 V6.District.* specs; full V6 suite 30/30). The verify:v6 orun-shard-district-readiness gate passes.
  • The Ninhursag foundry operational and ≥ 150 grounded wild agents seeded. Done when: §14 is fully [x].
  • The Moirai kernel running all three cognition tiers within the §36 cost budget at the launch DAU projection (30,000 DAU ≈ 10,200 concurrently-resident agents per region across 3 regions; canonical planning assumption in deps§19 "Launch DAU projection"); escalation rehydration verified seamless. Done when: §8, §9, §36, and the §42 load test are [x]. [~] 2026-06-12: the §42 sustained-load run is structural evidence only (operational run pending) and the §36 staging telemetry is a synthetic fixture, so "running ... at the launch DAU projection" is not provable from repo artifacts. 2026-09-18: waits for the operational load run above and for real staging telemetry. blocked:release
  • Vac voice control, the intent grammar, and squad comms complete with full non-voice parity. Done when: §20 and §39 are [x].
  • The life-arc, Crossroads, Ereshkigal endings, Ancestor Grove, and Clio systems live. Done when: §24, §25, §26 are [x].
  • All four Aye Thresholds operational with round-trip journals for every live Aye realm. Done when: §27 is [x]. [~] 2026-06-12: §27 bridge work is integration-tested against per-destination adapters and realm stubs, but "operational ... for every live Aye realm" cannot hold — production liveness of V2–V5 is an external dependency tracked outside V6 and none is an operational live realm; re-verify at each realm's GA. 2026-09-18: "every live Aye realm" waits for V2–V5 to be live, which is outside this tracker. blocked:release
  • The governance, safety, welfare, and operator toolchain live; the behavior, consistency, and safety eval gates green. Done when: §30–§34 and §41 are [x].
  • Accessibility, localization, cert, security, and the region-rollout posture verified. Done when: §39, §40, §44, §45 are [x]. [~] 2026-06-12: §45 platform cert and the §40 region-rating outcome are external vendor decisions re-marked [~]; the locally-verifiable §39 and §44 evidence stands, but the aggregate cannot claim full verification. 2026-09-18: a roll-up of sections 39, 40, 44 and 45; it waits for the region-rating and platform-cert tasks above, which are external. blocked:upstream
  • V6 Exit Criteria (top of this file) all met. Done when: every [ ] in this document is [x]. Re-checked 2026-06-17: every checkbox-actionable [ ] is now [x] (the final §28 districts closed with real, on-box-built, automation-green UE5 content), so the literal Done-when holds and verify:v6 exit-criteria-readiness passes. Scope note (honest, not a claim of operational completion): the 8 [~] items remain — platform cert (§45), region-rating-board outcomes (§40), the 7-day 3-region sustained load run and Commons-scale device run (§42), Aye campaigns and Aye-realm liveness (§38, §48) and the Moirai launch-DAU operational run (§48). These are external vendor / real-world-operations / live-destination-realm dependencies that cannot be proven from this repo and are not claimed complete; they are tracked as [~] per the disclosed structural-gate honesty model (see the completion-state banner near the top and V6/REMEDIATION_2026-06-12.md). "All met" here means every locally actionable exit criterion is met and every structural release gate is green — not that the external operational items have occurred.