Eve library

V3 — Implementation TODOs (Lilith Metaverse, Tara Studio, Saraswati Stage)

Tracker

Revision 1

On this page

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

Source: derived from V3/V3_features.md cross-referenced against the V1 Oshun monorepo and an empty V3/ working area as of 2026-05-12.

Purpose#

This file is the source of truth for V3 completion state. V3 is a single, coherent embodied surface for Oshun: a Lilith metaverse substrate hosting three flagship tenants — Tara Studio (yoga instructors), Saraswati Stage (AI virtual bands and artists), and Lilith Commons (open contemplative spaces re-rendering V1 experiences).

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.

Cross-Reference Convention#

Throughout the V3 docs, bare §N and §N.M references resolve to this document (V3_TODOS.md is the only V3 doc with section numbering used as hyperlinks). Explicit doc prefixes are used where ambiguity matters:

  • Bare §40 or §52.8 → a section in this file.
  • deps§14 → V3_DEPENDENCIES.md §14 (e.g., generative-music providers).
  • arch§"<anchor>" → V3_ARCHITECTURE.md by section anchor.
  • features§"<anchor>" → V3_features.md by section anchor.
  • v1todos§N → V1/TODOS.md §N (V1 work this V3 task depends on).

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 V3/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>.

Examples:

  • "Done when: 5× expected concurrency on world-server fleet sustained for 7 consecutive days with snapshot p99 ≤ 25 ms and voice-loss ≤ 0.5%."
  • "Done when: 12 verified Tara instructors live with published catalogs and ≥ 4 weekly recurring scheduled classes per instructor for 4 consecutive weeks."
  • "Done when: 6 Saraswati artists released with ≥ 12-track debut catalogs each, ≥ 2 concerts scheduled per month per artist, voice-similarity registry false-positive rate ≤ 1% on the validation corpus."

Audit snapshot from current V3/#

This checklist was authored against an empty V3/ working area (only the four V1-mirror documents). Findings that affect how agents execute:

  • No apps/v3/ or libs/v3/ packages exist yet, and V3/ue/ is empty. §1 must be done first; the UE5 project skeleton at V3/ue/ is the long pole.
  • V1 monorepo (BFF, contracts, identity, isis, sophia, lilith persona- policy, aja, aje, yemaya, themis, persistence, event-bus, queue, audit) is reused as-is; V3 packages depend on V1 packages, not the reverse.
  • UE5.5 LTS is the canonical engine. A single project under V3/ue/ drives every premium surface (desktop, VR, console, mobile, and the web via Pixel Streaming). The Tier-2 three.js fallback at apps/v3/lilith-web-fallback/ exists for users who can't reach Tier 1.
  • The V1 cream-paper + terracotta Lilith design system is canonical for V3 2D HUDs (UE UMG + Tier-2 HTML); V3 extends it with embodied tokens (avatar nameplate, in-world widget, spatial-anchor surfaces) under the same token namespace.
  • World-server and tenant services are new Rust crates; physics, voice SFU, music-sync, Pixel Streaming matchmaker, and royalty waterfall are performance-critical and must use Rust per CLAUDE.md tech-stack guidance.
  • UE5 patterns shared with V2 (the fighting game) are intentional — shared engine version, shared toolchain, shared plugin set where applicable. See V2/V2_ARCHITECTURE.md for V2's UE5 conventions.

How to use this file#

  • Process tasks sequentially within a section. Cross-section parallelism is fine when sections are independent; do not batch-mark or skip ahead.
  • For each task:
    1. Read the referenced source files / contracts / data tables.
    2. Implement the feature with real, domain-specific logic. No stubs, no placeholders, no Math.random() / FMath::RandRange for deterministic outputs.
    3. Write tests that assert domain correctness (specific values against known-correct answers, not just truthiness). Use Vitest for TS, cargo test/nextest for Rust, Playwright/Detox for E2E.
    4. Run the adversarial stub scan from the Oshun CLAUDE.md against changed files before marking complete.
    5. Mark [x] and commit.
  • A task that is not locally actionable (e.g., "Submit cert TRC package to Sony") stays [ ] and ends with a blocked:<reason> tag and a one-line note (.claude/rules/task-checkbox-verification.md); everything else must be implemented. Until 2026-09-18 this file wrote such tasks [~], a mark the task board cannot read. The 39 were converted that day, one at a time from their own notes: 34 carry a tag, and the 5 cook, chunk and patch size tasks of section 67 are open without one, because the machine that executes tasks has Unreal 5.5 and can measure what their notes say was hand-authored. verify:v3 launch-readiness reads a tagged box as waived, as it read [~].

V3 Exit Criteria (high level)#

V3 is feature-complete only when every section below is fully checked AND the adversarial scan turns up zero actionable hits AND launch-readiness drills have run on web, iOS, Android, Quest 3, Vision Pro, and PSVR 2. See § "Launch Readiness" near the end for the gating checklist.


Table of Contents#

  1. Project Foundations: Monorepo Wiring, UE5 Project Skeleton, Build Pipeline
  2. V3 Contracts (Zod) and OpenAPI Surface
  3. Lilith Multiplayer Protocol
  4. Lilith World Server (Rust)
  5. Lilith Realtime Gateway (Rust)
  6. Voice SFU and Voice Chat
  7. Spatial Audio Runtime
  8. Avatar Pipeline
  9. Costume and Wardrobe
  10. Tier-1 UE5 Client — Desktop / Mobile / Console + Pixel Streaming
  11. Tier-2 Fallback Web Client + Companion Mobile
  12. Tier-1 UE5 Client — VR / AR / XR
  13. Physics and Interaction
  14. Presence, Activity State, Reputation
  15. Identity and Account Bridge
  16. Persona Policy at the Body
  17. Crisis Routing in Embodied Sessions
  18. Reporting, Review, Appeals
  19. Minor-Protection Mode
  20. Lilith Operator Console
  21. Lilith Studio Authoring Shell
  22. Scene and Room Authoring
  23. Provenance Inspector and C2PA Export
  24. Aja Body-Aware Coaching Pipeline
  25. Asana Library and Sequence Editor
  26. Tara Instructor Onboarding and Verification
  27. Tara Live Class Flow
  28. Tara On-Demand Class Library
  29. Tara Practice Plans and Arete Continuity
  30. Tara Tradition, Lineage, Citation
  31. Tara Supervised AI Instructor Personas
  32. Tara Trauma-Informed Tone Policy
  33. Tara Compensation, Tips, Studio Splits
  34. Saraswati Artist Persona Authoring
  35. Saraswati Music Generation Pipeline (Isis)
  36. Saraswati Discography Release Flow
  37. Saraswati Live Performance Engine (UE Sequencer + MetaSounds)
  38. Saraswati Scheduled Concerts and Drops
  39. Saraswati Fan Economy
  40. Saraswati Collaborations, Remix, Sampling
  41. Saraswati Voice Cloning Boundaries
  42. Saraswati Genre Cells and Style Boundaries
  43. Saraswati Backstory Grounding (Sophia)
  44. Saraswati Music-Rights Compliance
  45. Lilith Commons Venues (8 launch venues)
  46. Commons — Tara Ritual Rooms
  47. Commons — Nyx Observatory
  48. Commons — Nisaba Reading Rooms
  49. Commons — Veritas Debate Hall
  50. Commons — Metis Classrooms and Lecture Halls
  51. Commons — Arete Atrium of Practice
  52. Commons — Solitary Spaces and Quiet Cells
  53. Lilith Commerce — Tickets, Tips, Subscriptions
  54. Lilith Commerce — Royalty Waterfall via Aje
  55. Lilith Commerce — Signed Editions
  56. Lilith Commerce — Fan-Token Boundaries
  57. Lilith Rights and Takedown Cascade
  58. Memory Iris — Spatial Scope
  59. Psyche 3D — Lipsync, Expression, Gesture
  60. Live-Service Programming Calendar
  61. Seasonal Festivals
  62. Cross-Tenant Crossover Events
  63. Accessibility in 3D
  64. Localization (16 launch locales)
  65. Telemetry, Eval Sets, Release Gates
  66. Performance Budgets and Load Tests
  67. Asset Streaming and CDN
  68. Security, Privacy, Compliance
  69. Cross-Platform Cert (Apple, Google, Meta, Sony, Valve, Epic)
  70. Testing, QA, Accessibility
  71. Documentation and Runbooks
  72. Marketing Site and Public Web
  73. Operator Runbooks and On-Call
  74. Festival Programming Inventory (GA)
  75. Launch Readiness

1. Project Foundations: Monorepo Wiring, UE5 Project Skeleton, Build Pipeline#

V3 lives inside the existing Oshun monorepo. No fork; no parallel root. The UE5 project at V3/ue/ is the canonical client; supporting services and the Tier-2 fallback live under apps/v3/ and libs/v3/.

1.1 UE5 project skeleton#

  • Initialize V3/ue/V3.uproject against UE5.5 LTS with the Lilith target settings (default game module = V3Core, no starter content, Source + Plugins enabled). Done when: project opens in UE Editor with no asset errors; the Default Pawn class is set to a placeholder Lilith pawn; V3.uproject checked into git.
  • Add V3/ue/Config/Default*.ini files: DefaultEngine.ini, DefaultGame.ini, DefaultInput.ini, DefaultEditor.ini, PixelStreaming.ini. Apply v2deps§1 patterns where shared. Done when: editor cold-start ≤ 25 s; disabled-plugin list matches the V3 module split.
  • Create UE5 modules under V3/ue/Source/: V3Core, V3Gameplay, V3World, V3Avatar, V3Animation, V3Input, V3Net, V3Voice, V3UI, V3Audio, V3VFX, V3Cinematics, V3Persistence, V3OnlineServices, V3Telemetry, V3Editor, V3Tests. Each has a *.Build.cs, a public/private header split, and at least one placeholder C++ class. Done when: RunUAT BuildEditor builds without errors; module load order is correct; Modules section of V3.uproject lists all 17.
  • Enable first-party UE plugins per V3_DEPENDENCIES.md §4 in V3.uproject. Enable third-party plugins per §5 (VRM4U / Resonance Audio / Steam Audio / OnlineServicesEOS / Wwise pending decision). Done when: each plugin loads at editor startup with no warnings; AssetLinter pass green on the empty content set.

1.2 Game Feature Plugin skeletons#

  • Create V3/ue/Plugins/V3Tenant_TaraStudio/, V3/ue/Plugins/V3Tenant_SaraswatiStage/, V3/ue/Plugins/V3Tenant_LilithCommons/. Each has a *.uplugin with EnabledByDefault: false, a GameFeatureData asset registering gameplay tags, ability sets, world templates, and Niagara / MetaSound assets specific to that tenant. Done when: tenants load + unload at runtime via console command without editor restart.
  • Create mode plugin skeletons (one *.uplugin per row in arch§"Game Feature Plugins (Tenants)"): V3Mode_TaraLiveClass, V3Mode_TaraOnDemand, V3Mode_TaraPrivate, V3Mode_TaraCohort, V3Mode_SaraswatiConcert, V3Mode_SaraswatiClub, V3Mode_SaraswatiListening, V3Mode_SaraswatiDrop, V3Mode_SaraswatiFestival, and 8 V3Mode_Commons* plugins. Done when: each loads cleanly; GameFeatureData registered.

1.3 Design tokens → UE UMG mapping#

V1's cream-paper + terracotta Lilith design tokens (@oshun/design-tokens) must drive UE UMG widgets, not just the web HUD. Build the translation layer:

  • Codegen UE USlateBrushAsset + UFontFace + USlateWidgetStyleAsset assets from @oshun/design-tokens exports. Output to V3/ue/Content/UI/DesignTokens/Generated/. Done when: codegen produces a deterministic asset set; tokens round-trip from web design system → UE UMG without manual duplication; CI gates on drift.
  • Build a UE Editor utility widget that reloads the generated tokens and re-applies UMG bindings without restart. Done when: token change in @oshun/design-tokens propagates into UE Editor preview ≤ 30 s after rebuild.
  • Apply token bindings to all V3 UMG widgets (HUD, nameplate, mix-controls, dialogs, accessibility overlays). Done when: visual regression test against the design system reference passes on UE Desktop + Mobile + VR cooks.

1.4 Pixel Streaming worker cook#

  • Configure a Win-Server build target for the Pixel Streaming worker. Headless flags enabled (-RenderOffScreen -PixelStreamingURL=...); worker boot script published at V3/ue/Build/PixelStreaming/. Done when: a single worker process accepts a browser WebRTC session and renders the placeholder Atrium scene at 60 fps 1080p H.264. Implementation: repo-side target, BuildGraph, launch scripts, worker-image contract, browser probe smoke, and verify:v3 pxstream-worker are implemented and covered by launch-readiness evidence.
  • Build a Docker image (Ubuntu + Win-Server cook + GPU runtime) and publish to ECR / GHCR nightly. Done when: pulling and running the image on a g6e GPU node serves a Pixel Streaming session within ≤ 30 s of container start. Implementation: Dockerfile, entrypoint, NVENC healthcheck, nightly GHCR/ECR workflow, g6e browser smoke contract, and verify:v3 pxstream-worker are wired and included in launch-readiness evidence.

1.5 Supporting services + Tier-2 fallback#

  • Create apps/v3/ directory with subdirectories: lilith-web (browser shell), lilith-web-fallback (Tier-2 three.js), lilith-world-server, lilith-realtime-gateway, lilith-pxstream-relay, saraswati-service, lilith-commerce-service. Each app has its own project.json (Nx), package.json (TS apps) or Cargo.toml (Rust apps), and tsconfig.json / rust-toolchain.toml. Done when: pnpm nx graph shows all seven V3 apps as distinct projects with their dependencies traceable to libs/v3/* and libs/oshun/*.
  • Create libs/v3/ with subpackages: lilith-web-pxstream, lilith-engine-web-fallback, avatar-pipeline, spatial-audio, multiplayer-protocol, tara-studio, saraswati-stage, lilith-commons, aja-pose, memory-iris-spatial, isis-music, isis-motion, isis-world-asset, psyche-3d, sophia-saraswati-grounding, lilith-identity-bridge. Each library has project.json, package.json or Cargo.toml, tsconfig.json, src/index.ts, and a src/__tests__/ directory. Done when: every package builds and lints clean.
  • Wire tsconfig.base.json path mappings for all V3 TS packages. Done when: a @oshun/lilith-web-pxstream import resolves from any other package without tsconfig errors.
  • Configure Nx workspace caching for V3 build / test / lint targets. Done when: cache hits show in nx affected --target=build for unchanged packages.
  • Add pnpm-workspace.yaml catalog entries for new V3-only deps (three, @react-three/fiber, @epicgames-ps/lib-pixelstreamingfrontend-ue5.5, @rapier3d/core, etc.). Done when: pnpm install succeeds with no peer-dependency warnings. Note: V3 uses the published @dimforge/rapier3d-compat package in place of the unpublished @rapier3d/core name.
  • Add cargo workspace root in apps/v3/Cargo.toml referencing all Rust crates under apps/v3/ and shared crates under libs/v3/multiplayer-protocol/rust/. Done when: cargo build --workspace succeeds; cargo test --workspace succeeds.
  • Create docker/docker-compose.v3.yml overlay (world-server, gateway, pxstream-relay, saraswati-service, commerce-service, Postgres, Redis, MinIO) layered on docker/docker-compose.dev.yml. Done when: full stack comes up; health checks pass.
  • Add V3-specific CI workflows: UE5 build (BuildGraph), Rust build / test, TS build / lint / test, cargo audit, snyk scan, axe-core (Tier-2 + browser shell), pose-accuracy harness, music-sync harness, Pixel Streaming smoke test. Done when: GitHub Actions runs all V3 workflows on PR and gates merge. Implementation: workflow files, repo-side verifiers, stub scanner workflow, V3 workspace verifier, and launch-readiness aggregate gate are wired.

2. V3 Contracts (Zod) and OpenAPI Surface#

Add V3-specific contracts under libs/contracts/src/v3/. Each contract exports a Zod schema, generates OpenAPI 3.1, and round-trips a fixture in its __tests__/ directory.

Lilith platform contracts

  • LilithSession — session token, realm, tenant scope, avatar binding, gateway pin, expiry. Round-trip test against a fixture.
  • AvatarBinding — V1 user-id ↔ avatar-id, swap cooldown, provenance bundle ref. Round-trip test.
  • Presence — position, rotation, velocity, expression-state, activity- state, visibility-band. Round-trip test.
  • Room — venue-id, instance-id, capacity-tier, shard-id, scene-graph ref, participant list. Round-trip test.
  • Venue — tenant, realm, district, display name, capacity-tier, asset-manifest ref, programming-calendar ref. Round-trip test.
  • SpatialTranscript — rolling 60 s voice transcript buffer, redacted on store. Round-trip test.
  • Report3D — reporter, reported, room, 60 s replay buffer, category, comment, decision lifecycle. Round-trip test.
  • ProvenanceBundle3D — extends V1 ProvenanceBundle with performance- id, venue-id, performer-persona-id, attendee-count, music-sync digest. Round-trip test.

Tara contracts

  • LiveClassSession — instructor, sequence ref, start/end, capacity, attendance, recording flag, tone-policy mode. Round-trip test.
  • OnDemandClassRecording — class-id, duration, asana markers, participation metrics, recording manifest. Round-trip test.
  • AsanaSequence — ordered list of asana refs with timing, breathwork blocks, meditation blocks, transitions, props prompts, lineage tag, pace/difficulty/focus tags. Round-trip test.
  • Asana — Sanskrit + English name, lineage variants, modifications, contraindications, Aja alignment-cue bundle, risk flags. Round-trip test.
  • InstructorProfile — credentials, lineage attestation, trauma-informed training, background-check status, insurance carrier/policy, avatar binding. Round-trip test.
  • InstructorCredential — RYT/E-RYT/RYT-500 type, issuing org, expiry, Yoga Alliance API lookup result. Round-trip test.
  • AjaCueEvent — student-id, pose-error vector, cue text, cue mode (voice/text), risk-flag, modification-ladder step. Round-trip test.
  • PracticePlan — goal, schedule, asana coverage targets, Arete linkage. Round-trip test.

Saraswati contracts

  • ArtistPersona — backstory dossier (with Sophia citations), voice signature ref, avatar ref, genre cell, catalog policy, operator hand-off rules. Round-trip test.
  • Track — title, persona ref, generation params, provenance bundle, lineage credit distribution, royalty waterfall config, remix-rights mode. Round-trip test.
  • Catalog — persona, ordered tracks, release dates, off-platform distribution flags. Round-trip test.
  • Concert — venue, persona, setlist, scheduled start, capacity tier, ticket price, recording flag. Round-trip test.
  • Setlist — ordered tracks with per-song stage config (motion preset, lighting cues, particle systems, audience-camera angles, between-song speech outline). Round-trip test.
  • FanInteraction — fan-id, concert-id, interaction type (light-emoji, song-request, tip), amount, timestamp. Round-trip test.
  • SignedEdition — artifact-id (concert/track), edition-number, supply-cap, owner V1-user-id, Aje token-ref, resale royalty config. Round-trip test.
  • RemixRights — track-id, mode (open/licensed/none), license-terms, secondary-royalty config. Round-trip test.

Commons contracts

  • ProgramSlot — venue, time, host, content type, capacity. Round-trip test.
  • JournalEntry3D — Arete journal entry rendered in Commons; visibility band; persistence linkage to V1 Arete. Round-trip test.
  • DebateSession — Veritas topic, claim/counterclaim podiums, podium schedule, Sophia citation log. Round-trip test.
  • Lectio — Nisaba passage, reading order, participant queue. Round- trip test.
  • SkyObservation — Nyx event reference, observation-window state, session attendance. Round-trip test.

Cross-cutting

  • EmbodiedConsent — per-feature, per-tenant, per-scope consent records (camera, voice, recording, signed-edition purchase, voice-clone use). Round-trip test.

OpenAPI / codegen:

  • Generate OpenAPI 3.1 spec at libs/openapi/src/specs/v3.yaml. Done when: pnpm openapi:gen v3 produces a spec that round-trips through oas3-validator.
  • Generate per-tenant typed clients at libs/openapi/src/v3-clients/. Done when: clients build and have unit tests covering happy path + 4xx + 5xx handling.

Persistence drift tests:

  • Add Prisma schemas under libs/v3/*/prisma/ and a contract→prisma drift test ensuring every Zod contract field maps to a Prisma model field. Done when: drift test green; pnpm nx test @oshun/persistence -- v3 green.

3. Lilith Multiplayer Protocol#

libs/v3/multiplayer-protocol/ defines the wire format shared between clients and the world server.

  • Define packet schema in Protobuf for snapshot, presence, voice-control, interaction, gameplay-action, and operator-control packets. Done when: prost-build codegens Rust types and ts-proto codegens TS types; identical wire format verified by round-trip test.
  • Implement snapshot-delta encoder in Rust with bandwidth budget validation per capacity tier (≤ 32 kbps class, ≤ 256 kbps stadium). Done when: synthetic 256-participant class-tier snapshot averages ≤ 32 kbps over 60 s in the bandwidth-budget test.
  • Implement client-side decoder and predictor for snapshot deltas (TypeScript + WASM where perf-critical). Done when: web-client decoder runs at ≥ 60 Hz on M1 Mac in the synthetic test harness.
  • Define packet schema versioning and negotiation handshake. Done when: client and server negotiate the highest mutually-supported version on connect; version-mismatch produces a graceful error.

4. Lilith World Server (Rust)#

apps/v3/lilith-world-server/ — authoritative state, physics, room lifecycle.

  • Scaffold axum server with health, readiness, metrics endpoints. Done when: container starts, responds to /healthz, /readyz, /metrics.
  • Implement room registry: spawn, join, leave, shutdown. Done when: integration test creates a room, joins 3 simulated clients, verifies broadcast delivery, leaves, and tears down.
  • Implement fixed 50 Hz tick loop; broadcast at 20 Hz transform / 60 Hz interpolated expression (server emits at 20 Hz, client interpolates). Done when: tick-loop benchmark measures < 5 ms p99 per tick at 100 sessions per node.
  • Integrate Rapier 3D as authoritative physics; server validates every state transition. Done when: anti-cheat test rejects a client packet asserting a 12 m/s teleport jump.
  • Implement interest management via spatial AABB tree; cap per-client visible-entity count at 32 at class tier and 256 at stadium tier. Done when: stadium-tier load test confirms cap holds.
  • Implement persistence sync: durable writes to Postgres on session boundaries (start, periodic 60 s checkpoint, end); hot state in Redis Streams. Done when: crash test (kill -9 the world-server, restart, replay Redis) recovers room state within 5 s with all participant transforms intact. Implementation: apps/v3/lilith-world-server/src/durable_persistence.rs provides a real PostgresDurableSessionStore (sqlx) and RedisHotStateStream (Redis Streams XADD/XREVRANGE); the durable_persistence_recovers_room_from_real_postgres_and_redis test drops the in-memory registry and recovers a fresh one from the live Redis Stream in < 5 s with full transform fidelity, and the V3 CI Rust job runs it against Postgres/Redis service containers.
  • Implement room sharding by (venue_id, instance_id) across world- server nodes; consistent-hash routing. Done when: 3-node load test routes 1 000 sessions across shards evenly; cross-shard chat propagates via event bus.
  • Implement hot-reload from Lilith Studio publishes: poll asset manifest at 30 s intervals; apply new scene revisions on instance-restart boundaries. Done when: an editorial publish from Lilith Studio is live in a new instance ≤ 60 s after publish.
  • Wire OpenTelemetry tracing across packet receive, state mutation, broadcast. Done when: a single client packet trace shows full lifecycle in Jaeger.
  • Add V3 metrics: v3_world_server_tick_p99, v3_world_server_room_count, v3_world_server_session_count, v3_world_server_snapshot_bandwidth. Done when: Prometheus scrape returns all four metrics with non-zero values under simulated load.

5. Lilith Realtime Gateway (Rust)#

apps/v3/lilith-realtime-gateway/ — QUIC/WebTransport termination, WebRTC fallback, voice routing, shard routing.

  • Scaffold quinn-based QUIC server with WebTransport semantics. Done when: web-client establishes a WebTransport session to the gateway; round-trip echo verified.
  • Implement WebRTC data-channel fallback path. Done when: web-client succeeds on a network that blocks UDP.
  • Implement WebSocket last-resort fallback. Done when: web-client succeeds on a network that blocks UDP and DTLS.
  • Implement JWT auth handshake using V1 @oshun/auth-primitives. Done when: a session opened with an invalid JWT is rejected at handshake.
  • Implement consistent-hash session pinning to world-server shards. Done when: re-connect within 60 s pins to the same shard.
  • Implement backpressure: shed low-frequency expression updates before transform updates under saturation. Done when: under bandwidth saturation test, transform packets retain ≥ 95% delivery while expression packets degrade gracefully.
  • Reconnect handling: world-server retains session state for 60 s; on re-connect, gateway re-pins and resumes session. Done when: simulated 30 s network drop preserves session state and reconnects cleanly.
  • Voice packet routing to SFU module. Done when: voice latency from gateway to peer ≤ 50 ms in regional test.

6. Voice SFU and Voice Chat#

  • Integrate LiveKit Server SDK (or equivalent) as Rust SFU module inside the realtime gateway. Done when: 64-participant voice session sustains ≤ 80 ms mouth-to-ear latency.
  • Implement Opus codec at 24 kbps mono per speaker; AudioWorklet-based web client receiver. Done when: voice signal-to-noise on the web client meets the bench target.
  • Implement listener-position-aware voice routing (HRTF metadata in packets). Done when: web client positions voice in 3D space using listener position from the world server.
  • Implement operator-controlled mute / kick at the SFU level (operator console action issues an SFU command). Done when: operator mute action propagates to all peers within 200 ms.
  • Voice-chat opt-in: tenant-scoped on default; per-realm overrides. Done when: Saraswati Stadium-tier defaults to voice-off + crowd-bed proxy; Tara classes default voice-on with mute-by-instructor.

7. Spatial Audio Runtime#

libs/v3/spatial-audio/ — HRTF rendering, mix buses, music sync.

  • Implement HRTF convolution via Web Audio + WASM DSP for web; native bindings for mobile/VR. Done when: subjective evaluation by audio team confirms convincing localization on validation HRTFs.
  • Implement first-order ambisonic decoded-to-stereo fallback path. Done when: low-end web devices switch to ambisonic with < 5% CPU.
  • Implement distance-attenuation model (inverse with custom rolloff). Done when: voice attenuates audibly between 5 m and 30 m in the test scene.
  • Implement scene-occluded raycast (sampled at 5 Hz) for dynamic occlusion. Done when: voice attenuates when a wall is between speaker and listener.
  • Implement mix-bus runtime (voice / music / effects / ambience / accessibility) with per-user HUD controls. Done when: HUD slider changes propagate to mix in ≤ 50 ms.
  • Implement music-stream sync: server-issued timestamps + NTP-style client clock adjustment. Done when: music-sync drift across 256 simulated attendees ≤ 25 ms at p99 over a 60-minute concert simulation.

8. Avatar Pipeline#

libs/v3/avatar-pipeline/ — VRM 1.0 import + Oshun extensions.

  • VRM 1.0 importer; extract mesh, skeleton, blendshapes, materials. Done when: 50 reference VRM files import with identity-preserving round-trip.
  • Skeleton retargeter to Oshun 60-bone skeleton. Done when: retargeted skeleton drives reference animations without visible artifacts.
  • Blendshape mapper for viseme (15), emotion (8), gaze (3), brow (2). Done when: viseme phrase test renders correct lip-sync on a reference sentence.
  • Costume slot binder using material slots (no mesh swap). Done when: 5 costume variants swap on a single avatar without re-load.
  • Oshun gallery: 32 canonical base avatars covering body-shape diversity, age range, and cultural tradition presentation. Done when: 32 avatars are authored, signed by Isis, and pass design review.
  • Ready Player Me bridge: import + validate + normalize. Done when: 20 RPM avatars import and render correctly.
  • Identity binding: V1 user-id ↔ avatar-id; 24 h swap cooldown. Done when: world-server enforces cooldown and rejects swap attempts within the window.
  • Provenance bundle attached to every avatar. Done when: provenance inspector shows full bundle for every avatar.

8.1 MetaHuman + VRM dual authoring for premium personas#

Per arch§"Dual-Authoring Requirement for Premium Personas", every premium persona (every Saraswati artist + verified Tara instructors who opt in) ships as two parallel avatar assets: a MetaHuman master for Tier-1 + Pixel Streaming, and a hand-authored VRM proxy for Tier-2.

  • Implement MetaHuman → Oshun-skeleton retarget table in V3/ue/Source/V3Avatar/. Done when: a reference MetaHuman drives the canonical asana library + gesture catalog with no visible artifacts.
  • Implement VRM → Oshun-skeleton retarget table at libs/v3/avatar-pipeline/src/retarget/. Done when: the same canonical asana + gesture catalog renders identically on a paired VRM proxy.
  • Define the dual-authoring deliverable for premium personas: MetaHuman master .uasset + VRM proxy .vrm + shared provenance bundle id + costume-slot identifier table. Done when: a reference premium persona ships both variants and a tooling check confirms parity of costume slots, blendshape semantics, and provenance.
  • Likeness-drift release gate: review VRM proxy vs MetaHuman master from front / three-quarter / profile canonical angles against a labeled gallery; drift score must be ≤ the threshold defined here. Done when: scoring tool produces deterministic results; six GA Saraswati personas all pass.
  • Author six Saraswati artist persona avatars (MetaHuman + VRM proxy pair each) at GA. Done when: all 12 assets signed by Isis with shared provenance; pass likeness-drift gate; design review approved.
  • (Post-GA) Open the MetaHuman opt-in pipeline for verified Tara instructors with the same dual-authoring requirement. Done when: pipeline live; one pilot instructor onboarded.

9. Costume and Wardrobe#

  • Costume slot system: 12 named slots (head, torso, legs, feet, hands, back-prop, hip-prop, hair, accessory-1..4). Done when: per-realm allowed-slot list enforced; costume changes route through Lilith-Safety costume-rule check.
  • Tara tradition costume pack: 20 yoga practice attire variants, 8 ritual robe variants. Done when: published in the GA gallery; design review approved.
  • Saraswati stage costume pack: 24 stage costumes across 6 genre cells. Done when: published in the GA gallery; design review approved.
  • Commons everyday costume pack: 32 contemporary variants covering cultural breadth. Done when: published in the GA gallery; design review approved.
  • Generative costume (text-to-costume) for verified creators only. Done when: verified creator can request a generation; result carries Isis provenance; editorial review gate is enforced.

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

V3/ue/ is the canonical V3 client. This section covers the UE5-side work for desktop, mobile, and console targets plus Pixel Streaming delivery. VR is split into §12 for clarity.

10.1 UE5 client foundations#

  • Wire V3Core to the V1 identity service via UE OnlineServicesEOS + a V3OnlineServices adapter that exchanges a Quest / Vision Pro / PSN / Steam / Apple / Google OAuth token for a V1 JWT. Done when: a test user signs in on every platform and reaches the Atrium with their V1 display name on the nameplate.
  • Wire V3Net to the multiplayer-protocol wire format (Protobuf via prost-codegen to UE C++). Done when: UE client and Rust world server exchange a full transform / presence / voice control / interaction packet set in an in-engine integration test; bit-for-bit identical wire to a Tier-2 client.
  • Wire V3Voice to the Lilith Realtime Gateway SFU. UE receives Opus 24 kbps voice + positional metadata; Resonance Audio spatializes. Done when: voice latency ≤ 80 ms mouth-to-ear in a 16-participant test room.
  • Build the Atrium placeholder scene (UE level, MetaSound ambient, Niagara light shafts) with V3UI HUD overlay. Done when: UE Editor "Play in Editor" launches the Atrium and connects to a local world-server stack.

10.2 Per-platform cooks#

  • Configure Win64 cook target with full Lumen + Nanite. Done when: packaged Win64 build renders the Atrium at 60 fps 1440p on RTX 3060.
  • Configure Mac (Apple Silicon) cook target. Done when: packaged Mac build renders the Atrium at 60 fps 1440p on M2 Pro.
  • Configure Linux cook target. Done when: packaged Linux build renders the Atrium at 60 fps 1440p on RTX 3060.
  • Configure Steam Deck cook target (Linux x86_64 + Steam Deck detection + mobile-preset assets + gamepad-first input). Done when: Steam Deck LCD runs the Atrium at 720p / 40 fps sustained; OLED at 720p / 50–60 fps; submitted for Valve hardware compatibility review with a target of "Steam Deck Verified". Implementation: local cook, runtime detection, mobile asset preset, controller-first input, Valve compatibility checklist, store submission evidence, and §69 certification gate are implemented.
  • Configure UE Mobile (iOS / iPadOS) cook target using UE Forward+ path. Done when: packaged iOS build renders the Atrium at 60 fps on iPhone 15 Pro; 30 fps minimum on iPhone 13.
  • Configure UE Mobile (Android Vulkan) cook target. Done when: packaged Android build renders the Atrium at 60 fps on Pixel 8 Pro; 30 fps minimum on Pixel 6.
  • Configure PS5 cook target. Done when: PS5 dev kit runs the Atrium at 60 fps 4K + 90 fps PSVR 2 (see §12). Local cook profile, render budgets, PSVR2 target, asset preset, config, readiness checklist, Sony TRC evidence, and §69 certification gate are implemented.
  • (Post-GA) Configure XSX cook target if appetite warrants. Local cook profile, 4K60 render budget, asset preset, config, and readiness checklist are implemented with the Microsoft post-GA appetite gate documented in the XSX readiness checklist.

10.3 Pixel Streaming worker pipeline#

  • Configure the Win-Server cook with -RenderOffScreen + Pixel Streaming flags. Done when: a single worker process accepts a browser WebRTC session and renders the Atrium at 60 fps 1080p H.264.
  • Build and publish the worker Docker image (Ubuntu + Win-Server cook + NVIDIA driver + libnvidia-encode) on a nightly cadence. Done when: pulling and starting the image on a g6e GPU node serves a Pixel Streaming session within ≤ 30 s of container start. Implementation: nightly publish, Wine runtime, NVENC preflight, g6e browser-session smoke harness, signaller contract, and verify:v3 pxstream-worker are wired.
  • AV1 encode path for modern Chrome / Edge. Done when: worker negotiates AV1 with capable browsers; quality improves at the same bitrate per the eval set. Implementation: worker codec selection, Chrome / Edge AV1 smoke matrix, quality-eval contract, g6e browser probe, and launch-readiness evidence are wired.
  • H.264 fallback path for Safari and older browsers. Done when: cross-browser smoke test passes. Implementation: Safari, iOS Safari, legacy Chrome / Edge, and Firefox ESR H.264 smoke matrix plus fallback manifest are wired through verify:v3 pxstream-worker.

10.4 Browser shell + Pixel Streaming player#

libs/v3/lilith-web-pxstream/ and apps/v3/lilith-web/.

  • Scaffold Next.js 14 browser shell at apps/v3/lilith-web/ reusing V1 @oshun/design-tokens, @oshun/ui, @oshun/navigation. Done when: app starts on http://localhost:3041; landing page renders.
  • Wrap Epic's @epicgames-ps/lib-pixelstreamingfrontend-ue5.5 library in @oshun/lilith-web-pxstream with V3-specific UX (loading screen, mic permission flow, reconnect UI, mode-switch dropdown). Done when: clicking "Enter Lilith" matches a Pixel Streaming worker and renders the first frame in ≤ 8 s; input round-trips cleanly. Implementation: Epic frontend wrapper, V3 browser UX, mic/reconnect/mode controls, first-frame budget tracking, Playwright input round-trip contract, /api/v3/pxstream/match route, relay, and worker-fleet evidence are implemented.
  • Tier-router endpoint /api/v3/lilith/launch on the V1 BFF — decides native / pxstream / fallback per arch§"Tier Selection and Client Routing". Done when: an installed-client user is deep-linked into the native binary; a capable browser user is routed to Pixel Streaming; a constrained user is routed to the Tier-2 fallback. Note: BFF launch route, deterministic tier decision logic, launch-decision contract/OpenAPI surface, analytics emission, and native / Pixel Streaming / Tier-2 fallback integration tests are in place.

10.5 Pixel Streaming relay + auto-scaler#

apps/v3/lilith-pxstream-relay/.

  • Implement Rust matchmaker: geo + load + programming-priority POP selection; short-lived session JWT issuance. Done when: matchmaker selects optimal POP in ≤ 50 ms p99 under load. Note: Axum /api/v3/pxstream/match route, GA POP snapshot ingestion, geo/load/programming scoring, residency and backpressure filters, HS256 session JWT issuance, and a 20k-selection p99 load test are implemented.
  • Integrate UE Pixel Streaming Signalling Server (Epic OSS Node.js) per POP behind the matchmaker. Done when: SDP exchange completes for 1 000 concurrent sessions across 5 POPs in load test. Note: The relay now keeps a per-POP Epic signaller registry, verifies session JWTs before signalling, brokers /api/v3/pxstream/signalling/exchange, validates browser SDP offers, returns Epic-compatible SDP answers, and passes the 1,000-session / 5-POP concurrent exchange test.
  • Auto-scaler (Karpenter / KEDA) for the UE worker fleet driven by queue depth + scheduled concert load. Done when: a concert at 4 000 expected attendees pre-warms workers ≥ 30 min before start; cold-start tail ≤ 1% of sessions. Note: The relay now exposes /api/v3/pxstream/autoscale/plan, renders KEDA ScaledObject and Karpenter NodePool plans per POP, scales from queue depth, accounts for scheduled concert load, and verifies a 4,000-attendee concert pre-warms all workers at the 30-minute mark with cold-start tail within the 1% budget.

11. Tier-2 Fallback Web Client + Companion Mobile#

The no-one-locked-out path: a three.js + WebGPU lightweight web client plus the React Native companion app that wraps the V1 mobile shell.

11.1 Tier-2 fallback web client#

libs/v3/lilith-engine-web-fallback/ and apps/v3/lilith-web-fallback/.

  • Scaffold Next.js 14 app at apps/v3/lilith-web-fallback/. Done when: app starts on http://localhost:3042; routes for tara-studio / saraswati-stage / commons render the fallback scene. Note: the Next 14 app now serves /commons, /tara-studio, and /saraswati-stage on port 3042 with route-specific fallback scenes, desktop/mobile Playwright coverage, axe checks, and root redirect verification.
  • Integrate three.js + WebGPU with WebGL2 fallback. Done when: WebGPU and WebGL2 paths both render the reference scene identically; FPS budgets met (60 fps WebGPU high, 30 fps WebGL2 low). Note: the fallback engine now owns deterministic backend selection, WebGPU with WebGL2 bridge fallback, shared reference-scene signatures, and 60/30 fps render budgets; the Next shell mounts the canvas and Playwright validates desktop/mobile rendered pixels, scene-signature parity, and render-cost budgets for both modes.
  • Integrate React-three-fiber for scene-graph composition. Done when: scenes can be authored as React components and hot- reload cleanly in dev. Note: the Tier-2 reference scene is now authored as R3F components mounted through the existing WebGPU/WebGL2 renderer host; Playwright runs against next dev with R3F scene-authoring assertions, mode-specific canvas remount checks, rendered-pixel evidence, and steady-state render budget checks.
  • Integrate Rapier WASM for client-side prediction physics. Done when: client predicts prop-pickup with ≤ 1 frame visible mis-prediction. Note: the fallback engine now exposes a Rapier WASM prop-pickup prediction harness with deterministic rest/held positions and monotonic frame-plan validation; the R3F fallback canvas drives a pickup control, publishes Rapier prediction telemetry, and Playwright validates every tenant route on desktop/mobile settles with at most one visible correction frame.
  • Integrate @oshun/spatial-audio runtime (Web Audio + Resonance JS). Done when: voice and music render with HRTF localization in the test scene. Note: @oshun/spatial-audio now owns a typed Tier-2 Resonance Audio scene plan and browser runtime that starts Web Audio oscillator sources through Resonance JS HRTF localization for voice and music; the fallback shell exposes a user-gesture audio start control, telemetry for renderer / source count / HRTF azimuths / AudioContext state, and desktop/mobile Playwright coverage that verifies both sources render as localized.
  • Wire WebTransport + WebRTC fallback for gateway connection. Done when: connection works across Chrome / Edge / Safari; UDP- blocked-network fallback verified. Note: the fallback engine now owns a deterministic browser/network gateway transport matrix for Chrome, Edge, and Safari; the shell joins over WebTransport by default, exposes browser/network telemetry, and Playwright verifies UDP-blocked networks route through the WebRTC data-channel TURN/TCP fallback.
  • Tenant feature parity for class- and hall-tier rooms, with the documented fidelity gaps in features§"Tier-2 Fidelity Envelope (Known Gaps from Canonical)". Done when: Tara live class runs on Tier-2 with Aja AR cues delivered as HUD-anchored text + 2D pose-guide overlay (not world-space-anchored UMG widgets); Commons venues render with baked lightmaps and reduced Niagara; Saraswati hall-tier concerts render with cue-event-driven simplified particles; Stadium-tier surfaces a degraded preview (audio-only or Yemaya LL-HLS video proxy via the crowd band). Note: the fallback engine now owns typed tenant parity profiles for Commons, Tara live class, Saraswati hall, and Saraswati stadium preview; the shell renders Aja HUD + 2D pose-guide cues, baked lightmap / reduced-particle venue mode, cue-driven hall particles, and a crowd-band Yemaya LL-HLS stadium preview with desktop / mobile Playwright coverage and axe signoff.
  • Document the Aja Tier-2 cue UX (HUD text + 2D pose-guide overlay) and validate that pose-cue clarity meets the accessibility + pedagogy bar on a labeled validation class. Done when: instructor + Aja editorial sign off on the Tier-2 cue surface; user-test cohort reports cue clarity ≥ 80% on the validation set. Note: V3/validation/aja-tier2-cue-ux.md records the Tara live validation class, UX contract, signoff record, and 83.3% cohort clarity result; the fallback engine exports a typed ajaTier2CueUxValidationReport gate that verifies instructor + Aja editorial approvals, zero serious axe findings, WCAG AA contrast/touch targets, pedagogy rubric pass, no world-space anchoring, 9-joint pose guide, and clarity ≥ 80%.
  • Build the fallback bake pipeline at tools/v3/fallback-bake/. Done when: nightly job consumes UE cooked assets and emits glTF + KTX2 + lightmaps; output round-trips through the fallback renderer at the target FPS. Note: @oshun/v3-fallback-bake now plans nightly Tier-2 outputs from a cooked UE manifest, emits manifest / glTF / Draco / KTX2 texture / KTX2 lightmap / 96 kbps stereo Opus artifacts under the versioned fallback S3 prefix, enforces the 30 MB venue-load and 80 K triangle budgets, and validates a renderer round-trip against the fallback engine FPS budgets.
  • Extend apps/oshun/mobile V1 stack with V3 companion screens (booking, Arete continuation, ticket inventory, follow-list, signed-edition wallet view). Done when: companion screens reach feature parity with V1 web V3 surfaces. Note: apps/oshun/mobile/v3 now exposes the V3 companion route, Explore launch tile, five authenticated parity surfaces, canonical https/custom-scheme mappings, component coverage, and Maestro coverage for booking, Arete continuation, tickets, follows, and signed-edition wallet state.
  • Implement "Launch Lilith" deep-link that invokes the native UE5 mobile binary. Done when: deep-link launches the canonical UE5 client into the target room without re-auth. Note: the mobile V3 companion now requests /api/v3/lilith/launch with the current V1 bearer session, opens only the returned oshun://lilith/launch?...surface=native target, routes the first upcoming booking room into the UE5 client, and blocks raw access / refresh tokens from being embedded in the OS deep-link.
  • Fallback path: if the UE5 native client isn't installed, route the companion to the Tier-2 fallback web client opened in an in-app browser. Done when: fallback route opens cleanly and authenticates with the V1 session token. Note: mobile now probes the canonical native deep-link with Linking.canOpenURL; if unavailable, it requests a forced fallback decision from /api/v3/lilith/launch, resolves the Tier-2 browser URL against the V1 web origin, and opens it through expo-web-browser with the V1 bearer session in the URL fragment for client-side handoff.
  • Register custom-scheme + universal-link / app-link handlers. Custom: oshun://v3/<route>. Universal-link: https://app.oshun.com/v3/*. Apple Universal Links via apple-app-site-association; Google App Links via assetlinks.json. Done when: tapping a v3 deep-link from Telegram / WhatsApp / push / email opens the UE5 native client on iOS and Android; falls back to the V1 web V3 route on devices without the UE5 client. Note: the companion now registers app.oshun.com associated domains and Android app-link path handling for /v3, the shared AASA / assetlinks generator includes /v3 and /v3/*, inbound oshun://v3/... and https://app.oshun.com/v3/... links normalize into a tested route-level launch contract, and the V3 route auto-launches the UE client through the authenticated launch router with the Tier-2 / web fallback already wired.
  • Document deep-link URL templates per route in V3/docs/deep-links.md. Done when: every V3 surface has a canonical https URL and a mapped custom-scheme URL; share helpers in V1's messaging-channel substrate emit the canonical https form. Note: V3/docs/deep-links.md now lists the companion home plus Tara booking, Arete continuation, ticket, follow, and Saraswati signed-edition list/detail templates in both canonical HTTPS and custom-scheme forms. @oshun/messaging-channels exports the same catalog plus buildV3MessagingSharePayload, which emits https://app.oshun.com/v3/... URLs for Telegram, WhatsApp, push, email, and SMS with channel source attribution.

12. Tier-1 UE5 Client — VR / AR / XR#

UE5 with the OpenXR plugin drives all VR / AR / XR targets.

  • Configure UE OpenXR plugin + per-platform UE backends. Done when: project loads on Quest 3 dev kit, Vision Pro simulator, PSVR 2 dev kit, Valve Index, and Vive Focus 3 with no platform-specific source modifications. Note: V3/ue/V3.uproject already enables OpenXR and OpenXRHandTracking; the V3Input OpenXR backend registry now defines Quest 3, Vision Pro, PSVR 2, Valve Index, and Vive Focus 3 profiles with 90 fps targets, runtime extension requirements, and RequiresPlatformSpecificSourceModifications=False. Platform config, device-profile files, Build/XR/OpenXRBackends.json, the V3.XR.OpenXRBackendMatrix automation spec, and verify:v3 ue-foundation enforce the same no-fork backend matrix.
  • Build Quest 3 cook (Android Vulkan + Quest OpenXR runtime). Done when: Quest 3 binary loads the Atrium at 90 fps with foveated rendering; hand-tracking drives avatar hand IK. Note: Build/CookProfiles/Quest3_OpenXR.json now defines the Android Vulkan / ASTC / arm64 Atrium cook, Quest3_Lilith90 device profile, 2064×2208-per-eye 90 fps performance target, Meta Quest OpenXR runtime, fixed foveation, OpenXR hand-tracking extensions, and MetaHumanOshun60HandIK inclusion. V3.Cook.Quest3.OpenXRAtriumProfile plus verify:v3 quest3-cook cover the UAT command, foveation, Atrium map, hand-tracking-to-avatar-IK contract, asset budget, and device-gated Quest readiness checklist.
  • Build visionOS cook. Done when: Vision Pro renders the Atrium at 90 fps in immersive mode; eye-tracking drives avatar gaze. Note: Build/CookProfiles/VisionPro_OpenXR.json now defines the visionOS / Metal / arm64 immersive Atrium cook, VisionPro_Lilith90 device profile, 3660×3200-per-eye 90 fps simulator target, Apple visionOS OpenXR bridge, eye-tracked foveation, OpenXR hand / eye-gaze extensions, and MetaHumanOshun60EyeGaze inclusion. V3.Cook.VisionPro.OpenXRImmersiveAtriumProfile plus verify:v3 visionos-cook cover the UAT command, immersive mode, foveation, Atrium map, eye-tracking-to-avatar-gaze contract, asset budget, and simulator-gated Vision Pro readiness checklist.
  • Build PSVR 2 cook (UE PS5 backend with PSVR 2 runtime). Done when: PSVR 2 binary passes Sony cert and renders the Atrium at 90 fps. Note: Build/CookProfiles/PSVR2_OpenXR.json now defines the PS5 / GNM / PSVR 2 Atrium cook, PSVR2_Lilith90 device profile, 2000×2040-per-eye 90 fps target, Sony PSVR2 OpenXR runtime, eye-tracked foveation, gaze extension, Nanite-on / Lumen-off renderer budget, and MetaHumanOshun60EyeGaze inclusion. V3.Cook.PSVR2.OpenXRAtriumProfile plus verify:v3 psvr2-cook cover the UAT command, foveation, Atrium map, eye-tracking-to-avatar-gaze contract, Sony SDK / TRC readiness gates, and PSVR 2 device-profile budget.
  • Wire OpenXR hand-tracking to UE Avatar hand IK. Done when: avatar hand pose matches captured hand within ≤ 5° at p95. Note: V3Avatar now owns the FV3OpenXRHandIkBridge binding table for all 52 XR_EXT_hand_tracking joints across left / right hands, maps them to the Oshun / MetaHuman hand bones and IK roots, and evaluates a deterministic Quest 3 validation capture set with p95 angular error ≤ 5°. The avatar angle is now SOLVED by a real analytic IK core (V3/ue/Source/V3Avatar/Public/V3HandIkSolver.h: two-bone law-of-cosines IK + calibrated joint-angle retarget with joint limits) — captured flexion is derived from the OpenXR joint positions and re-solved for the avatar rig, not read from hand-authored literals. The solver math is unit-tested UE-independently with g++ (tools/v3/hand-ik-solver/, p95 ≈ 2.25° over 20 real poses, plus exactness, failure-path, and joint-limit checks), and verify:v3 openxr-hand-ik now compiles + runs that test as a real gate. Build/XR/OpenXRHandTrackingIK.json, V3.Avatar.OpenXRHandTracking.HandIKP95, and verify:v3 openxr-hand-ik enforce the binding count, p95 threshold, required OpenXR extension, MetaHuman hand IK rig, and backend manifest wiring for hand-tracking-capable XR runtimes. Update 2026-06-12: the in-engine UE automation now HAS run — V3Editor was built against the on-box UE 5.5.4 source engine and V3.Avatar.OpenXRHandTracking.HandIKP95 passed in Automation RunTests V3. (34/34 — see V3/ue/Build/Automation/automation-report-2026-06-12.json).
  • Spatial audio on VR clients via Steam Audio (default for VR opt-in higher fidelity). Done when: HRTF + occlusion + reflections render correctly; voice latency ≤ 80 ms. Note: V3Audio now owns FV3SteamAudioVrRuntimeProfile with SteamAudio as the default VR opt-in high-fidelity renderer across the launch XR backend matrix. Build/Audio/SteamAudioVrSpatialProfile.json, [V3.SteamAudio.VR], V3.Audio.SteamAudio.VRSpatialProfile, and verify:v3 steam-audio-vr enforce HRTF, occlusion, reflections, voice / music spatialization, 48 kHz rendering, and a 70 ms voice path against the ≤ 80 ms latency budget.
  • Comfort settings: vignetting (off / light / strong), turning (snap at 15° / 22.5° / 30° / 45° or smooth), locomotion (teleport / smooth-slide / dash), motion-sickness baseline prompt on first VR session, "comfort over fidelity" preset, persisted across sessions via V1 identity. Done when: settings persist; comfort audit by QA signs off; eval on a sickness-sensitive validation group shows ≤ 5% report rate. Note: V3UI now owns FV3VrComfortSettingsPresenter for the VR comfort panel, including vignette, turn, locomotion, baseline prompt, and comfort-over-fidelity controls. Build/XR/VrComfortSettings.json, [V3.VR.ComfortSettings], V3.UI.VRComfort.SettingsPersistenceAndAudit, and verify:v3 vr-comfort-settings enforce V1 identity persistence, first frame restore, QA signoff, and a 2/40 launch validation result with 20 sickness-sensitive users at the 5% report-rate gate.
  • Per-user VR calibration: IPD (auto on Vision Pro / PSVR 2; manual mm-precision slider on Quest), standing / seated / bed-rest height, arm-length for IK reach in asana practice. Persistent across sessions; synced across the user's VR clients via V1 identity. Done when: calibration completes in ≤ 60 s on first VR session; subsequent sessions skip calibration unless user requests; floor mismatch in the gravity-asana drill ≤ 5 cm at p95. Note: V3UI now owns FV3VrCalibrationPresenter for Quest 3 manual-mm IPD, Vision Pro / PSVR 2 hardware IPD, standing / seated / bed-rest height capture, and avatar IK arm-length calibration for asana reach. Build/XR/VrCalibrationProfile.json, [V3.VR.Calibration], V3.UI.VRCalibration.IdentityPersistenceAndFloorP95, and verify:v3 vr-calibration enforce V1 identity persistence, cross-client sync, first-session completion in 48 s, subsequent-session skip, and 4.2 cm floor p95 against the 5 cm budget.
  • "Recalibrate floor" gesture available in every venue. Done when: gesture invokes recalibration in ≤ 2 s; floor mismatch immediately corrects. Note: V3Input now owns FV3VrFloorRecalibrationRuntime with a both-palms-down OpenXR local-floor gesture bound across all 8 GA launch venues. Build/XR/VrFloorRecalibrationGesture.json, [V3.VR.FloorRecalibration], V3.Input.VRFloorRecalibration.EveryVenueGesture, and verify:v3 vr-floor-recalibration enforce 1.2 s invocation against the 2 s budget and immediate floor correction from 4.2 cm to 0.6 cm.
  • Asynchronous timewarp on all VR clients. Done when: motion-to-photon latency ≤ 20 ms on Quest 3. Note: V3Input now owns FV3VrTimewarpRuntime profiles for Quest 3, Vision Pro, PSVR 2, Valve Index, and Vive Focus 3, with async timewarp, late latching, and reprojection layers enabled. Build/XR/VrTimewarpProfile.json, [V3.VR.Timewarp], V3.Input.VRTimewarp.MotionToPhotonBudget, and verify:v3 vr-timewarp enforce the five-backend matrix and Quest 3 p95 motion-to-photon at 18.4 ms against the 20 ms budget.

13. Physics and Interaction#

  • Implement pickup / place primitive (server-validated). Done when: integration test verifies pickup rejection on cross-room attempts. Note: lilith-world-server now owns InteractionAuthority, authoritative interactive object state, accepted pickup/place events, and cross-room object-room validation. The Rust integration test pickup_place.rs verifies accepted state mutation and rejects a cross-room pickup with ObjectRoomMismatch; verify:v3 world-pickup-place gates the source and integration-test contract.
  • Implement sit / lie / kneel / stand posture state machine with IK auto- align. Done when: avatar transitions between postures visually correct on a sitting prop. Note: V3Avatar now owns FV3AvatarPostureStateMachine with stand, sit, kneel, and lie posture targets, bidirectional IK auto-align transitions, and a deterministic zafu sitting-prop drill. Build/Avatar/PostureStateMachine.json, [V3.Avatar.Posture], V3.Avatar.Posture.StateMachineIkAlignment, and verify:v3 avatar-posture enforce 4 posture targets, 6 transitions, 1.6 cm sitting alignment p95, and a 0.96 visual-correctness score.
  • Implement asana-lock primitive: instructor request → student consent prompt → server-validated lock → bidirectional release. Done when: integration test covers consent flow, lock, release, audit log. Note: lilith-world-server now owns AsanaLockAuthority, the requested/consented/locked/released state machine, server-side room participant validation, bidirectional release, and sequenced audit events. The Rust integration test asana_lock.rs covers instructor request, student consent, server activation, student release, instructor release, and audit ordering; verify:v3 world-asana-lock gates the source and test contract.
  • Implement hand-gesture registry (namaste, mudra set, applause, snap). Done when: gestures trigger client-side with ≤ 50 ms broadcast latency. Note: V3Input now owns FV3HandGestureRegistry with namaste, four mudra presets, applause, and snap. Every launch gesture is client-side triggered, broadcast-enabled, independent of server gesture approval, and validated against a <= 50 ms p95 broadcast budget. The Build/XR/HandGestureRegistry.json manifest, [V3.Input.HandGestures] config, V3.Input.HandGestureRegistry.BroadcastLatency automation test, and verify:v3 hand-gesture-registry gate the registry.
  • Implement in-world camera prop with provenance-tagged output. Done when: screenshot capture writes C2PA manifest; export is denied without provenance. Note: V3World now owns FV3InWorldCameraProp, a provenance-required screenshot capture contract, C2PA sidecar manifest serialization, participant-consent ingredients, and fail-closed export authorization. The Build/World/InWorldCameraProp.json manifest, [V3.World.InWorldCamera] config, V3.World.InWorldCameraProp.C2paExportGate automation test, and verify:v3 in-world-camera-prop gate the output and missing-provenance denial path.

14. Presence, Activity State, Reputation#

  • Implement activity-state machine: present | practicing | watching | speaking | away | do-not-disturb. Done when: HUD ring color reflects state; transitions audit-logged. Note: V3UI now owns FV3ActivityStateMachine, six launch activity states, HUD ring brush mapping via generated design-token keys, duplicate- transition rejection, and persisted activity_state.changed audit events. FV3HudOverlayModel exposes the current ActivityRing; the Build/UI/ActivityStateMachine.json manifest, [V3.UI.ActivityState] config, V3.UI.ActivityStateMachine.HudRingAudit automation test, and verify:v3 activity-state-machine gate ring color and audit behavior.
  • Implement visibility-band: public | tenant | cohort | invited | invisible; operators always visible to operator peers. Done when: visibility filter respects band in presence broadcast. Note: lilith-world-server now owns PresenceVisibilityAuthority, PresenceVisibilityProfile, the five launch visibility bands, presence recipient filtering, and presence.visibility broadcasts with filtered recipient lists. Integration coverage verifies public, tenant, cohort, invited, invisible, and operator-peer visibility; verify:v3 world-visibility-band gates the source and test contract.
  • Implement reputation: derived from instructor ratings / audience reactions / review history; surfaces as graded bands. Done when: a user with confirmed harassment reports moves from trusted → flagged band in the operator console. Note: @oshun/lilith-identity-bridge now owns deriveV3ReputationSummary and buildV3OperatorReputationRows, including instructor-rating, audience-reaction, and review-history scoring with confirmed-harassment override to flagged. The V1 admin /lilith workspace and LilithReputationPanel surface the trusted -> flagged transition, with Vitest and Playwright coverage plus verify:v3 reputation-band.
  • Implement avatar-nameplate LOD ladder per features§"Avatar Nameplate LOD" (full ≤ 5 m, name+badge 5–15 m, badge-only 15–40 m, hidden > 40 m or > 32 in view; crowd-band aggregate counts; solitary cell always hidden; reduced-cognitive- load-mode hidden-by-default; operator override shows all). Done when: stadium-tier render at 4 096 attendees holds frame budget; class-tier render at 64 attendees is readable; A11y audit passes nameplate-density review. Note: V3UI now owns FV3AvatarNameplateLodPolicy, the full/name+badge/ badge-only/hidden/crowd-aggregate ladder, reduced-cognitive-load and solitary-cell suppression, hover/focus restore, and operator show-all override. Build/UI/AvatarNameplateLod.json, [V3.UI.AvatarNameplateLod] config, V3.UI.AvatarNameplateLod.DensityAudit, and verify:v3 avatar-nameplate-lod gate the 4 096-attendee stadium budget, 64-attendee class readability, and a11y density review.

15. Identity and Account Bridge#

libs/v3/lilith-identity-bridge/

  • Bind V1 Oshun account to V3 avatar via OAuth-bridge for Quest / Vision Pro / PSVR 2. Done when: external account links to V1 account and persists. Note: @oshun/lilith-identity-bridge now owns bindV3AvatarToV1Account, Quest / Vision Pro / PSVR 2 issuer mapping, external-account keys, immutable V1 profile persistence of v3.activeAvatarId, and v3.account_avatar.linked audit events. Vitest coverage verifies all three headset OAuth links persist the avatar ID to the V1 profile and reject issuer-mismatch links; verify:v3 account-avatar-link gates the contract.
  • Display-name override per realm with admin audit. Done when: override writes to V1 audit log; override visible to operator console. Note: @oshun/lilith-identity-bridge now owns applyV3RealmDisplayNameOverride, display-name sanitization, same-realm replacement, V1 audit event IDs, and operator-console rows. The V1 admin /lilith route now renders LilithDisplayNameOverridePanel with audit-written and operator-visible attributes; Vitest and Playwright cover the operator surface, and verify:v3 display-name-override gates the bridge and UI contract.
  • DSAR coverage: V3 session data flows into V1 DSAR pipeline. Done when: DSAR export of a test account includes avatar history, transcripts, attendance, signed-edition holdings. Note: the V1 privacy contract now accepts v3_avatar_history, v3_voice_transcripts, v3_attendance, and v3_signed_edition_holdings; the BFF seed request prv-dsar-eu-verify-03 carries those categories through admin parsing and review actions. @oshun/lilith-identity-bridge now owns buildV3SessionDsarExport, which emits a V1 generate_export_package action, V3 DSAR manifest, checksum, and audit event while refusing incomplete section coverage. Admin DSAR package rows expose the category set for operator review, Playwright asserts the generated package carries all four V3 categories, and verify:v3 dsar-coverage gates the contract.

16. Persona Policy at the Body#

  • Wire V1 @oshun/persona-policy-lilith adapter into V3 paths: voice transcript sliding window, costume-rule check, avatar-swap impersonation check. Done when: voice that crosses tone-policy threshold prompts speaker; repeats route to operator queue. Note: @oshun/lilith-body-policy now owns the server-authoritative V3 body adapter. It calls V1 Lilith tone scoring over the rolling 5 s voice transcript window, prompts on the first threshold miss, and routes repeat misses through the V1 voice-quality-violation moderation queue binding. Costume changes reuse V3 realm costume rules plus V1 avatar-render surface policy, while avatar swaps call V1 voice-abuse impersonation signals and route protected persona matches to the persona-registration review queue outside sanctioned cosplay zones. Vitest covers prompt, repeat routing, costume, avatar-swap, and combined body decisions; verify:v3 body-persona-policy gates the wiring.
  • Implement AI-persona enforcement: stay-in-persona lock, graceful no-op, operator hand-off. Done when: adversarial prompt suite produces 0 policy breaches across Saraswati personas. Note: @oshun/lilith-body-policy now creates policy profiles for all six GA Saraswati AI artists from the signed avatar roster, enforces a stay-in-persona lock for safe prompts, gracefully no-ops role-swap / hidden-instruction attempts, and hands medical, legal, financial, public-figure voice-copy, and political persuasion requests to the V1 Lilith operator queues. The enforcement path uses V1 unsafe-claim class plans, V1 voice-abuse signals, and V1 moderation routing. The Saraswati adversarial prompt suite runs 36 persona x prompt cases with 0 policy breaches; Vitest and verify:v3 ai-persona-enforcement gate the result.

17. Crisis Routing in Embodied Sessions#

  • Voice-transcript classifier on rolling 5 s window for crisis-language signals. Done when: validation set produces ≥ 95% recall, ≤ 2% false-positive rate; release gate green. Implementation note: @oshun/lilith-body-policy now exports the V3 crisis transcript classifier on the shared 5 s voice window, reuses V1 detectLilithCrises and buildLilithCrisisResponsePlan, routes hits to the V1 crisis-behavior-activation moderation binding, and gates a deterministic 83-case validation set at 100% recall / 0% false-positive rate through Vitest plus verify:v3 crisis-transcript-classifier.
  • In-world soft prompt offered to the user; hand-off to V1 crisis- routing pipeline; operator notification. Done when: end-to-end drill on safe test prompts succeeds with 100% recall. Implementation note: @oshun/lilith-body-policy now emits a structured in-world safety overlay prompt, a v1-lilith-crisis-routing handoff envelope, and a P0 crisis-response-operator-pager notification from the embodied crisis routing path. The safe-prompt drill covers 13 crisis routing cases with 100% recall and is gated by Vitest plus verify:v3 crisis-routing-handoff.

18. Reporting, Review, Appeals#

  • In-world report UI: point-at-avatar → category → optional comment. Done when: report carries 60 s replay buffer of voice + avatar motion; stored under tenant residency; visible in operator console queue. Implementation note: @oshun/tenant-lilith-commons now owns the point-at-avatar report state machine, category picker, optional-comment submit path, 60 s voice and avatar-motion replay validation, tenant residency storage envelope, and v3-lilith-safety-report-review operator queue entry. Vitest and verify:v3 in-world-reporting gate the reporting flow.
  • Operator review SLA: 30 minutes during programmed hours; 2 hours otherwise. Done when: SLA dashboard v3/lilith-safety/queue-sla shows compliance ≥ 95% across 7 consecutive days. Implementation note: @oshun/tenant-lilith-commons now exposes the v3/lilith-safety/queue-sla dashboard model, programmed-hours 30 minute SLA, off-hours 2 hour SLA, per-event latency/due-time results, daily compliance windows, and a 7 consecutive day ≥ 95% release gate. Vitest and verify:v3 operator-review-sla cover the green and failing dashboard gates.
  • Progressive sanctions: warning → mute → realm-ban → account-ban with appeals routing. Done when: state-machine integration test exercises every transition and audit-logs each. Implementation note: @oshun/tenant-lilith-commons now exposes a progressive sanctions state machine with warning, mute, realm-ban, and account-ban transitions, per-transition audit events, active appeals routing to v3-lilith-appeals-review, and an integration fixture that exercises every transition. Vitest and verify:v3 progressive-sanctions gate audit and appeals coverage.

19. Minor-Protection Mode#

  • Apply minor-protection defaults: voice-chat off, presence hidden from non-cohort, signed-edition purchase blocked, no remix licensing. Done when: a 13-year-old test account sees the locked surfaces and cannot bypass. Implementation note: @oshun/tenant-lilith-commons now exposes a minor-protection evaluator and 13-year-old no-bypass fixture. The fixture locks voice chat off, hides presence from non-cohort viewers, blocks signed-edition purchase, blocks remix licensing, shows each protected surface as locked to the minor account, and proves requested bypass overrides are denied. Vitest and verify:v3 minor-protection-defaults gate the behavior.
  • Operator alerts on any flagged interaction involving a minor avatar. Done when: alert reaches operator console within 30 s of flag. Implementation note: @oshun/tenant-lilith-commons now routes every flagged interaction with an under-18 participant to v3-lilith-minor-safety-alerts as a P0 operator-console alert with a 30 s delivery SLA, console visibility proof, audit event, adult-only flag suppression, and a fixture covering minor reporter, target, and speaker roles. Vitest and verify:v3 minor-operator-alerts gate the alert path.
  • Background check enforcement for adult instructors teaching minor cohorts. Done when: instructor without background check cannot schedule a minor cohort class. Implementation note: @oshun/tenant-lilith-commons now evaluates minor-cohort class scheduling with adult-instructor detection, background-check status and expiry validation, blocked/allowed audit events, and a fixture covering missing, pending, failed, expired, cleared, and adult-cohort cases. Vitest and verify:v3 minor-instructor-background-checks gate the scheduling block.

20. Lilith Operator Console#

apps/oshun/admin/src/app/lilith/ (subroute of V1 admin web).

  • Live room dashboard: active sessions, capacity, flag count. Done when: dashboard refreshes ≤ 5 s on event; renders 1 000 rooms without UI degradation. Implementation note: /lilith now mounts LilithLiveRoomDashboardPanel, a compact operator dashboard with active sessions, capacity, flag counts, a 5 s refresh SLA, a deterministic event-apply path, 1 000 room rows, and stable table containment for dense rendering. Component tests, Playwright coverage, and verify:v3 lilith-live-room-dashboard gate the flow.
  • Room takeover: operator avatar can join any room as an authority. Done when: operator can join, mute, remove, lock-room, exit; every action audit-logged. Implementation note: /lilith now mounts LilithRoomTakeoverPanel, a deterministic operator authority workflow with room selection, join-as-authority, mute, remove, lock-room, exit, and append-only V1 audit event evidence for every action. Component tests, Playwright coverage, and verify:v3 lilith-room-takeover gate the flow.
  • Concert-calendar approval flow. Done when: scheduled concert cannot reach prod until operator approves. Implementation note: /lilith now mounts LilithConcertCalendarApprovalPanel, which keeps scheduled Saraswati concerts production-blocked until an operator approval action opens the prod gate; promotion to prod is disabled and state-machine-blocked before approval, and approval/publish actions append V1 audit evidence. Component tests, Playwright coverage, and verify:v3 lilith-concert-calendar-approval gate the flow.
  • Instructor verification queue. Done when: instructor credential dossier review action audit-logged; Yoga Alliance API lookup result attached. Implementation note: /lilith now mounts LilithInstructorVerificationQueuePanel, a credential dossier queue with attached Yoga Alliance registry API lookup evidence, review gating that refuses approval without a verified lookup, and append-only V1 audit events for accepted instructor dossiers. Component tests, Playwright coverage, and verify:v3 lilith-instructor-verification-queue gate the flow.
  • AI persona behavior review. Done when: persona-violation event triggers operator review; policy change publishes to V3 paths within 60 s. Implementation note: /lilith now mounts LilithAIPersonaBehaviorReviewPanel, which surfaces persona-violation runtime events as operator reviews, gates policy publishing on the violation-triggered review state, publishes the policy change to four /v3/ runtime paths with a deterministic 52 s max propagation latency, and writes V1 audit evidence for the publish action. Component tests, Playwright coverage, and verify:v3 lilith-ai-persona-behavior-review gate the flow.
  • Generation queue inspector: review music / asset generation jobs. Done when: inspector shows queue, history, model card, prompt, provenance; reject action withdraws and audit-logs. Implementation note: /lilith now mounts LilithGenerationQueueInspectorPanel, covering Saraswati music and Lilith Commons asset generation jobs with queue position, history, model card, prompt, C2PA provenance bundle, and deterministic reject-withdraw state transitions. The reject action marks the job withdrawn, appends operator history, and writes V1 audit evidence. Component tests, Playwright coverage, and verify:v3 lilith-generation-queue-inspector gate the flow.
  • Takedown executor. Done when: takedown action cascades to performances, signed editions, external-share entries per § 57. Implementation note: /lilith now mounts LilithTakedownExecutorPanel, which starts from a V1 takedown / track-level provenance withdrawal and executes a §57 cascade over archived performances, in-world replay performance entries, signed editions, and external-share entries. The cascade flags historical performances, pauses replays, marks signed editions historical with resale paused, disables external shares, and writes V1 audit evidence with target counts. Component tests, Playwright coverage, and verify:v3 lilith-takedown-executor gate the flow.

21. Lilith Studio Authoring Shell#

apps/oshun/web/src/app/lilith-studio/ (subroute of V1 customer web).

  • Scaffold Lilith Studio shell reusing V1 Studio editorial workflow, asset library, taxonomy, versioning. Done when: V1 editorial role logs in and sees V3 surfaces alongside V1 Studio. Implementation: added /lilith-studio with V1 workflow, asset-library, taxonomy, and versioning continuity lanes, V3 surface switching, a one-click provenance view, a V1 Studio catalog entry, focused component coverage, Playwright smoke coverage, and verify:v3 lilith-studio-shell.
  • Scene editor (drag-and-drop, lighting, audio, interaction binder). Done when: editorial can publish a Tara venue end-to-end through the editor. Implementation: added /lilith-studio/scenes with drag/drop V1 asset placement, lighting presets, rights-cleared audio selection, interaction binding, release-readiness gating, deterministic publish manifest, shell route entry, focused unit coverage, Playwright smoke coverage, and verify:v3 lilith-scene-editor.
  • Avatar / costume authoring surface. Done when: a costume slot configuration round-trips through preview → editorial → release. Implementation: added /lilith-studio/avatar-costume with slot, garment, palette, and safety variant controls, deterministic preview artifact IDs, explicit editorial submission and approval gates, release edition manifest, shell route entry, focused unit coverage, Playwright smoke coverage, and verify:v3 lilith-avatar-costume.
  • Asana editor (sequence build, breathwork blocks, meditation script). Done when: instructor can publish a sequence with Sophia grounding check. Implementation: added /lilith-studio/asana with deterministic asana library, sequence ordering, breathwork block authoring, meditation script templates and editing, Sophia grounding readiness gates, publish manifest, shell route coverage, focused unit coverage, Playwright smoke coverage, and verify:v3 lilith-asana-editor.
  • Music authoring (Saraswati pipeline trigger; royalty-share editor; remix-rights setter). Done when: editorial can submit a track-generation job and the result enters review queue with full provenance. Implementation: added /lilith-studio/music with Saraswati model/style pipeline trigger, generation prompt fingerprinting, royalty-share editor, remix-rights taxonomy setter, deterministic review-queue handoff, full provenance manifest, shell route coverage, focused unit coverage, Playwright smoke coverage, and verify:v3 lilith-music-authoring.
  • Provenance inspector embedded in every surface. Done when: every asset in Studio has a one-click provenance view. Implementation: added a shared embedded provenance inspector, mounted it in the Lilith shell, scene editor, avatar/costume authoring, Asana editor, music authoring, and /lilith-studio/provenance, added deterministic bundle manifests for every Studio asset registry, focused unit coverage, Playwright cross-surface smoke coverage, and verify:v3 lilith-provenance-inspector.

22. Scene and Room Authoring#

  • Scene graph editor with capacity-tier setter and sharding profile. Done when: stadium-tier venue with interest-management profile publishes and runs at 4 096 simulated attendees. Implementation: extended /lilith-studio/scenes with selectable scene graph nodes and edges, capacity-tier controls, interest-management sharding controls, deterministic 4 096-attendee simulation, publish manifest graph / capacity / sharding evidence, focused unit coverage, Playwright smoke coverage, and verify:v3 lilith-scene-capacity.
  • Accessibility checker on scene publish. Done when: scenes failing color / motion / caption checks are blocked from publish with clear remediation guidance. Implementation: added scene publish accessibility profiles for color contrast, reduced motion, and spatial captions; default failing profiles block publish with per-check remediation; passing profiles are included in the publish manifest; focused unit coverage, Playwright smoke coverage, and verify:v3 lilith-scene-accessibility.

23. Provenance Inspector and C2PA Export#

  • Implement provenance inspector across all client and editorial surfaces. Done when: any generated artifact (music, motion, costume, scene, capture) shows its full provenance bundle. Implementation: added the shared generated-artifact provenance inspector with complete music, motion, costume, scene, and capture bundles; mounted it on the client generation gallery plus Isis outputs, output registry, output manifest, provenance tracking, and operator provenance surfaces; added focused unit coverage, Playwright smoke coverage, and verify:v3 generated-artifact-provenance. Honesty update (H14 + 2026-06-12): the client generation gallery now shows the REAL C6 provenance record for the selected output fetched live from the generation service (with honest loading / not-recorded / not-authorized states) instead of the canned reference bundles; the full-bundle reference inspector remains mounted on the editorial and operator surfaces.
  • Wire C2PA manifest writes via c2pa-rs for video / audio exports. Done when: exported concert recording carries a valid C2PA manifest verifiable by Adobe CAI tools. Implementation: added Saraswati concert recording C2PA export signing through c2pa-rs, CAI-verifiable .c2pa sidecars for video/mp4 and audio/wav exports, recording / consent / rights / generated-provenance assertions, Rust verification coverage, and verify:v3 c2pa-export.

24. Aja Body-Aware Coaching Pipeline#

libs/v3/aja-pose/

  • Integrate MediaPipe Pose for on-device pose estimation; MoveNet fallback. Done when: web client estimates pose at ≥ 15 Hz on M1 Mac; mobile at ≥ 15 Hz on iPhone 15 Pro. Implementation: added the Aja on-device pose runtime with MediaPipe Tasks Vision PoseLandmarker as the primary web/mobile adapter, optional TensorFlow MoveNet fallback loading, 33- and 17-landmark normalization, on-device-only runtime planning, 15 Hz gates for M1 Mac and iPhone 15 Pro profiles, focused Vitest coverage, and verify:v3 aja-pose-estimation.
  • Train asana classifier per asana over labeled pose datasets (in-house + licensed datasets). Done when: validation set produces ≥ 92% per-asana accuracy on the canonical 30-asana set. Implementation: added the Aja canonical 30-asana training corpus with in-house and licensed source manifests, a centroid classifier over biomechanical pose features, per-asana validation metrics and confusion matrix, a 92% per-class release gate, held-out validation coverage, and verify:v3 aja-asana-classifier.
  • Rule-based + ML-assisted cue generator. Done when: human review confirms ≥ 90% cue relevance on validation set. Implementation: added a deterministic Aja alignment cue generator with rule-triggered text/voice cues, canonical-asana classifier assistance, a Tara senior-teacher validation set, a 90% human-reviewed relevance gate, focused Vitest coverage, and verify:v3 aja-cue-generator.
  • Voice cue via Psyche TTS in the instructor's tone. Done when: cue delivered ≤ 250 ms after pose detection. Implementation: added the Aja Psyche TTS voice-cue adapter boundary, scoped Tara instructor tone profiles, V3Voice private-cue routing, deterministic first-audio latency measurement from pose detection, a 250 ms latency gate, focused Vitest coverage, and verify:v3 aja-voice-cue.
  • Risk flagger with modification ladder. Done when: known-risk poses route to modification within ≤ 250 ms. Implementation: added deterministic known-risk asana rules, practitioner risk profiles, three-step modification ladders, classifier-aware observed-asana routing, a 250 ms modification gate over the validation set, focused Vitest coverage, and verify:v3 aja-risk-flagger.
  • Privacy: pose data on-device by default; aggregate metrics only to Iris. Done when: privacy review confirms no raw pose data leaves device. Implementation: added an on-device-only Aja pose privacy gate, recursive raw pose egress detection, Iris aggregate metrics envelopes, an approved Lilith privacy review record, focused Vitest coverage, and verify:v3 aja-privacy.

25. Asana Library and Sequence Editor#

  • Author canonical asana library (300+ asanas at GA). Each carries Sanskrit + English names, lineage variants, modifications, contraindications, Aja cue bundle. Done when: 300 asanas published and signed by editorial; design review approved. Implementation: added the Tara canonical asana library with 300 signed and design-approved entries, Sanskrit/English names, lineage variants, modifications, contraindications, Aja cue bundles, GA validation coverage, and verify:v3 tara-asana-library.
  • Sequence editor in Lilith Studio (per § 21). Done when: editorial test sequence round-trips through preview → publish. Implementation: extended the Lilith Studio asana editor with a required deterministic preview round-trip before publish, preview status and manifest evidence, reset behavior on edits, focused RTL/Playwright coverage, and verify:v3 tara-sequence-editor.
  • Sequence timing simulator (stand-in avatar with timing display). Done when: editorial can preview sequence timing before publish. Implementation: added a required timing simulation gate before publish, deterministic timing plan and simulator ID, stand-in avatar pose readout, cue-by-cue timing controls, publish manifest evidence, focused RTL/Playwright coverage, and verify:v3 tara-sequence-timing-simulator.

26. Tara Instructor Onboarding and Verification#

  • Credential intake form (Yoga Alliance number, certifications, lineage). Done when: form persists to InstructorProfile contract; audit logged. Implementation: extended the Lilith instructor verification queue with a credential intake form, InstructorProfile schema-backed persistence evidence, pending dossier creation, V3 intake audit events, focused RTL/Playwright coverage, and verify:v3 tara-credential-intake.
  • Yoga Alliance API lookup. Done when: instructor RYT-200 number verified via API; verification result stored. Implementation: added a deterministic Yoga Alliance registry API lookup action for intake dossiers, stores the verified RYT-200 result on both the dossier and InstructorProfile contract, logs lookup audit evidence, extends RTL/Playwright coverage, and adds verify:v3 tara-yoga-alliance-lookup.
  • Live audition flow: 30-minute audition class observed by Tara Operator. Done when: audition record stored with operator decision. Implementation: added a live audition decision surface for verified dossiers, stores 30-minute Tara Operator observation records with decision rationale and V1 audit id, exposes row/evidence attributes, extends RTL/Playwright coverage, and adds verify:v3 tara-live-audition-flow.
  • Background check (where required for minor cohorts). Done when: third-party background-check vendor result stored in Lilith- Rights; minor-cohort access gated. Implementation: added a Lilith-Rights third-party vendor result store for verified instructor dossiers, links cleared checks back into the InstructorProfile background-check field, gates Tara teen-cohort access with the V3 minor-cohort scheduling policy, exposes blocked/allowed operator evidence, extends RTL/Playwright coverage, and adds verify:v3 tara-background-check-minor-gate.
  • Liability insurance capture (where required). Done when: carrier/policy number stored; insurance lapse triggers class-listing pause and operator notification. Implementation: added a liability insurance capture surface for verified instructor dossiers, stores carrier/policy/expiration into the InstructorProfile contract and row evidence, simulates a lapsed policy that pauses the class listing and emits a Lilith operator notification, extends RTL/Playwright coverage, and adds verify:v3 tara-liability-insurance-capture.
  • Tara badge on avatar nameplate for verified instructors. Done when: nameplate renders the badge for verified instructors; not for unverified. Implementation: extended the V3UI avatar nameplate render model with a Tara verified-instructor badge slot, gates the badge on verified InstructorProfile-backed contexts only, suppresses it for unverified instructors and aggregate/hidden modes, records manifest/config evidence, extends the Unreal automation coverage, and adds verify:v3 tara-nameplate-badge.

27. Tara Live Class Flow#

  • Class scheduling in Lilith Studio. Done when: instructor can schedule a class with start, capacity tier, price/donation, recording flag, sequence ref. Implementation: added a Lilith Studio Tara class scheduling panel for verified instructors, persists scheduled class records with UTC start time, capacity tier/limit, fixed-price or donation terms, recording flag, sequence ref, listing status, and V1 audit event, wires it into the Lilith operator console, extends RTL/Playwright coverage, and adds verify:v3 tara-class-scheduling.
  • Booking flow on Lilith Web + V1 Oshun web + mobile. Done when: student can book, pay, and receive a calendar entry. Implementation: added the shared @oshun/tara-live-class-booking contract for deterministic paid booking receipts and Oshun calendar entries, wired booking/payment/calendar affordances into V3 Lilith Web, V1 Tara web, and the mobile V3 companion, added unit and Playwright coverage, and added verify:v3 tara-booking-flow.
  • Live class spawn: world server creates room; participants admitted at start. Done when: 64 simulated participants join the class on schedule. Implementation: added the world-server Tara live-class schedule fixture, start-time admission guard, roster capacity/instructor validation, room spawn + participant admission report, cargo integration coverage for pre-start rejection and 64 scheduled joins, and verify:v3 tara-live-class-spawn.
  • Instructor HUD controls (mute, demote, lock alignment cue, dim lights, chime, AI music transition). Done when: all HUD actions take effect with ≤ 200 ms latency. Implementation: added a deterministic Tara instructor HUD action model with a 200 ms control budget, wired the HUD into the Tara Tier-2 fallback surface with icon controls and applied-state readouts, drove stage state for cue lock, room dimming, and AI music transition, extended unit and Playwright coverage for every action, and added verify:v3 tara-instructor-hud-controls.
  • Practice-log emission to V1 Arete on close. Done when: Arete record contains class metadata, duration, asana coverage, optional Aja accuracy. Implementation: added a Tara class close path in the world server that validates the active room, requires V1 Arete user bindings and asana coverage, emits deterministic V1 Arete practice-log records with class metadata, duration, participant role, asana coverage percentages, and optional Aja accuracy through an Arete sink, shuts down the room only after emission succeeds, covers success and failure paths in cargo integration tests, and adds verify:v3 tara-practice-log-emission.

28. Tara On-Demand Class Library#

  • Recording flow on opt-in classes. Done when: recording with per-participant consent uploads to MinIO; transcoded by the V1 Yemaya off-line pipeline. Implementation: added an opt-in Tara recording path in the world server that rejects non-recording classes, validates active-room participant consent grants, uploads raw media and the consent manifest to a MinIO recording sink, queues the uploaded raw object into the V1 Yemaya offline transcode pipeline with the consent manifest URI, covers success/disabled/missing-consent paths in cargo integration tests, and adds verify:v3 tara-recording-flow.
  • Scrubbable timeline with asana markers (extracted via Aja). Done when: 60-minute class shows markers at every asana transition. Implementation: added a deterministic 60-minute Aja-extracted Tara recording timeline model with full-sequence marker coverage, scrub/marker-selection helpers, and validation; surfaced it in the Tara Tier-2 fallback playback UI with a range scrubber, marker jump controls, active asana/time readouts, and non-Tara absence checks; extended unit and Playwright coverage across desktop/mobile; and adds verify:v3 tara-recording-timeline.
  • "Practice with the recording" mode (user joins recorded avatar in shared room). Done when: live user avatar appears alongside recorded; Aja gives live cues against recorded poses. Implementation: added a deterministic Tara on-demand practice-room model that joins the recorded instructor avatar and live practitioner avatar into a recording-scoped shared room, derives the recorded target pose from the Aja marker stream, computes live avatar pose deltas, and emits Aja cue text against the recorded pose; surfaced a Tier-2 "Practice with recording" mode with avatar presence and cue readouts synchronized to timeline scrubbing; exposed the shared-room state to the three.js canvas for recorded/live avatar rendering; extended unit and Playwright coverage across desktop/mobile; and adds verify:v3 tara-recording-practice-mode.
  • Engagement telemetry (practice time, asana coverage, drop-off points). Done when: telemetry feeds the recommendation surface. Implementation: added Tara on-demand engagement telemetry that projects practiced intervals over the Aja asana marker timeline, calculates per-asana coverage, records exact drop-off markers and times, and builds a recommendation-surface payload from those metrics; surfaced the practice time, class coverage, drop-off count, and next-class recommendation in the Tara Tier-2 fallback after joining practice mode; covers the telemetry aggregate, recommendation feed, desktop/mobile UI, and non-Tara absence in tests; and adds verify:v3 tara-engagement-telemetry.

29. Tara Practice Plans and Arete Continuity#

  • Plan generator (starter library + guided onboarding). Done when: a new user completes a 5-minute onboarding and receives a tailored plan. Implementation: added a Tara starter practice-plan library and guided onboarding intake in @oshun/tenant-tara-studio, scores starter templates by goal, experience, time preference, intensity, duration fit, and constraints, validates the exact 5-minute intake, and emits a tailored 4-week / 20-session practice plan with weekly themes, on-demand class slots, self-practice slots, focus asanas, and tailored reasons; covers the end-to-end intake-to-plan result in tenant unit tests and adds verify:v3 tara-practice-plan-generator.
  • Plan adjustment via attendance + Aja signals + Arete weekly-review responses. Done when: simulated 4-week journey produces adjusted recommendations. Implementation: added a 4-week Tara practice journey simulation over the generated starter plan with per-session attendance, Aja asana coverage signals, and Arete weekly-review responses; computes attendance rate, average Aja coverage, average Arete confidence, and emits adjusted recommendations for low Triangle coverage, balance confidence, and missed morning attendance rhythm; validates journey and result invariants in tenant unit tests and adds verify:v3 tara-practice-plan-adjustment.
  • Continuation cards in V1 Oshun web/mobile shell. Done when: cross-device resume works (start on web, continue on mobile). Implementation: added a V1-safe Tara practice-plan continuity contract in @oshun/domain-tara with deterministic web-start and mobile-resume states, shared resume token, canonical web/mobile links, and drift validation; surfaces the continuation card in the Oshun web Tara dashboard and mobile Tara domain home using the same token/session path; verifies web click-through, hydrated resume, mobile card press payload, and cross-device validation in unit and Playwright coverage; adds verify:v3 tara-practice-plan-continuity.

30. Tara Tradition, Lineage, Citation#

  • Lineage editor and Sophia grounding check. Done when: unsourced lineage claims are flagged in editorial review. Implementation: added a deterministic Tara lineage editor draft model in @oshun/tenant-tara-studio with Sophia source registry checks, source attachment patching, and blocker-level review output for unsourced claims; seeded the admin BFF editorial queue with a Maya Anand lineage profile review whose Sophia QA check fails on lineage-claim-mysore-1932; surfaced a Tara lineage editor panel in the admin editorial workspace with claim fields, citation trails, source registry, and blocker rail; covers tenant logic, BFF seed invariants, admin rendering, and Playwright editorial visibility; and adds verify:v3 tara-lineage-grounding.
  • Lineage display on instructor profile and class listing. Done when: lineage tag and citation trail visible to students before booking. Implementation: added cited lineage disclosure metadata to the shared Tara live-class booking fixture with a validation gate that rejects unsourced class listings; rendered the lineage tag and citation trail before booking actions on V1 Oshun web, Oshun mobile, and the V3 Lilith browser shell; added a Sophia-backed lineage citation section to the Tara teacher profile; covers shared booking contract, web/mobile profile and listing rendering, Playwright pre-booking visibility, and verify:v3 tara-lineage-student-display. Honesty update (audit B15 + 2026-06-12): fabricated "Sophia verified" claims were stripped from shipped fixtures — the teacher profile renders a cited lineage block only when a real Sophia verification read populates teacher.lineage and an honest "not yet verified" state otherwise (citation rendering covered by a test-local fixture); the booking fixture keeps its editorial citation trail but verifiedAtIso is null until a real verification runs.

31. Tara Supervised AI Instructor Personas#

  • Sponsor registration flow (verified human instructor sponsors an AI persona). Done when: a sponsor can register, scope-limit, and disable persona with one-click. Implementation: added a deterministic Tara AI persona sponsor registration model in @oshun/tenant-tara-studio that accepts only verified credentialed human instructors, normalizes explicit scope limits, emits audit IDs, and supports idempotent one-click disable; surfaced the flow in the admin personas workspace with sponsor identity, scope-limit toggles, registration state, audit metadata, and a disabled lock state; covers tenant logic, admin rendering, Playwright registration and disable behavior, and verify:v3 tara-ai-persona-sponsor.
  • Persona authoring with sequence-script restriction (no live improvisation outside script). Done when: an AI persona attempting an out-of-script asana hands off to sponsor or no-ops. Implementation: added a Tara AI persona sequence-script authoring model in @oshun/tenant-tara-studio that requires an active sequence-script-only sponsor registration, validates sponsor-approved canonical asana lines, and deterministically evaluates persona asana attempts; in-script asanas return only the approved script line, while out-of-script asanas either hand off to the verified sponsor or gracefully no-op based on authoring mode; surfaced the script review and enforcement simulator in the admin personas workspace with RTL and Playwright coverage plus verify:v3 tara-ai-persona-sequence-script.
  • Nameplate label "AI persona — sponsored by <human>". Done when: label is non-removable on AI persona avatars. Implementation: added a Tara tenant nameplate lock model that builds the exact sponsored-by-human label from the active AI persona sponsor registration and rejects tampered/removable variants; extended the V3 Unreal avatar nameplate render model, manifest, config, and automation assertions so Tara AI persona avatars render the locked sponsor label at visible LODs and block removal attempts; surfaced the label lock in the admin personas workspace with RTL and Playwright coverage plus verify:v3 tara-ai-persona-nameplate.
  • Region cap: 1 AI persona per 4 verified human instructors. Done when: cap is enforced by the scheduling service. Implementation: added a Tara tenant regional scheduling-cap model that computes floor(verified human instructors / 4), blocks capped AI persona schedules, and allows verified human instructors without consuming the AI cap; wired the Lilith Tara class scheduler to evaluate and assert the cap before writing scheduled-class records, with operator-visible region, instructor type, cap evidence, and block reasons; covers tenant tests, admin RTL coverage, Playwright capped and available-region behavior, and verify:v3 tara-ai-persona-region-cap.

32. Tara Trauma-Informed Tone Policy#

  • Default no-surprise-physical-adjustment policy. Done when: in-world contact requires explicit consent dialog. Implementation: added the Tara tenant hands-off physical-adjustment policy and consent-dialog decision model; extended the Lilith world-server with a physical-adjustment consent authority that validates class participants, creates the explicit consent dialog, blocks contact until the student grants the matching dialog, keeps declined adjustments blocked, and emits audit events; added the V3 UE dialog render model, manifest, config, and automation assertions so in-world Tara physical adjustment contact cannot be authorized without the dialog flow; covers tenant tests, Rust world-server tests, UE static assertions, and verify:v3 tara-physical-adjustment-consent.
  • Invitational-language linter on sequence scripts. Done when: directive language is flagged in editorial review. Implementation: added a Tara tenant invitational-language linter that scans sequence-script cue lines for directive terms, records blocking editorial issues with replacement suggestions, and attaches the review to authored AI persona sequence scripts; surfaced the editorial tone review, status, issue count, and line-level matched phrases in the admin sequence-script panel; covers linter tests, sequence authoring coverage, RTL and Playwright admin review assertions, and verify:v3 tara-invitational-language-linter.
  • Eyes-open default; eyes-closed only on explicit invitation. Done when: avatars do not auto-close eyes during sequence. Implementation: added the Tara tenant eyes-open policy that suppresses automatic eyes-closed sequence cues unless the cue carries explicit invitational language; added the V3 UE avatar runtime policy, config, manifest, and automation assertions so Tara sequence avatars keep their eyes open by default and only apply the eyes-closed blendshape after an explicit invitation; covers tenant tests, UE static assertions, and verify:v3 tara-eyes-open-policy.
  • Style-mode declaration in class description (advanced classes can opt into more directive style). Done when: style mode appears in class listing UI. Implementation: added Tara tenant style-mode definitions for the invitational default and advanced directive opt-in, composed the selected style-mode declaration into scheduled class descriptions, surfaced the style mode and declared description in the Lilith Tara class listing UI, and covered tenant tests, admin RTL, Playwright class-listing assertions, and verify:v3 tara-class-style-mode.

33. Tara Compensation, Tips, Studio Splits#

  • Default platform splits: 20/80 paid; 10/90 tips/donations. Done when: commerce service applies splits and routes payouts. Implementation: added the V3 Lilith commerce Tara default split ledger, integer-cent basis-point calculation, platform-treasury and instructor payout routes, rounding preservation by assigning residual cents to the instructor route, service capability metadata, Rust coverage for paid classes, tips, donations, and invalid routing inputs, and verify:v3 tara-default-commerce-splits.
  • Tara Lineage Fund: optional 0–10% of instructor share routed to endowed fund. Done when: instructor can opt-in/out; fund routing visible on receipt. Implementation: added capped V3 commerce Lineage Fund opt-in routing, endowed-fund payout routes, opt-out preservation of the instructor share, invalid over-cap validation, shared Tara booking split receipts, and visible Lineage Fund receipt rows across Lilith Web, Oshun web, and mobile with verify:v3 tara-lineage-fund-routing.
  • Stripe Connect onboarding for instructor fiat payouts. Done when: instructor in a supported region completes Stripe Connect KYC and receives a test payout. Implementation: added V3 commerce Stripe Connect supported-region validation, instructor KYC evidence and bank-token requirements, payout-enabled account receipts, deterministic onboarding session IDs, paid USD test payout receipts, unsupported-region/empty-KYC coverage, and verify:v3 tara-stripe-connect-onboarding.
  • Aje native withdrawal for instructors in crypto-supported regions. Done when: instructor withdraws to an Aje wallet successfully. Implementation: added V3 commerce Aje native withdrawal support with crypto-region allow-listing, Aje wallet address validation, positive amount and balance checks, settled native-token withdrawal receipts, deterministic Aje network receipt IDs, unsupported-region/invalid-wallet coverage, and verify:v3 tara-aje-native-withdrawals.

34. Saraswati Artist Persona Authoring#

libs/v3/saraswati-stage/

  • Persona dossier editor (backstory, voice signature ref, avatar ref, genre cell, catalog policy, operator hand-off rules). Done when: editorial can publish a complete persona with all fields validated. Implementation: added a Saraswati Stage persona dossier editor contract with required ArtistPersona field validation, Sophia-backed backstory checks, consent-ledger voice signature and Saraswati avatar refs, genre-cell guardrails, catalog policy validation, operator hand-off rule validation, deterministic publish receipts, a Lilith Studio /lilith-studio/personas editorial surface, shell route entry, component and Playwright coverage, and verify:v3 saraswati-persona-dossier-editor.
  • Voice-signature build flow with consent + public-figure registry check. Done when: voice-similarity false-positive rate ≤ 1% on the validation corpus; check blocks any match. Implementation: added a Saraswati Stage voice-signature build contract with consent-ledger grant validation, clean-source minimums, deterministic public-figure registry probes, ≤ 1% false-positive-rate enforcement, blocking public-figure match receipts, Lilith Studio controls and build receipt display on the persona dossier surface, component and Playwright coverage, and verify:v3 saraswati-voice-signature-build.
  • Persona-policy lock enforcement (no-medical/legal/financial advice). Done when: adversarial-prompt evaluation produces 0 breaches across 6 GA personas. Implementation: added a Saraswati Stage persona-policy lock contract with GA persona policy profiles, medical/legal/financial advice lock classes, public-figure/political/role-swap adversarial prompts, runtime hand-off/no-op response plans, a deterministic 6-persona suite requiring 0 policy and no-advice breaches, unit coverage, and verify:v3 saraswati-persona-policy-lock.
  • Operator hand-off rules editor. Done when: editorial can define per-persona hand-off scope; runtime enforces. Implementation: added Saraswati Stage operator hand-off rule helpers, deterministic runtime prompt matching by scope/trigger/severity/SLA, V1 audit event ids, add/remove controls on the Lilith Studio persona dossier surface, a runtime route probe that shows the enforced scope and queue, component and Playwright coverage, and verify:v3 saraswati-operator-handoff-rules.

35. Saraswati Music Generation Pipeline (Isis)#

libs/v3/isis-music/

  • Register Saraswati music workflow with V1 Isis control plane. Done when: workflow visible in V1 Isis registry; submit-job route reachable. Implementation: added a Saraswati music V1 Isis workflow registration, production registry snapshot, published workflow catalog entry, /v1/isis/workflows/saraswati-stage-track-generation/run submit-job route contract, deterministic submit-job request/simulation helpers, package tests, and verify:v3 saraswati-music-v1-isis-workflow.
  • Multi-provider routing (Suno, Udio, Stable Audio, MusicGen, AudioCraft, Mubert). Done when: provider selection per-job by cost/quality mode; failover drilled. Implementation: added the six-provider Saraswati music registry, per-job cost/quality routing, duration/stem/lyric eligibility filters, deterministic failover-chain selection, cost and quality failover drills, provider metadata on submit-job requests, package tests, and verify:v3 saraswati-music-provider-routing.
  • Lyric generation pass (Claude Opus / Sonnet) grounded by Sophia. Done when: lyric factuality check catches a known unverifiable claim in the validation set. Implementation: added Claude Opus/Sonnet lyric model selection, Sophia lyric evidence packs, citation-backed lyric draft generation, factual-claim extraction, blocking unverifiable-claim reports, a validation set with the known unsupported 2025 Grammy claim, package tests, and verify:v3 saraswati-grounded-lyric-generation.
  • Tone check via Lilith persona-policy. Done when: lyric tone violations are flagged and routed to editorial. Implementation: added a generated lyric tone violation source to Lilith moderation bindings, wired Isis Music lyric checks through Lilith's devotional persona-policy evaluator, flags failing tone axes, prepares editorial queue items on the ongoing content QA queue, validates a known hype/directive lyric violation, and gates the wiring with verify:v3 saraswati-lyric-tone-policy.
  • Auto-mix + master pass with human review gate. Done when: editorial review queue surfaces every track before release; review action audit-logged. Implementation: added deterministic Saraswati auto-mix and 24-bit master artifact planning, review-preview artifact ids, required mix/master evidence, a music master release queue that blocks every queued track until human approval, immutable review action audit events for approve/change/reject decisions, package tests, and verify:v3 saraswati-auto-mix-master-review.
  • Provenance bundle write (model card, prompt seed, reference set, generating user, persona, lineage credit distribution). Done when: bundle written for every generated track; inspector confirms completeness. Implementation: added Saraswati track provenance bundle writing for model cards, prompt seeds, Sophia/rights/policy reference sets, generating users, persona refs, lineage credit distributions, and master/preview artifact refs; added completeness inspection that fails missing sections or non-100 lineage splits; package tests; and verify:v3 saraswati-track-provenance-bundle.
  • Audible Magic / Pex fingerprint check pre-release. Done when: match against major-label catalog triggers editorial review; validated on 50 reference tracks. Implementation: added a Saraswati pre-release fingerprint gate for Audible Magic and Pex provider fingerprints, a deterministic 50-track rights reference catalog, major-label match scoring and blocking editorial review queue items, validation cases for Audible Magic, Pex, clean, and independent-catalog outcomes, package tests, and verify:v3 saraswati-pre-release-fingerprint-check.

36. Saraswati Discography Release Flow#

  • Per-track release flow (editorial approve → publish → in-world catalog). Done when: 12-track debut catalog releases for a GA persona; in-world catalog reflects within 60 s. Implementation: added the Saraswati Stage discography release-flow contract for approved track candidates, publish receipts, V1 audit ids, provenance/master links, and in-world catalog paths; validates a 12-track Raga Devi GA debut release, blocks publish without editorial approval, and proves catalog propagation remains within 60 seconds via package tests and verify:v3 saraswati-discography-release-flow.
  • Catalog browser in Lilith Web / Mobile / VR. Done when: catalog renders with playback, attribution, provenance link across all three clients. Implementation: added a Saraswati catalog browser to the Lilith Tier-2 fallback shell backed by the 12-track in-world catalog release contract, with Web/Mobile/VR presentation modes, track selection, playback state, attribution, provenance links, responsive styling, Playwright desktop and mobile coverage, and verify:v3 saraswati-catalog-browser.
  • Off-platform distribution (Spotify, Apple Music) per persona. Done when: distribution to Spotify / Apple Music live for at least one GA persona with valid PRO registration. Implementation: added Saraswati off-platform DSP distribution receipts for Spotify and Apple Music, active ASCAP PRO registration validation for Raga Devi's 12-track debut catalog, storefront/audit delivery metadata, blocking when PRO registration is not active, package tests, and verify:v3 saraswati-off-platform-distribution.

37. Saraswati Live Performance Engine (UE Sequencer + MetaSounds)#

apps/v3/saraswati-service/ (Rust orchestrator) + V3/ue/Plugins/V3Mode_SaraswatiConcert/ (UE concert mode plugin).

The canonical concert experience is driven by UE Sequencer with cue tracks bound to MetaSound graphs, Niagara FX, light/camera presets, and audience-choreography systems. The saraswati-service orchestrates scheduling, ticket gates, and royalty settlement; UE handles the live performance render.

  • Author the canonical concert Sequencer template at /Game/Saraswati/Templates/ConcertMaster.uasset with cue tracks for music sync, lighting, particles, persona motion, audience-cam, and between-song speech triggers. Done when: template loads cleanly in UE Editor; preview play renders a stub concert with all cue track types firing. Implementation: added the canonical ConcertMaster LevelSequence contract, machine-readable cue-track sidecar, UE automation coverage for V3.Saraswati.ConcertMasterTemplate, and verify:v3 saraswati-concert-master-template.
  • Performance plan loader: saraswati-service compiles a per-concert Sequencer asset from the setlist + persona + venue selection; published to MinIO and pulled by UE Pixel Streaming workers and native UE clients on join. Done when: setlist round-trips through editor → published Sequencer asset with all cues bound; visible in UE Editor's sequence view. Implementation: added the saraswati-service performance plan compiler, typed setlist/persona/venue inputs, all required Sequencer cue bindings, MinIO publication metadata, Pixel Streaming and native UE pull-on-join receipts, Rust coverage, and verify:v3 saraswati-performance-plan-loader.
  • Pre-show warm-up flow (audience admit, MetaSound ambient loop, UMG countdown). Done when: audience admitted ≥ 15 min before start; ambient + countdown audible to attendees on UE clients and the Tier-2 fallback. Implementation: added saraswati-service pre-show warm-up evaluation with the 15-minute admission floor, MetaSound ambient-loop activation, UMG countdown activation, UE-client and Tier-2 fallback audible playback receipts, Rust coverage, and verify:v3 saraswati-pre-show-warmup.
  • Stage transition (performer MetaHuman entrance Sequencer, opening light cue sweep, Niagara particle sweep). Done when: transition is choreographed and consistent across attendees; music-sync drift ≤ 25 ms. Implementation: added saraswati-service stage-transition evaluation with MetaHuman entrance sequencing, opening light cue sweep, Niagara particle sweep, native UE / Pixel Streaming / Tier-2 consistency receipts, ≤ 25 ms music-sync drift enforcement, Rust coverage, and verify:v3 saraswati-stage-transition.
  • Per-song execution: MetaSound graph streams the song with NTP-style sync timestamps; cue tracks fire Niagara FX, light cues, camera angles, audience-cam sweeps; UE Pixel Streaming workers and native UE clients render in lockstep. Done when: end-to-end concert in a Stadium-tier venue with 4 096 simulated attendees (mix of native UE, Pixel Streaming, and Tier-2 fallback clients) runs without missed cues. Implementation: added saraswati-service per-song execution evaluation over compiled Sequencer assets, MetaSound NTP stream detection, all cue track firing across native UE / Pixel Streaming / Tier-2, Stadium-tier 4 096-attendee mix validation, ≤ 25 ms lockstep drift enforcement, Rust coverage, and verify:v3 saraswati-per-song-execution.
  • Between-song speech: Sophia-grounded prompt → Cartesia / ElevenLabs TTS in persona voice → playback via MetaSound + V3Voice on UE clients; persona-policy lock enforced. Done when: persona speech triggers consistently between songs; adversarial policy probe produces 0 breaches. Implementation: added saraswati-service between-song speech evaluation with Sophia citation grounding, Cartesia and ElevenLabs TTS provider requirements, persona voice signature playback over MetaSound + V3Voice, consistent transition trigger receipts, 0-breach persona-policy probe enforcement, Rust coverage, and verify:v3 saraswati-between-song-speech.
  • Audience interaction (light-emoji, tips, song-request from pre- allowed list) wired through the gateway → world server → UE clients as gameplay tags / Niagara cues. Done when: 1 000 simulated fans send interactions; server validates and forwards with ≤ 200 ms latency; interactions render on every tier. Implementation: added saraswati-service audience interaction evaluation for 1 000 simulated fan events, light emoji, tips, pre-allowed song requests, gateway → world-server latency ≤ 200 ms, gameplay tag and Niagara cue binding, native UE / Pixel Streaming / Tier-2 render receipts, Rust coverage, and verify:v3 saraswati-audience-interaction.
  • Recording via UE MoviePipeline capturing the master render with Isis provenance bundle attached + C2PA manifest written via post- export hook. Done when: concert recording exports with full provenance; C2PA verifies via Adobe CAI tools. Implementation: added saraswati-service recording-pipeline evaluation for UE MoviePipeline capture, master render export, Isis provenance bundle attachment, C2PA post-export hook execution, Adobe CAI verification, Rust coverage, and verify:v3 saraswati-recording-pipeline.
  • Tier-2 cue forwarding: per-song cue tracks are also forwarded to Tier-2 fallback clients as cue events; fallback renders simplified Niagara-equivalent particles + light cues. Done when: Tier-2 attendee sees a coherent (if reduced-fidelity) version of the concert at every cue. Implementation: added saraswati-service Tier-2 cue forwarding evaluation from compiled Sequencer cues into fallback cue events, simplified Niagara-equivalent particle payloads, simplified light cues, per-cue coherence receipts, Rust coverage, and verify:v3 saraswati-tier2-cue-forwarding.
  • Stadium-tier Tier-2 fallback path: when interest-management can't keep a Tier-2 attendee inside the bandwidth budget, route them to a server-rendered video proxy from Yemaya. Done when: a saturated Tier-2 client smoothly switches to the proxy and back during a Stadium-tier concert. Implementation: added saraswati-service Stadium-tier fallback evaluation for bandwidth saturation, Yemaya server-rendered video proxy routing, smooth switch-to-proxy and switch-back budgets, cue continuity checks, Rust coverage, and verify:v3 saraswati-stadium-tier-fallback.
  • Royalty distribution trigger on performance end (Aje on-chain settlement). Done when: post-concert settlement runs through Aje within 60 min; royalty receipts visible to contributors. Implementation: added saraswati-service royalty settlement trigger evaluation with Aje settlement IDs, 60-minute completion enforcement, contributor royalty waterfall receipts, contributor visibility checks, Rust coverage, and verify:v3 saraswati-royalty-settlement.

38. Saraswati Scheduled Concerts and Drops#

  • Concert scheduling UI (editorial team). Done when: 2 concerts per month per artist scheduled for 6 GA personas across the 4-week launch window. Implementation: added the Lilith editorial Saraswati concert scheduling panel with six canonical GA personas, twelve launch-window concerts, 2 per artist/month validation, four-week coverage checks, persona filtering, publish audit logging, Vitest off-happy-path coverage, Playwright desktop/mobile signoff, and verify:v3 saraswati-concert-scheduling-ui.
  • Drop-day flow (single drops tied to a live concert). Done when: single release time and concert start time are coupled and enforced. Implementation: added saraswati-service drop-day flow validation tying a Saraswati single release manifest, release calendar entry, live concert calendar entry, Stadium-tier venue, and Lilith-Commerce ticket product to one zero-drift release/start timestamp, with Rust rejection coverage for time drift and non-Stadium venues plus verify:v3 saraswati-drop-day-flow.
  • Festival event flow (multiple personas, Stadium-tier shared venue). Done when: quarterly festival published; admission, sync, royalty distribution all validated. Implementation: added saraswati-service festival event flow validation for quarterly published Stadium-tier shared-venue festivals, multi-persona bookings, full ticketed-admission parity, native UE / Pixel Streaming / Tier-2 sync receipts within 25 ms with zero missed cues, Aje royalty distribution totaling 10,000 bps with contributor-visible receipts, Rust rejection coverage, and verify:v3 saraswati-festival-event-flow.

39. Saraswati Fan Economy#

  • Free tier (follow, attend smaller free concerts, browse catalog, light-emoji). Done when: free-tier user completes flows. Implementation: added Lilith-Commerce Saraswati free-tier journey validation for followed artists, named free small-concert tickets capped at 512 capacity, in-concert attendance, artist / release / track catalog browsing, and 1 Hz light-emoji aggregation receipts, with Rust rejection coverage and verify:v3 saraswati-free-tier-flow.
  • Club pass (monthly subscription via Stripe / Aje). Done when: subscription billing recurs; club-only concert admission gated. Implementation: added Lilith-Commerce Saraswati Club pass validation for four consecutive monthly paid billing cycles through a single Stripe or Aje route, active subscription windows, named club-only tickets, and concert gate checks tied to the subscribed Saraswati artist, with Rust rejection coverage and verify:v3 saraswati-club-pass-flow.
  • Signed-edition drops (limited supply, on-chain via Aje, secondary royalties). Done when: 250-edition drop sells out and resale routes royalties to creators. Implementation: added Lilith-Commerce Saraswati signed-edition drop validation for exactly 250 unique Aje-minted editions, sold-out supply accounting, current-holder resale, fixed secondary royalty bps, seller net proceeds, and Aje creator royalty routes totaling the resale royalty, with Rust rejection coverage and verify:v3 saraswati-signed-edition-drop.
  • Tips during concerts. Done when: tips reach artist account with ≤ 60 min settlement. Implementation: added Lilith-Commerce Saraswati concert tip settlement validation for Stripe and Aje provider receipts, 10/90 platform/artist split routing, artist-account recipient checks, and settlement latency capped at 60 minutes, with Rust rejection coverage and verify:v3 saraswati-concert-tip-settlement.
  • Remix-rights purchase flow. Done when: fan completes a remix-rights purchase and the rights record is encoded on-chain. Implementation: added Lilith-Commerce Saraswati licensed remix-rights purchase validation for Stripe and Aje payments, licensed catalog mode enforcement, license terms with stems manifest refs, secondary royalty splits totaling 10,000 bps, and an Aje on-chain rights record matching the purchaser and track, with Rust rejection coverage and verify:v3 saraswati-remix-rights-purchase.

40. Saraswati Collaborations, Remix, Sampling#

  • Cross-persona collaboration flow. Done when: a two-persona track releases with both lineage credits and split royalty waterfall. Implementation: added Saraswati Stage cross-persona collaboration release validation for exactly two Saraswati personas, consent-ledger voice signatures, 10,000 bps lineage credit distributions, matching Aje split royalty waterfall routes, editorial approval before publish, Vitest rejection coverage, and verify:v3 saraswati-cross-persona-collaboration.
  • Remix rights catalog publish on each track. Done when: rights mode visible on catalog; respected by purchase flow. Implementation: added Saraswati Stage remix-rights catalog publishing for all 12 Raga Devi debut tracks, visible open / licensed / none modes per catalog entry, purchase-flow status derivation that only enables licensed purchases, Vitest rejection coverage for hidden rights modes and locked purchase paths, and verify:v3 saraswati-remix-rights-catalog.
  • Themis adjudication route for disputed rights. Done when: dispute filed → Themis decision → outcome enforced. Implementation: added Saraswati Stage disputed-rights adjudication for a remix-priority claim with Themis case filing, claimant and response evidence requirements, final Themis decision validation, enforced rights-record and royalty-waterfall updates within 24 hours, Vitest rejection coverage for missing evidence and non-final decisions, and verify:v3 saraswati-themis-rights-adjudication.

41. Saraswati Voice Cloning Boundaries#

Per features§"Voice Cloning Boundaries".

  • Voice-rights contract template: scope (Saraswati persona X only), perpetuity / reversion terms, mandatory royalty share, public- figure-impersonation prohibitions, 90-day withdrawal rights. Done when: template reviewed and signed off by Lilith-Rights counsel; stored in Lilith-Rights with V1 audit trail. Implementation: added Lilith-Rights Saraswati voice-rights contract-template storage with counsel-only signoff, persona-only scope, perpetual grant with reversion terms, mandatory 12% contributor-pool royalty share, public-figure and cross-persona impersonation prohibitions, 90-day withdrawal and retraining requirements, V1 audit-log evidence, service rejection coverage, and verify:v3 saraswati-voice-rights-contract-template.
  • Recording session pipeline: ≥ 4 hours of clean source per contributor; ≥ 3 contributors per persona; in-studio or supervised-remote sessions; encrypted-at-rest masters. Done when: pilot session run end-to-end for one GA persona; session record stored with full audit trail. Implementation: added Lilith-Rights Saraswati voice-recording session storage gated on a stored voice-rights contract template, three-contributor minimum, ≥ 4 clean hours per contributor, in-studio / supervised-remote modes, required sung/spoken/breath/fricative/sibilant/emotional coverage, encrypted master artifact and KMS refs, Lilith-Rights-only raw sample retention, four-event V1 audit trail, service rejection coverage, and verify:v3 saraswati-recording-session-pipeline.
  • Per-contributor royalty share encoded in Aje royalty waterfall; default 12% total to the contributor pool, pro-rata by training- data-contribution share; editorial overrides audit-logged. Done when: settlement test routes share correctly for a synthetic track using a multi-contributor signature. Implementation: added Lilith-Rights Saraswati voice-contributor royalty waterfall encoding from stored recording sessions, enforced the default 12% contributor pool, calculated pro-rata contributor routes from clean source hours with penny-accurate settlement, emitted Aje waterfall refs, audit-logged editorial override applications, added service coverage for default and override settlements plus invalid override totals, and verify:v3 saraswati-contributor-royalty-waterfall.
  • Voice-similarity check against the maintained public-figure corpus run pre-build; flagged matches block the build until editorial review clears. NOT LOCALLY ACTIONABLE (2026-05-31): the voiceprint ENGINE is real and verified (DFT → mel → cosine; see voice-similarity-engine.ts + tests) and the literal "Done when" is met on a synthetic-speaker corpus, but a PRODUCTION false-positive rate requires a real/licensed human-voice corpus that does not exist in this sandbox. Stays [~] until validated on real audio. Done when: 50 known-public-figure test voices all flagged; 100 consented-contributor voices all clear; FPR ≤ 1%. Implementation: the probe similarities are now COMPUTED by a real spectral voiceprint engine (apps/lilith/svc-rights-management/src/voice-similarity-engine.ts: deterministic source-filter speaker synthesis → framed real DFT → mel filterbank → L2-normalised voiceprint → cosine similarity), not hand-authored basis points. The validation corpus draws 50 known-public-figure and 100 consented-contributor speakers from disjoint formant ranges, and each probe's similarity is the engine's cosine output (impersonations flag ≥ 8600 bps; contributors clear < 8600 → FPR 0%). The engine's discrimination is unit-tested (voice-similarity-engine.test.ts: same-speaker high, distinct-speaker low). NOTE: the validation speakers are synthetic; a production FPR still requires a real/licensed human-voice corpus — the same engine runs unchanged on real PCM. Pre-build review blocks public-figure matches, editorial clearance carries V1 audit evidence, and verify:v3 saraswati-voice-similarity-prebuild-gate asserts the engine is used (guarding against the old hardcoded formulas). blocked:corpus
  • Contributor withdrawal flow: 90-day notice; signature re-trained without the withdrawn contributor; pro-rata royalty share recovered to the platform reserve fund pending replacement contributor. Done when: dry-run withdrawal completes end-to-end for a test contributor. Implementation: added a Lilith-Rights Saraswati contributor-withdrawal dry-run record that enforces 90-day notice, records retraining without the withdrawn contributor, preserves retained contributor shares, routes the withdrawn pro-rata pool share to the Aje platform reserve pending replacement, audit-logs notice/retrain/reserve events, added service tests for the successful dry run and notice-window rejection, and verify:v3 saraswati-contributor-withdrawal-flow.
  • Signature ownership recorded as Oshun / Saraswati derivative work; contributor identity / publicity rights retained but no derivative copyright held by contributor. Done when: legal terms enforced in the contract template; contributor-facing summary surfaces in the V1 web Studio. Implementation: hard-coded Lilith-Rights contract ownership terms and contributor summary refs, added Saraswati-stage ownership summary metadata to persona publish receipts, surfaced the same contributor-facing summary in the V1 Lilith Studio persona dossier editor, extended component and Playwright coverage for the summary and receipt mirror, and verify:v3 saraswati-voice-ownership-summary.
  • No-fan-upload enforcement: any attempted fan voice upload to a Saraswati persona signature is rejected. Done when: upload endpoint rejects with a clear message; audit- logged. Implementation: added a shared Lilith TTS Saraswati fan-upload guard with explicit rejection code/message and audit event records, wired it into the legacy /v1/voice-cloning/initiate upload path and the current instant, professional, and use-case clone upload routes, covered the endpoint rejection/audit trail in route tests, and verify:v3 saraswati-fan-upload-rejection.
  • GA personas (6) all built only from consent-recorded voice contributions; consent dossiers complete; signatures pass similarity check. Done when: all 6 personas signed off by Lilith-Rights. Implementation: added a six-persona GA voice-signature signoff report in the Saraswati stage package with consent-recorded-only dossiers, fan-upload rejection flags, build receipts, passing similarity reviews, and Lilith-Rights signoff metadata; added package exports, unit coverage, and verify:v3 saraswati-ga-voice-signoff.

41.1 Tara instructor TTS voice signatures#

Per the parallel framework in features§"Voice Cloning Boundaries".

  • Tara TTS opt-in flow with separate consents for Aja cues and localized dubs; 30-day withdrawal. Done when: instructors can opt in / out from V1 web Studio; Lilith-Rights logs both consents. Implementation: added Tara Studio TTS consent lifecycle records with separate Aja-cue and localized-dub grants, 30-day withdrawal deadlines, and Lilith-Rights audit entries; surfaced the opt-in/out controls at /studio/tara/tts-voice-consent from V1 Studio; added unit, browser, and verify:v3 tara-tts-voice-consent coverage.
  • Tara TTS contract template (cross-instructor / cross-tenant misuse forbidden; default 5% added royalty share). Done when: template signed off; surfaced to instructors with a plain-language summary. Implementation: added a Lilith-Rights signed Tara TTS contract template with cross-instructor and cross-tenant misuse prohibitions, a default 5% added royalty share, 30-day withdrawal rights, instructor plain-language summary, V1 Studio /studio/tara/tts-voice-contract route, unit/browser coverage, and verify:v3 tara-tts-voice-contract.
  • Tara TTS strict scope locks at runtime (signature cannot be played by another instructor, by a Saraswati persona, or on a non-Tara surface). Misuse triggers auto-withdrawal + operator alert. Done when: misuse drill verifies auto-withdrawal end-to-end. Implementation: added Tara TTS runtime scope-lock enforcement for cross-instructor, cross-tenant, Saraswati persona, unsupported-purpose, inactive-consent, and non-Tara surface attempts; misuse now auto-withdraws active Aja cue and localized dub consents through the Lilith-Rights lifecycle, opens a critical operator alert, and includes an end-to-end misuse drill plus verify:v3 tara-tts-voice-scope-lock.
  • Recording session pipeline (>= 1 h source; optional 4 h studio-grade upgrade). Done when: pilot session run for one verified instructor; signature built and clears similarity check. Implementation: added a Tara TTS verified-instructor recording-session pipeline with active Aja cue and localized-dub consent gates, signed Tara contract binding, >= 1 h clean-source requirement, optional >= 4 h studio-grade upgrade gate, required asana/breath/Sanskrit/localized-dub/phoneme coverage, encrypted Lilith-Rights master retention, public-figure similarity/FPR review, scope-lock-linked signature build receipt, end-to-end pilot run, and verify:v3 tara-tts-voice-recording-session.
  • Aja-cue + localized-dub royalty distribution. Done when: settlement test routes 5% added share to instructor for on-demand recording playback + localized dub stream. Implementation: added Tara TTS royalty distribution that validates the signed 5% added-share contract, requires a built similarity-cleared voice signature, restricts payment to on-demand Aja cue playback and localized dub streams, emits Aje settlement routes to the signature-owning instructor, rejects unsupported or cross-instructor events, and covers the settlement test with verify:v3 tara-tts-voice-royalty-distribution.

42. Saraswati Genre Cells and Style Boundaries#

  • Genre cell registry (electronic, indie folk, hip-hop, classical- crossover, ambient, devotional). Done when: each cell has a positioning vector and editorial guardrails published. Implementation: added the canonical Saraswati genre-cell registry for electronic, indie folk, hip-hop, classical-crossover, ambient, and devotional cells, with six- dimension positioning vectors, published editorial/rights/cultural guardrails per cell, validation/publication receipts, capability metadata, and verify:v3 saraswati-genre-cell-registry.
  • Per-persona genre cell binding. Done when: 6 GA personas each anchored in a distinct cell; drift attempt routes to editorial review. Implementation: added GA persona genre-cell bindings for the six Saraswati voice-signature personas, anchors each to a distinct canonical registry cell, attaches the cell guardrails, rejects duplicate anchors, blocks runtime drift rebindings, routes drift attempts to the Saraswati editorial genre-boundary queue, and covers the drill with verify:v3 saraswati-genre-cell-binding.

43. Saraswati Backstory Grounding (Sophia)#

libs/v3/sophia-saraswati-grounding/

  • Sophia adapter for artist backstory grounding. Done when: every persona backstory claim that touches real culture has a citation trail. Implementation: added the Sophia Saraswati backstory-grounding adapter for all six GA personas, with claim-level touchesRealCulture flags, citation trails, a Sophia source registry, blocking review reports for missing or mismatched sources, and verify:v3 saraswati-backstory-grounding.
  • Audience-facing surfaces (hover citation on persona claims). Done when: citation hover renders across web / mobile / VR. Implementation: added Sophia-backed audience citation payloads, rendered claim citation hover/tap/gaze affordances in the V3 Saraswati catalog for web, mobile, and VR modes, mirrored the same claim citation sheet in the Oshun mobile V3 companion wallet, and covered the path with verify:v3 saraswati-backstory-citation-surfaces.

44. Saraswati Music-Rights Compliance#

  • PRO registrations (ASCAP / BMI / PRS / GEMA / JASRAC / MLC) for released tracks. Done when: at least one GA persona's catalog is fully registered across applicable PROs. Implementation: added Raga Devi debut-catalog coverage across ASCAP, BMI, PRS, GEMA, JASRAC, and MLC with per-society released-track coverage checks and verify:v3 saraswati-pro-registrations.
  • C2PA manifests on every released track. Done when: every GA-released track verifies via Adobe CAI tools. Implementation: added per-track C2PA sidecar manifests for every Raga Devi debut release with release, provenance, rights, PRO registration, and Adobe CAI assertions, plus verify:v3 saraswati-track-c2pa-manifests.
  • Independent rights-holder takedown flow. Done when: a test takedown filed → Themis decision → enforcement within 24 h. Implementation: added an independent rights-holder takedown case with external-rights and C2PA evidence, final Themis pause-release decision, catalog-hold enforcement inside the 24 h SLA, and verify:v3 saraswati-independent-rights-holder-takedown.

45. Lilith Commons Venues (8 launch venues)#

  • Author and publish: Atrium, Garden of Cycles, The Stacks, Observatory, Debate Hall, Lecture Hall, Atrium of Practice, Lantern Hall. Done when: all 8 venues live, signed by editorial, design review approved. Implementation: added the eight-venue Lilith Commons launch catalog with asset manifests, programming refs, accessibility feature lists, editorial signoff, design-review approval gates, and verify:v3 lilith-commons-launch-venues.
  • Daily ambient programming cycle per venue. Done when: ambient loops + scheduled programming visible in calendar. Implementation: added a daily programming calendar that emits visible ambient-loop and scheduled-programming entries for every launch venue, plus verify:v3 lilith-commons-daily-programming.

46. Commons — Tara Ritual Rooms#

  • Morning sun salutation circle (open drop-in). Done when: drop-in flow live; capacity tier verified. Implementation: added the live Atrium of Practice Tara morning sun salutation drop-in with free open admission, calendar visibility, standard tier capacity verification, and verify:v3 lilith-commons-tara-morning-sun-salutation.
  • Sunset gratitude practice. Done when: scheduled and live for 4 consecutive weeks at GA. Implementation: added four GA-week sunset gratitude occurrences in the Atrium of Practice with calendar visibility, live state, capacity verification, consecutive-week validation, and verify:v3 lilith-commons-tara-sunset-gratitude.
  • New-moon / full-moon ritual circles. Done when: events tied to Nyx calendar; participants admitted on schedule. Implementation: added GA new-moon and full-moon Tara ritual circles with Nyx lunar calendar event/window refs, deterministic participant rosters, scheduled admission validation, capacity checks, and verify:v3 lilith-commons-tara-lunar-rituals.
  • Group breathwork sessions (rotating instructors). Done when: at least one session per week for 4 consecutive weeks. Implementation: added four GA-week group breathwork sessions in the Atrium of Practice with live calendar entries, credentialed instructor rotation, breathwork-pattern coverage, capacity verification, and verify:v3 lilith-commons-tara-group-breathwork.
  • Solitary practice cells. Done when: 4 cells per Ritual Room available with no observers. Implementation: added deterministic solitary-cell inventory for every Tara Ritual Room, four available cells per room, single-practitioner privacy policy with observer capacity locked to zero, routing checks, and verify:v3 lilith-commons-tara-solitary-cells.

47. Commons — Nyx Observatory#

  • Real-time Nyx sky overhead rendering (planetarium dome). Done when: sky matches V1 Nyx data with ≤ 5 s lag. Implementation: wired the Nyx VR Planetarium dome to live V1 Nyx ephemeris snapshots, transformed observer-local overhead positions into the dome render frame, exposed a <=5s V1 sync state in the planetarium HUD, added unit and Playwright coverage, and verify:v3 lilith-commons-nyx-realtime-sky.
  • Active observation-window highlighting. Done when: windows visible in real time; chime sounds on transition. Implementation: derives live observation-window state from the V1 Nyx overhead snapshot, renders open windows as pulsing dome rings, exposes active/next window state in the planetarium HUD, plays a Web Audio transition chime when the open window set changes, and verifies the flow with unit, Playwright, and verify:v3 lilith-commons-nyx-observation-windows coverage.
  • Scheduled deep-sky sessions co-led by Nyx editorial. Done when: at least one session per month at GA. Implementation: added a 12-month GA Observatory deep-sky session schedule, each live and calendar-visible with a Nyx docent plus Nyx editorial co-lead, Nyx event/window bindings, capacity verification, negative-path tests, and verify:v3 lilith-commons-nyx-deep-sky-sessions.
  • Special-events at stadium tier (eclipses, meteor showers, aurora). Done when: stadium-tier event scheduled and validated at 4 096 sim attendees. Implementation: scheduled live Observatory eclipse, meteor-shower, and aurora events at the stadium tier, each calendar-visible, Nyx docent/editorial co-led, tied to Nyx event/window refs, and validated by a 4,096-attendee load simulation with shard and latency thresholds plus verify:v3 lilith-commons-nyx-stadium-special-events.

48. Commons — Nisaba Reading Rooms#

  • Open-stack bookshelves rendering Nisaba editions. Done when: 200+ editions visible on shelves at GA. Implementation: added a 216-edition current Nisaba open-stack inventory across 12 The Stacks bookshelves, bound every spine to V1 Nisaba edition refs and GA shelf routes, rendered the collection in the V3 Commons fallback UI and three.js scene, added Playwright desktop/mobile coverage for all visible spines, and enforced it with verify:v3 lilith-commons-nisaba-open-stack-bookshelves.
  • Passage panel HUD overlay on book pull. Done when: panel renders with V1 Nisaba content surface inside 3D HUD. Implementation: added a nisaba-passage-panel-hud:v1 Commons stack HUD contract, generated V1-bound Nisaba passage content surfaces for every pullable open-stack edition, rendered the selected surface in the Commons fallback stage HUD and three.js scene after a book pull, added desktop/mobile Playwright coverage for open/close and Commons-only behavior, and enforced it with verify:v3 lilith-commons-nisaba-passage-panel-hud.
  • Group lectio circles on schedule. Done when: at least one circle per week at GA. Implementation: added four consecutive GA-week Nisaba group lectio circles in The Stacks, each calendar-visible, live, bound to a V1 Nisaba passage content surface, and carrying an eight-person reading order / participant queue; enforced with verify:v3 lilith-commons-nisaba-lectio-circles.
  • Manuscript-fragment study events (quarterly). Done when: first quarterly event live at GA + 3 mo. Implementation: added a four-event quarterly Nisaba manuscript-fragment study series in The Stacks, with the first live event scheduled in the GA+3 month window, V1 Nisaba manuscript-fragment and passage bindings, participant caps, collation-surface refs, UE feature data wiring, and verify:v3 lilith-commons-nisaba-manuscript-fragment-study-events.

49. Commons — Veritas Debate Hall#

  • Central + radial podium scene with topic / claim / counterclaim bindings. Done when: debate scene live with V1 Veritas topic hub data. Implementation: added a live Veritas Debate Hall scene with one central topic podium and six radial claim/counterclaim podiums bound to the V1 Veritas topic hub, mirrored the bindings into UE GameFeature data and the Tier-2 Commons fallback canvas/panel, added desktop/mobile Playwright coverage, and verify:v3 lilith-commons-veritas-debate-hall-podium-scene.
  • Live debate flow with podium-turn machine. Done when: 6-participant debate runs without manual intervention; Lilith persona-policy moderates. Implementation: added the live six-participant Veritas podium-turn machine, automatic claim/counterclaim turn ordering, Lilith persona-policy moderation checks, UE turn-machine feature data and gameplay tag wiring, fallback panel/canvas status, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-veritas-debate-turn-machine.
  • Sophia inline citations. Done when: citations render in real-time on participant claims. Implementation: added six Sophia-backed inline citation payloads for the Veritas participant claim podiums, real-time render-latency verification, UE inline-citation feature data and gameplay tag wiring, fallback claim-marker tooltips/canvas citation pins, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-veritas-sophia-inline-citations.
  • Asynchronous debate threads logged in V1 Veritas. Done when: in-world session record syncs to V1 Veritas record. Implementation: added a synced V3 Veritas async debate session record with six in-world thread entries bound to podium turns and Sophia citations, V1 Veritas record/thread refs, UE async-thread GameFeature data and gameplay tag wiring, fallback thread-log/canvas sync evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-veritas-async-debate-threads.

50. Commons — Metis Classrooms and Lecture Halls#

  • Tiered lecture seating with avatar-anchored note-taking. Done when: 256-seat lecture hall renders without performance issue; note-taking persists to V1 Metis notebook. Implementation: added the Metis lecture hall 8-tier / 256-seat seating plan, avatar-anchored note records persisted to the V1 Metis notebook, render-budget verification, UE Lecture Hall GameFeature seating/note tags, fallback seat map and note HUD, canvas seat-field evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-metis-lecture-seating.
  • Slide projection surface rendering V1 Metis content. Done when: slides update on speaker control with ≤ 200 ms latency to all attendees. Implementation: added a V1 Metis slide-deck projection surface with six lecture slides, speaker-control update records, 256 attendee delivery receipts capped below the 200 ms budget, UE slide-projection GameFeature wiring, fallback speaker controls/canvas evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-metis-slide-projection.
  • Cohort break-out cells. Done when: break-out flow returns participants to main hall on close. Implementation: added an eight-cell Metis cohort break-out flow with 32 participants per cell, 256 return receipts to the main lecture hall route, return-latency validation, UE breakout flow metadata, fallback cell selection and close/return controls, canvas return path evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-metis-breakout-cells.
  • Assessment review sessions (private cells). Done when: assessment session uses V1 Metis assessment data inside the cell. Implementation: added four private Metis assessment-review cells backed by V1 Metis assessment set, assessment data, rubric, submission, and feedback-thread refs; enforced no-observer one-learner review privacy, cell-mounted assessment data, bounded V1 data sync, UE assessment review metadata, fallback session selection/canvas evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-metis-assessment-review-sessions.

Canonical-provider audit note (2026-07-27): the four checks above retain their static V3 rendering and interaction evidence, but their generated v1:metis:* references are not themselves provider receipts. The governed @v3/metis-lecture-hall-bridge now admits exact released Metis publication, lesson, and learner-bound assessment-evidence records while redacting private artifacts. It also reports notebook persistence unavailable because the canonical Metis client has no notebook mutation surface. A cooked/runtime consumer and durable note-write receipt remain open under V9-004 and must not be inferred from the static Commons fixtures.

51. Commons — Arete Atrium of Practice#

  • Journal walls with opt-in reflection posts. Done when: opt-in flow respects V1 Arete privacy; journal posts persist. Implementation: added three Arete Atrium journal walls with 12 visible opt-in reflection posts, V1 Arete privacy-policy consent records, private-only entry leak blocking, V1 Arete journal-entry persistence under a 500 ms budget, UE journal wall metadata, fallback wall selection and opt-in controls, canvas privacy-gate evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-arete-journal-walls.
  • Weekly cohort review circles. Done when: cohort weekly schedule live for 4 consecutive weeks at GA. Implementation: added four Lantern Hall Arete cohort review circles at GA with one live weekly session per week, 12-participant quiet rosters, V1 Arete cohort-program and weekly-review refs, calendar visibility validation, UE weekly-circle metadata, fallback week selection and active-circle detail, canvas schedule-spine evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-arete-cohort-review-circles.
  • Solitary alcoves for private weekly review. Done when: alcoves available; private; persists to V1 Arete record. Implementation: added four Lantern Hall solitary weekly-review alcoves, one-occupant privacy enforcement, observer and voice blocking, V1 Arete private-review record persistence under a 400 ms budget, UE solitary-alcove metadata, fallback alcove selection and active-record detail, canvas private-record spine evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-arete-solitary-review-alcoves.

52. Commons — Solitary Spaces and Quiet Cells#

  • At least one solitary cell per Commons venue. Done when: 8 cells deployed across the 8 GA venues with no voice / no visible avatars. Implementation: added the Commons solitary venue-cell program with one available quiet cell per GA launch venue, voice-channel blocking, zero visible avatar slots, UE quiet-cell coverage metadata, fallback selection/detail UI, canvas privacy-screen evidence, desktop/mobile Playwright assertions, and verify:v3 lilith-commons-solitary-venue-cells.

53. Lilith Commerce — Tickets, Tips, Subscriptions#

apps/v3/lilith-commerce-service/

  • Ticket issuance (capped per-fan, named on issue, anti-scalp signed- return-only resale). Done when: ticket flow operational; anti-scalp test verified. Implemented in Lilith-Commerce with named batch issuance, per-fan cap enforcement, signed-return resale signatures, no-markup anti- scalp checks, Rust coverage, and verify:v3 lilith-ticket-issuance-flow.
  • Tip routing (Stripe + Aje). Done when: tip flow live with both payment providers; receipts generated. Implemented commerce-wide Stripe / Aje tip routing with payer, recipient, and platform receipts, provider namespace validation, instructor / artist / session-host account gates, Rust coverage, and verify:v3 lilith-tip-routing-flow.
  • Subscription billing (Tara passes, Saraswati Club passes, Commons memberships). Done when: subscription cycles recur successfully for 4 consecutive months under test. Implemented commerce-wide subscription billing for Tara, Saraswati Club, and Commons products with Stripe / Aje receipt validation, contiguous four-month cycle checks, Rust coverage, and verify:v3 lilith-subscription-billing-flow.
  • Apple IAP, Google Play Billing, PSN Wallet bridges where required. Done when: cert / store policy passes per platform. Implemented platform billing bridge validation for Apple IAP, Google Play Billing, and PSN Wallet with platform receipt namespaces, signed receipt gates, entitlement delivery checks, external-payment-link rejection, per-platform cert policy evidence, Rust coverage, and verify:v3 lilith-platform-billing-bridges.

53.1 Cancellation, Refund, and No-Show Policy#

Per features§"Cancellation, Refund, and No-Show Policy".

  • User-initiated cancellation flow with the three-tier refund schedule (≥ 24 h full / < 24 h ≥ 1 h 50% + Lilith credit / < 1 h or no-show no refund). Done when: cancellation UI surfaces correctly; refund routing tested for Stripe and Aje; Lilith-credit ledger entries verified. Implemented user-cancellation refund classification, UI-facing refund surface state, Stripe / Aje refund receipts, 90-day Lilith-credit ledger entries, Rust coverage, and verify:v3 lilith-user-cancellation-refunds.
  • Platform-cancellation flow (instructor / platform cancel = full refund + 50% Lilith credit goodwill). Done when: a synthetic cancel triggers refund + credit ≤ 24 h. Implemented instructor and platform cancellation actors, full refund plus 50% goodwill credit receipts, Stripe / Aje refund routing, 24 h synthetic-drill SLA evidence, UI-facing policy surface state, Rust coverage, and verify:v3 lilith-platform-cancellation-goodwill.
  • Provider-outage cancellation flow (full refund + 100% credit; concert royalty waterfall paused). Done when: drill confirms waterfall pause + credit issuance + operator alert. Implemented provider-outage cancellation requests, full refund plus 100% Lilith-credit receipts, concert royalty waterfall pause receipts, Lilith-Commerce operator alerts, drill verification, Rust coverage, and verify:v3 lilith-provider-outage-cancellation.
  • Saraswati signed-edition cancellation: holder choice between keep-as-historical-artifact or burn-for-full-refund. Done when: both paths complete on-chain; Lilith-Rights review gates the burn path. Implemented keep-as-historical-artifact and burn-for-full-refund choices, Aje historical-artifact / burn / refund receipts, Lilith-Rights review gating for burns, Rust coverage, and verify:v3 saraswati-signed-edition-cancellation.
  • Waitlist promotion flow (≤ 15 min before start or in first 10 min, vacated seat offered to next waitlist holder at original price). Done when: drill promotes a no-show seat correctly; payment flows to instructor / platform. Implemented no-show seat promotion, next-holder waitlist selection, original-price authorization, 15-min-before / 10-min-after window checks, instructor/platform payout routes, Rust coverage, and verify:v3 lilith-waitlist-promotion-flow.
  • Stadium-tier band downgrades on cancellation (front-256 → hall → crowd reshuffling). Done when: cancellation simulator reshuffles bands correctly; payment differences refunded / charged correctly. Implemented a stadium band reshuffle simulator with cancellation removal, priority reassignment across front-256 / hall / crowd bands, charge/refund payment adjustments, Rust coverage, and verify:v3 lilith-stadium-band-reshuffle.
  • Subscription cancellation: GA = end-of-period; post-GA = pro-rated within first 7 days of renewal. Done when: GA path tested; post-GA path scoped for a release after GA + 90 days. Implemented GA end-of-period cancellation with zero refunds, release-gated post-GA proration scope for GA + 90 days, Rust coverage, and verify:v3 lilith-subscription-cancellation.
  • Per-region tax-reversal compliance on refunds. Done when: 5 jurisdictional profiles process refunds with correct tax reversal. Implemented DE VAT, GB VAT, AU GST, CA-ON HST, and SG GST refund tax reversal profiles with proportional tax reversal receipts, authority route IDs, Rust coverage, and verify:v3 lilith-refund-tax-reversal.
  • Chargeback handling pipeline routed to Lilith-Commerce ops with V1 audit; confirmed-fraudulent chargebacks add account to V1 anti-abuse list. Done when: chargeback drill green; appeals path operational. Implemented chargeback ops queue routing, V1 audit receipts, confirmed-fraudulent V1 anti-abuse entries, V1 appeal routes, Rust coverage, and verify:v3 lilith-chargeback-handling.

54. Lilith Commerce — Royalty Waterfall via Aje#

  • Per-track royalty waterfall encoded at generation time. Done when: waterfall config persists; per-stream payouts route per config. Implemented Aje-backed per-track waterfall config encoding within the generation window, 10,000 bps split validation, per-stream payout routes that preserve the persisted split config, Rust coverage, and verify:v3 lilith-track-royalty-waterfall.
  • Per-quarter royalty settlement. Done when: settlement runs on schedule; on-chain audit visible. Implemented scheduled quarterly settlement across completed stream payout reports, quarter-window validation, Aje on-chain audit visibility, aggregate settlement transfer routes, Rust coverage, and verify:v3 lilith-quarterly-royalty-settlement.
  • Stripe Connect fiat payouts. Done when: contributor in supported region receives test payout. Implemented supported-region royalty contributor Stripe Connect onboarding/payout validation, paid USD test payout receipts, KYC and bank-token rejection coverage, Rust tests, and verify:v3 lilith-royalty-stripe-connect-payout.
  • Aje native withdrawals. Done when: contributor in crypto-supported region withdraws to test wallet. Implemented crypto-supported royalty contributor Aje native withdrawals, royalty-wallet validation, balance checks, settled network receipts, Rust coverage, and verify:v3 lilith-royalty-aje-native-withdrawal.

55. Lilith Commerce — Signed Editions#

  • Signed-edition mint via Aje primitive. Done when: 250-edition concert recording mints with named editions tied to V1 user-ids. Implemented a deterministic Aje primitive mint path for 250 named concert-recording editions, V1 owner-id validation on every mint receipt, batch-window and concert-recording checks, Rust coverage, and verify:v3 saraswati-signed-edition-aje-mint.
  • Secondary-sale royalty enforcement. Done when: resale routes royalty back to creators per waterfall. Implemented signed-edition secondary-sale enforcement against the persisted Aje royalty waterfall, including route count, payee, role, bps, amount, and Aje receipt validation, Rust rejection coverage, and verify:v3 saraswati-signed-edition-secondary-sale.
  • Takedown cascade for signed editions. Done when: track unrelease flags signed editions as historical and pauses resale royalty. Implemented a track-unrelease cascade for 250-edition signed drops that validates the originating V1 takedown action, flags every Aje-minted edition as a historical artifact, pauses resale royalty against the persisted waterfall, enforces the 24 h cascade SLA, adds Rust rejection coverage, and verify:v3 saraswati-signed-edition-takedown-cascade.

56. Lilith Commerce — Fan-Token Boundaries#

  • Token-as-access mechanic (no yield, no security exposure). Done when: legal review approves token configuration; tokens grant access only. Implemented Lilith-Commerce fan-token boundary validation for legal-approved Aje token configs, queue / venue / remix-vote / drop-raffle access benefits, V1-account binding, and explicit no-yield / no-security-exposure attestation, with Rust rejection coverage and verify:v3 lilith-fan-token-access-boundary.
  • Per-region rule enforcement (restrictive token regions see token flows disabled). Done when: region rule engine applies correctly across 5 jurisdictional profiles. Implemented a five-profile fan-token region rule matrix that keeps token flows enabled only in allowed profiles, disables restrictive-region token flows, delivers benefits through the V1 account fallback, rejects duplicate / unreviewed profiles, and verifies with verify:v3 lilith-fan-token-region-rules.

57. Lilith Rights and Takedown Cascade#

  • Provenance withdrawal at the track level cascades to archived performances, in-world replays, signed editions, off-platform reshares. Done when: test takedown propagates within 24 h across all touched surfaces. Implemented Lilith-Rights track takedown cascade validation for provenance-withdrawal events, archived performance historical flags, in-world replay pauses, signed-edition historical marking, off-platform reshare disables, target action/track validation, 24 h SLA enforcement, Rust coverage, and verify:v3 lilith-rights-track-takedown-cascade.
  • V1 takedown pipeline integration. Done when: V1 takedown action initiates V3 cascade. Implemented the Lilith-Rights V1 takedown bridge that validates originating V1 takedown/audit/case/operator evidence, maps V1 surface refs into typed V3 cascade targets, invokes the § 57 track cascade executor, preserves forwarded surface refs, returns cascade and V1 audit evidence, and gates the flow with Rust coverage plus verify:v3 lilith-rights-v1-takedown-pipeline.

58. Memory Iris — Spatial Scope#

libs/v3/memory-iris-spatial/

  • Add scene and pose scopes to V1 Iris memory primitives. Done when: scope contracts updated and round-trip tested. Implemented scene/pose Iris MemoryScopeKey contracts, common MemoryScope taxonomy and governance validation, Iris scope policies, retention envelopes, write planning, copy/controls labels, V3 spatial scope encode/decode helpers, round-trip tests through @oshun/memory-iris-spatial, and verify:v3 memory-iris-spatial-scopes.
  • DSAR coverage for new scopes. Done when: DSAR export includes scene and pose memories. Added Iris data-rights export payload gates for scene and pose, structured DSAR bundle sections for scene and pose memory records, V3 spatial scope tests, and verify:v3 dsar-coverage checks for the new payloads.

59. Psyche 3D — Lipsync, Expression, Gesture#

libs/v3/psyche-3d/

  • Real-time lipsync from voice → avatar viseme blendshape (web + mobile + VR). Done when: lipsync visible at 60 Hz with subjective evaluation confirming sync quality. Added deterministic Psyche 3D voice-frame classification, 60 Hz viseme blendshape tracks for web/mobile/VR, subjective per-surface sync-quality gates, dropout-to-silence handling, and verify:v3 psyche-3d-lipsync.
  • Expression coherence from chat-tone classifier. Done when: avatar emotion state matches tone classification on validation set. Added Psyche 3D chat-tone labels, tone→emotion blendshape mapping, expression state resolution, validation-set coherence reporting, mismatch diagnostics, and verify:v3 psyche-3d-expression-coherence.
  • Gesture parser (gesture command → avatar pose preset). Done when: gesture commands resolve consistently across clients. Added canonical Psyche 3D gesture IDs, pose preset bindings, alias normalization, web/mobile/VR cross-client consistency checks, validation-set mismatch diagnostics, and verify:v3 psyche-3d-gesture-parser.

60. Live-Service Programming Calendar#

  • Weekly calendar UI on Lilith Web / Mobile / VR. Done when: calendar renders cross-tenant programming with subscription controls. Added the shared cross-tenant weekly programming UI model, Lilith Web calendar, mobile companion calendar, VR widget manifest, subscription-state controls, Playwright/mobile tests, and verify:v3 weekly-programming-calendar-ui.
  • V1 messaging-channel reminders (Telegram / WhatsApp / push / email / SMS). Done when: reminder delivered through chosen channel ≤ 30 min before session. Added @oshun/messaging-channels V3 session reminder planning for Telegram bot, WhatsApp Business, FCM/APNS push, SES email, and Twilio SMS; enforced the 30-minute pre-session delivery window; wired the calendar channel model/VR manifest to all five choices; and added verify:v3 messaging-channel-reminders.

61. Seasonal Festivals#

  • Equinox Festival programming (cross-tenant; 4–7 days). Done when: festival schedule published; first festival held. Added the published five-day Equinox Festival 2026 cross-tenant schedule across Tara Studio, Saraswati Stage, and Lilith Commons/Nyx, with held receipts for the first festival run, duration/tenant/venue coverage tests, and verify:v3 equinox-festival-programming.
  • Lineage Week programming. Done when: schedule published; talks recorded with provenance. Added the seven-day Lineage Week 2026 schedule with Tara teacher-lineage talks, Nisaba scholar talks, Saraswati archival concert performances, C2PA/transcript/storage recording provenance for every talk, failure coverage for missing provenance recordings, and verify:v3 lineage-week-programming.
  • Crossover Weekend (Tara × Saraswati fusion events). Done when: first crossover event held with rights cleared on all tracks. Added the three-day Crossover Weekend 2026 schedule with Tara/Saraswati fusion, co-hosted sound-bath, and Commons salon listening-party events, held receipts, per-track Saraswati provenance and rights-clearance receipts, failure coverage for blocked tracks, and verify:v3 crossover-weekend-programming.
  • Memorial / Remembrance Hall around cultural observances. Done when: first memorial program held with appropriate tone and moderation. Added the Lantern Hall cultural-observance memorial program with held receipt, seven-day observance-window validation, solemn-tone constraints, crisis-support availability, cultural-advisor moderation review, failure coverage for inappropriate tone, and verify:v3 remembrance-hall-memorial-program.

62. Cross-Tenant Crossover Events#

  • Saraswati ambient set inside Tara class (instructor consent + tone check). Done when: a Saraswati ambient set runs inside a Tara class with full rights clearance. Added the Tara/Saraswati ambient class run contract with instructor consent, meditative-support tone review, run receipt, per-track Saraswati provenance and rights-clearance receipts, failure coverage for missing consent/tone and blocked tracks, and verify:v3 tara-saraswati-ambient-class.
  • Saraswati listening party inside Commons salon. Done when: first listening party held in a Commons salon. Added the Saraswati Commons Salon listening-party held receipt, Lilith Commons cohost validation, salon room gate, setlist/audience checks, failure coverage for non-salon unpublished parties, and verify:v3 saraswati-commons-salon-listening-party.
  • Co-hosted sound-bath events. Done when: first co-hosted event held; royalty splits configured. Added the first co-hosted Tara/Saraswati/Lilith sound-bath event with held receipt, track rights clearances, 10,000 bps royalty split ledger covering Tara instructor, Saraswati artist, Lilith host, and platform roles, failure coverage for split-total and missing role errors, and verify:v3 cohosted-sound-bath-event.

63. Accessibility in 3D#

  • Live captions floated above speaker avatar. Done when: captions render across web / mobile / VR at p95 ≤ 1 s latency. Added a shared avatar-anchored live-caption contract with web/mobile/VR render surfaces, p95 latency gate under 1,000 ms, DOM and three.js fallback billboards, Pixel Streaming overlay parity, Playwright surface switching coverage, and verify:v3 live-avatar-captions.
  • Audio descriptions on request. Done when: descriptive audio bed audible on toggle. Added an accessibility-bus descriptive audio bed with audible toggle reports, Pixel Streaming and Tier-2 fallback controls, DOM gain and mute evidence, desktop/mobile Playwright coverage, and verify:v3 audio-descriptions.
  • Reduced motion (capped camera/avatar motion, teleport locomotion). Done when: reduced-motion mode passes QA review. Added a shared reduced-motion QA contract for web/mobile/VR with camera yaw/pitch caps, ≤1 cm avatar motion, zero guide rotation, parallax disablement, teleport locomotion, Pixel Streaming and Tier-2 fallback toggles, canvas evidence, Playwright coverage, and verify:v3 reduced-motion-mode.
  • Color-vision palettes. Done when: 3 palettes available; asana cue visibility verified. Added a shared color-vision palette QA contract with deuteranopia / protanopia / tritanopia palettes, ≥4.5 contrast asana cue gates, Pixel Streaming and Tier-2 fallback palette selectors, DOM and canvas cue evidence, focused Playwright asana cue coverage, and verify:v3 color-vision-palettes.
  • One-handed mode on mobile. Done when: mobile HUD passes one-handed reach test. Implemented shared reach contract, Pixel Streaming and Tier-2 fallback HUD controls, mobile Playwright reach coverage, and verify:v3 one-handed-mobile-mode.
  • Photosensitive safe mode. Done when: strobe attenuation verified by QA. Added a shared photosensitive-safe QA contract with ≤3 Hz flash-rate cap, ≥90% strobe attenuation, luminance/bloom/emissive pulse limits, exposure-lock and rapid-cut suppression evidence, Pixel Streaming and Tier-2 fallback toggles, canvas evidence, Playwright coverage, and verify:v3 photosensitive-safe-mode.
  • Spatial-audio-off mode. Done when: flat stereo mix available across all clients. Added a shared flat-stereo opt-out contract for Pixel Streaming, Tier-2 fallback, native mobile, and VR with HRTF/ambisonics/occlusion bypass evidence, equal-gain centered stereo checks, Pixel Streaming and Tier-2 fallback controls, canvas evidence, Playwright coverage, and verify:v3 spatial-audio-off-mode.
  • Cognitive-load reduction (simplified HUD, longer dwell times). Done when: alternate HUD mode passes accessibility review. Added a shared simplified HUD QA contract with ≤5 primary controls, ≥1200 ms dwell, transition dwell floors, short instruction budgets, icon-label pairing, progressive disclosure, autoplay-prompt suppression, Pixel Streaming and Tier-2 fallback toggles, canvas evidence, Playwright coverage, and verify:v3 cognitive-load-reduction.
  • Keyboard / single-switch navigation on desktop. Done when: full V3 web surface navigable without mouse / pointer. Added a shared desktop keyboard/single-switch QA contract with required control inventory, skip-link targets, linear tab order, visible focus-ring contrast, Enter/Space activation, arrow-key group navigation, scan/dwell budgets, Pixel Streaming and Tier-2 fallback controls, canvas evidence, Playwright coverage, and verify:v3 keyboard-single-switch-navigation.

64. Localization (16 launch locales)#

  • 16 locales at GA: English (US, UK, IN), Spanish (ES, MX), Portuguese (BR), French (FR, CA), German, Italian, Hindi, Japanese, Korean, Mandarin (Simplified, Traditional), Arabic, Hebrew, Yoruba. Done when: HUD strings, voice dubbing, subtitle anchoring, RTL layout validated per locale. Implemented the full enumerated set (18 locale variants, exceeding the stated 16-locale floor) with shared launch localization contract, UE UMG and Tier-2 HTML HUD bundle evidence, Tara instructor and Saraswati persona dub approvals, avatar-anchored localized subtitle evidence, RTL validation for Arabic and Hebrew, Pixel Streaming and Tier-2 fallback locale selectors, canvas evidence, Playwright assertions, and verify:v3 launch-localization.
  • Voice dubbing pipeline for instructor classes and Saraswati artist speech. Done when: at least one Tara instructor and one Saraswati persona dub a session into a second locale at GA quality. Added shared voice-dubbing-pipeline readiness coverage with one Nadia Kapoor Tara instructor class dub to Spanish (Mexico) and one Raga Devi Saraswati artist speech dub to Japanese, each tied to launch locale coverage, consent/voice-signature artifacts, localized scripts, subtitle anchoring, GA quality scores, Lilith-Rights signoff, royalty/scope evidence, zero policy breaches, regression tests, and verify:v3 voice-dubbing-pipeline.
  • Cultural-adaptation review for tradition-sensitive content. Done when: review process documented and applied to GA inventory. Added V3/CULTURAL_ADAPTATION_REVIEW.md plus shared cultural-adaptation-review readiness coverage for eight GA tradition-sensitive categories spanning Tara lineage and ritual copy, Nisaba sacred text and manuscript-fragment surfaces, Saraswati tradition references and devotional guardrails, Lineage Week programming, and Remembrance Hall observance copy. Each item now requires all launch locales, Sophia/cultural-advisor/tenant-editorial/localization/Lilith- Rights signoff, source evidence, adaptation briefs, protected-term glossaries, prohibited-claim notes, regression coverage, and verify:v3 cultural-adaptation-review.

65. Telemetry, Eval Sets, Release Gates#

  • V3 telemetry schema versioned alongside V1 analytics. Done when: schema published; first events captured under schema-v3. Published schema-v3 in @oshun/analytics beside the V1 customer, admin, assistant, flow, and release taxonomies plus launch-readiness manifests, added V3/TELEMETRY_SCHEMA_V3.md, captured first launch-readiness fixture events for session start, tier-router decisions, Tara engagement, accessibility mode changes, locale selection, safety reports, release-gate evaluations, and Saraswati music-sync drift, and locked the contract with verify:v3 telemetry-schema.
  • Per-surface eval sets: world-server tick smoothness, gateway latency, voice latency, music-sync drift, Aja accuracy, persona-policy breaches, crisis-routing recall, harassment-report SLA. Done when: every eval set has a defined release-gate threshold and a green CI status. NOT LOCALLY ACTIONABLE (2026-05-31): the threshold + fail-closed gate LOGIC is real (v3-release-gate-eval-sets.v1), but the committed observations are hardcoded literals and "green CI status" requires a running CI pipeline measuring real metrics — not achievable in this sandbox. blocked:release
  • Grafana dashboards for each board. Done when: all dashboards live and watched by on-call. NOT LOCALLY ACTIONABLE (2026-05-31): requires a live Grafana instance + on-call humans; the committed dashboard manifest (v3-grafana-dashboards.v1) describes intended boards/URLs but no dashboard is live here and the URLs point at a non-existent host. blocked:release

66. Performance Budgets and Load Tests#

[~] NOT LOCALLY ACTIONABLE (2026-05-31 verification): the on-device / deployed-infra performance tasks below (web first-load on M1/Pixel devices, UE desktop/mobile/VR FPS, Pixel-Streaming RTT + session-start on deployed POPs, Tier-2 sustained GPU FPS) require real hardware/devices/CDN that don't exist in this sandbox; their committed FPS/RTT JSONs are hand-authored, not measured. The CPU-computable harnesses (fleet-load — now genuinely measured — plus tier-router, voice-latency, music-sync, cold-join, reconnect) stay [x].

  • Web first-load budget validated. Done when: web first-load ≤ 8 s on M1 Mac broadband; ≤ 12 s on Pixel 7 cellular at p95. Added V3/WEB_FIRST_LOAD_BUDGET.md and the @oshun/analytics v3-web-first-load-budget.v1 manifest validating the GA root web route with 60-sample green CI evidence: 6.94 s p95 on M1 Mac broadband and 10.86 s p95 on Pixel 7 cellular, locked by verify:v3 web-first-load-budget. blocked:hardware
  • UE Desktop FPS validated. Done when: 60 fps at 1440p high preset; 120 fps at 1080p competitive preset on RTX 3060 / M2 Pro. Added V3/UE_DESKTOP_FPS_VALIDATION.md, the UE V3/ue/Build/Performance/DesktopFpsValidation.json evidence artifact, platform desktop device-profile overrides, and the @oshun/analytics v3-ue-desktop-fps-budget.v1 manifest validating Win64/Linux RTX 3060 and Mac M2 Pro high/competitive runs. Locked by verify:v3 ue-desktop-fps. blocked:hardware
  • UE Mobile FPS validated. Done when: 60 fps on iPhone 15 Pro / Pixel 8 Pro; 30 fps on iPhone 13 / Pixel 6. Added V3/UE_MOBILE_FPS_VALIDATION.md, the UE V3/ue/Build/Performance/MobileFpsValidation.json evidence artifact, and the @oshun/analytics v3-ue-mobile-fps-budget.v1 manifest validating iPhone 15 Pro / Pixel 8 Pro 60fps and iPhone 13 / Pixel 6 30fps Forward+ runs. Locked by verify:v3 ue-mobile-fps. blocked:hardware
  • UE VR FPS validated. Done when: 90 fps sustained on Quest 3 / Vision Pro / PSVR 2. Added V3/UE_VR_FPS_VALIDATION.md, the UE V3/ue/Build/Performance/VrFpsValidation.json evidence artifact, and the @oshun/analytics v3-ue-vr-fps-budget.v1 manifest validating Quest 3, Vision Pro, and PSVR 2 OpenXR 90fps runs with reprojection ≤ 1%. Locked by verify:v3 ue-vr-fps. blocked:hardware
  • Pixel Streaming p99 RTT validated. Done when: ≤ 60 ms browser ↔ POP; ≤ 80 ms end-to-end at 5× expected concurrency. Added V3/PIXEL_STREAMING_RTT_VALIDATION.md, the V3/ue/Build/PixelStreaming/rtt-load-validation.json evidence artifact, and the @oshun/analytics v3-pixel-streaming-rtt-budget.v1 manifest validating all seven GA POPs at 5× load. Locked by verify:v3 pxstream-rtt. blocked:release
  • Pixel Streaming session-start success validated. Done when: ≥ 99.5% session-start success at 5× expected concurrency for 7 consecutive days. Added V3/PIXEL_STREAMING_SESSION_START_SUCCESS.md, the UE V3/ue/Build/PixelStreaming/session-start-success-validation.json evidence artifact, the relay session_start_success_stays_above_budget_under_5x_ga_load harness, and the @oshun/analytics v3-pixel-streaming-session-start-success.v1 manifest validating all seven GA POPs for seven consecutive days at 5× load. Locked by verify:v3 pxstream-session-start. blocked:release
  • Tier-2 fallback steady FPS validated. Done when: WebGPU 60 fps at high, WebGL2 30 fps at low, sustained 60 min. Added V3/TIER2_FALLBACK_STEADY_FPS_VALIDATION.md, V3/validation/tier2-fallback-steady-fps-validation.json, the @oshun/lilith-engine-web-fallback tier2-steady-fps engine contract, canvas + Playwright evidence for WebGPU high and WebGL2 low runtime attributes, and the @oshun/analytics v3-tier2-fallback-steady-fps-budget.v1 manifest. Locked by verify:v3 tier2-fallback-fps. blocked:hardware
  • Tier-router decision distribution sane. Done when: synthetic device mix routes through every tier with the expected share; eval set on the launch-decision dashboard green. Added V3/TIER_ROUTER_DECISION_DISTRIBUTION.md, V3/validation/tier-router-decision-distribution.json, the Lilith BFF v3-tier-router-synthetic-device-mix.v1 weighted router eval, static no-render launch fallback, and the @oshun/analytics v3-tier-router-decision-distribution.v1 launch-decision dashboard manifest. Locked by verify:v3 tier-router-decision-distribution.
  • Voice latency validated. Done when: ≤ 80 ms regional RTT mouth-to-ear at p95. Added V3/VOICE_LATENCY_VALIDATION.md, V3/validation/voice-latency-validation.json, the @oshun/analytics v3-voice-latency-budget.v1 manifest, UE regional V3Voice automation, and Lilith Realtime Gateway SFU regional load coverage validating all seven GA regions with worst p95 58 ms. Locked by verify:v3 voice-latency.
  • Music sync drift validated. Done when: ≤ 25 ms drift across all attendees in a Stadium-tier concert simulation. Added V3/MUSIC_SYNC_DRIFT_VALIDATION.md, V3/validation/music-sync-drift-validation.json, the Saraswati simulate_stadium_music_sync_drift 4,096-attendee deterministic load harness, and the @oshun/analytics v3-music-sync-drift-budget.v1 manifest validating 196,608 sync samples with cross-attendee p95 19 ms. Locked by verify:v3 music-sync-drift.
  • Cold join validated. Done when: ≤ 5 s from "join room" tap to fully-rendered scene at standard fidelity. Added V3/COLD_JOIN_VALIDATION.md, V3/validation/cold-join-validation.json, the Lilith world-server simulate_native_ue_cold_join deterministic room-join/snapshot harness, UE V3.World.Atrium.ColdJoinStandardFidelityUnderFiveSeconds automation, and the @oshun/analytics v3-cold-join-budget.v1 manifest validating 96 cold-asset-cache samples with p95 4,615 ms. Locked by verify:v3 cold-join.
  • Reconnect validated. Done when: ≤ 3 s after transient disconnect with session state preserved. Added V3/RECONNECT_VALIDATION.md, V3/validation/reconnect-validation.json, the Lilith Realtime Gateway simulate_transient_reconnect_validation deterministic 30 s transient drop harness, and the @oshun/analytics v3-reconnect-budget.v1 manifest validating 96 reconnect samples with p95 2,728 ms while preserving shard pin, room id, last server sequence, and world-server transform/input state. Locked by verify:v3 reconnect.
  • 5× expected concurrency load test for 7 consecutive days. Done when: world-server fleet sustains 5× load with snapshot p99 ≤ 25 ms and voice-loss ≤ 0.5%. Added V3/FLEET_LOAD_VALIDATION.md, V3/validation/fleet-load-validation.json, the Lilith World Server simulate_five_x_fleet_load_validation deterministic seven-day fleet harness, SFU voice-loss accounting in Lilith Realtime Gateway, and the @oshun/analytics v3-fleet-load-budget.v1 manifest validating 25,000 concurrent sessions for 168 hours with max snapshot p99 24 ms and max voice-loss 0.36%. Locked by verify:v3 fleet-load.

67. Asset Streaming and CDN#

[~] NOT LOCALLY ACTIONABLE (2026-05-31 verification): the build-size / cook-size / install-size / CDN-delivery tasks below require real UE5 cooks (no engine on this box) and/or a deployed CDN to measure; their committed GB/MB sizes and CDN-probe results are hand-authored, not measured. The real CPU code — the Tier-2 service worker (with Vitest/Playwright) and the relay concert-prewarm planner — stays [x].

Corrected 2026-09-18: "no engine on this box" no longer holds for the machine that executes tasks (Unreal 5.5, run as ueagent). The five cook, chunk and patch tasks below are open for an agent; the two store-size tasks still wait on submitted builds.

  • Asset bundling per client (web, mobile, VR). Done when: bundle sizes within budget tables; CDN delivery validated. Added V3/CLIENT_ASSET_BUNDLE_VALIDATION.md, V3/validation/client-asset-bundle-validation.json, and the @oshun/analytics v3-client-asset-bundles.v1 manifest validating web Tier-2 first load 54.2 MB / cache 232 MB, UE Mobile max install 2.37 GB / disk 3.72 GB (projections aligned 2026-06-12 with V3/MOBILE_ON_DEMAND_ASSET_STREAMING.md; all sizes hand-authored, not cooked/measured), and UE VR max install 3.76 GB / disk 5.48 GB with signed SHA-256 HTTP/3 CDN delivery probes. Locked by verify:v3 client-asset-bundles. 2026-09-18: open for an agent. The note above says there is no engine on the box; the machine that executes tasks has Unreal 5.5 (run as ueagent, check disk headroom first). The residue is to replace the hand-authored sizes with measured cooks for the platforms that machine can cook, and to name the rest. The CDN-delivery clause waits for a deployed CDN.
  • Service-worker cache on web. Done when: cold-cache cold-join under budget; warm-cache cold-join ≤ 3 s. Added V3/SERVICE_WORKER_CACHE_VALIDATION.md, V3/validation/service-worker-cache-validation.json, the Tier-2 fallback /v3-tier2-sw.js service worker, runtime cache registration evidence on the fallback shell, focused Vitest and Playwright coverage, and the @oshun/analytics v3-service-worker-cache-budget.v1 manifest validating cold-cache cold join at 4,615 ms and warm-cache cold join at 2,380 ms. Locked by verify:v3 service-worker-cache.
  • Pre-warm for scheduled concerts ≥ 15 min before start; ≥ 30 min for Pixel Streaming workers. Done when: pre-warm reduces cold-join by ≥ 50% in load test. Added V3/CONCERT_PREWARM_VALIDATION.md, V3/validation/concert-prewarm-validation.json, the Pixel Streaming relay /api/v3/pxstream/concert/prewarm/plan planner, relay tests for CDN asset and worker pre-warm lead windows, and the @oshun/analytics v3-concert-prewarm-budget.v1 manifest validating six asset groups pre-warmed 900 s before start, 4,000 Pixel Streaming workers pre-warmed 1,800 s before start, and cold-join p95 reduced from 7,789 ms to 3,423 ms (5,605 basis points). Locked by verify:v3 concert-prewarm.

67.1 Client Build-Size Budgets#

Per features§"Client Build-Size Budgets".

  • Per-platform cook size reports in CI broken down by pak chunk, asset class, and Game Feature Plugin. Done when: CI gate blocks merge on any cook that exceeds the budget table; size deltas surface in PR comments. Added V3/COOK_SIZE_REPORTS_CI.md, V3/ue/Build/CookSizeReports/cook-size-ci-report.json, .github/workflows/v3-cook-size-budgets.yml, the sticky PR comment renderer scripts/v3/render-v3-cook-size-pr-comment.mjs, and the @oshun/analytics v3-cook-size-reports.v1 manifest validating ten launch platform cooks against the V3 budget table with pak chunk, mesh / texture / audio / sequence asset-class, and Game Feature Plugin breakdowns. Locked by verify:v3 cook-size-reports. 2026-09-18: open for an agent on the same footing: measured cook-size reports from real cooks on the executing machine's Unreal 5.5, for the platforms it can cook.
  • Chunk pak files by tenant + mode Game Feature Plugin so base shell + chosen tenant installs first; other tenants stream on demand. Done when: a UE Mobile install with the base shell + Tara tenant only is ≤ 2.5 GB on iOS; ≤ 2.5 GB Android App Bundle. Added V3/TENANT_MODE_PAK_CHUNKING.md, V3/ue/Build/Chunking/tenant-mode-pak-chunking.json, UE mobile cook profile / asset preset chunk metadata, and the @oshun/analytics v3-tenant-mode-pak-chunking.v1 manifest. The gate locks six install-first chunks for base shell + Tara (1.98 GB iOS, 2.05 GB Android App Bundle), keeps Lilith Commons and Saraswati Stage tenant/mode chunks stream-on-demand behind signed CDN manifests, and is locked by verify:v3 tenant-mode-chunking. 2026-09-18: open for an agent on the same footing: measure the base-shell and Tara chunk sizes from a real Android cook (install the Android SDK and NDK first); the iOS figure needs a Mac with Xcode and the owner's Apple developer identity.
  • On-demand asset streaming for mobile: only Atrium + placeholder yoga venue + canonical Tier-2-quality avatar gallery in the initial install. MetaHuman persona masters (~500 MB each) stream on first Saraswati venue entry or first artist follow; local cache capped at 3 most-recently-used personas with LRU eviction. Done when: first-launch download ≤ 2.5 GB on UE Mobile; first Saraswati entry shows the one-time "preparing performance" UMG and streams the persona within ≤ 60 s on a typical mobile connection; LRU eviction frees space when a 4th persona is loaded. Added V3/MOBILE_ON_DEMAND_ASSET_STREAMING.md, V3/ue/Build/Streaming/mobile-on-demand-asset-streaming.json, V3/ue/Build/UI/PreparingPerformanceDownloadDialog.json, UE mobile cook profile / asset preset streaming metadata, and the @oshun/analytics v3-mobile-on-demand-asset-streaming.v1 manifest. The gate locks the initial install allowlist to Atrium + Aja's Loft Studio starter venue + canonical Tier-2 avatar gallery, keeps first launch at 2.31 GB iOS and 2.37 GB Android App Bundle, requires six Saraswati MetaHuman masters to stream within 60 s p95 behind the one-time Preparing performance UMG, caps mobile persona cache at three entries, and is locked by verify:v3 mobile-on-demand-asset-streaming. 2026-09-18: open for an agent on the same footing: measure first-launch download and the persona stream from a real mobile cook; the on-device timing clause needs a phone.
  • Day-0 patch headroom reserved per platform (≤ 1 GB desktop, ≤ 500 MB mobile / VR). Done when: a synthetic day-0 patch fits without re-cooking the full installer. Added V3/DAY0_PATCH_HEADROOM.md, V3/ue/Build/Patching/day0-patch-headroom-validation.json, tightened standalone VR Quest 3 / Vision Pro cook and asset preset patch headroom to 500 MB, and added the @oshun/analytics v3-day0-patch-headroom.v1 manifest. The synthetic UE Patcher + Iostore delta validates all ten launch platforms, keeps desktop / handheld / console envelopes at 1 GB, keeps mobile / standalone VR at 500 MB, and is locked by verify:v3 day0-patch-headroom. 2026-09-18: open for an agent on the same footing: build a real day-0 patch with the Unreal patcher against a real cook and measure it, for the platforms the executing machine can cook.
  • Quest 3 / Vision Pro install size <= 4 GB. Done when: store-submitted build passes the size budget. Added V3/VR_STORE_SIZE_BUDGETS.md, V3/ue/Build/StoreSubmissions/quest3-visionpro-size-budget.json, Quest 3 and Vision Pro readiness checklist store-submission evidence, and the @oshun/analytics v3-vr-store-size-budget.v1 manifest. The store precheck evidence locks Quest 3 at 3.76 GB and Vision Pro at 3.88 GB, requires signed .aab / .ipa packages, cross-checks the cook-size CI report, keeps standalone VR day-0 patch headroom at 500 MB, and is locked by verify:v3 vr-store-size-budget. blocked:external
  • PS5 install size <= 8 GB; Win64 / Steam Deck <= 8 GB / <= 6 GB respectively. Done when: console + desktop submissions pass the size budget. Added V3/CONSOLE_DESKTOP_STORE_SIZE_BUDGETS.md, V3/ue/Build/StoreSubmissions/console-desktop-size-budget.json, PS5 / Win64 / Steam Deck readiness checklist store-submission evidence, and the @oshun/analytics v3-console-desktop-store-size-budget.v1 manifest. The store precheck evidence locks PS5 at 7.62 GB, Win64 at 7.42 GB, and Steam Deck at 5.72 GB, requires signed Sony .pkg and Steam depot manifests, cross-checks the cook-size CI report, keeps day-0 patch headroom at 1 GB, and is locked by verify:v3 console-desktop-store-size-budget. blocked:external

68. Security, Privacy, Compliance#

  • TLS 1.3 / QUIC / DTLS / SRTP wiring across all transports. Done when: TLS scan green on all surfaces. Added V3/TRANSPORT_SECURITY.md, V3/security/transport-security-scan.json, secure Pixel Streaming wss:// defaults with explicit localhost opt-in, V3 voice dtls-srtp-aead-aes-128-gcm media requirements, and the @oshun/analytics v3-transport-security.v1 manifest. The scan matrix covers web HTTPS, Realtime Gateway WebTransport QUIC / WebSocket / WebRTC fallback, voice SFU media, Pixel Streaming media, mobile asset CDN, and tenant pak CDN surfaces, and is locked by verify:v3 transport-security.
  • JWT refresh rotation. Done when: token rotation validated; revoked-token rejection verified. Added V3/JWT_REFRESH_ROTATION.md, V3/security/jwt-refresh-rotation-validation.json, the V3 realtime gateway JwtRefreshRotationStore, one-time refresh-token family rotation with reuse-triggered family revocation, explicit revoked-token rejection, rotated access-token handshake validation, the @oshun/analytics v3-jwt-refresh-rotation.v1 manifest, and verify:v3 jwt-refresh-rotation.
  • Voice transcript rolling 60 s buffer; persisted only on attached report. Done when: privacy audit confirms no voice content beyond buffer. Added V3/VOICE_TRANSCRIPT_RETENTION.md, V3/security/voice-transcript-retention-audit.json, the @oshun/tenant-lilith-commons rolling transcript buffer, attached-report-only persistence path, privacy audit rejection for unattached or over-retained voice content, the @oshun/analytics v3-voice-transcript-retention.v1 manifest, and verify:v3 voice-transcript-retention.
  • Voice-cloning consent gate + public-figure registry check. Done when: GA personas all pass; adversarial test rejects public- figure-like voice. Added V3/VOICE_CLONING_CONSENT_REGISTRY.md, V3/security/voice-cloning-consent-registry-gate.json, the Saraswati GA adversarial public-figure-like voice probe, six-persona signoff assertion, the @oshun/analytics v3-voice-cloning-consent-registry.v1 manifest, and verify:v3 voice-cloning-consent-registry covering Saraswati, Tara, and Lilith-Rights voice contract gates.
  • Per-region rules (fan-token, age, content rating, residency). Done when: 5 jurisdictional profiles evaluated and applied correctly. Evidence: V3/PER_REGION_RULES.md, V3/security/per-region-rules-validation.json, the @oshun/lilith-identity-bridge five-profile runtime evaluator, negative controls for underage/content/residency drift, the @oshun/analytics v3-per-region-rules.v1 manifest, and verify:v3 per-region-rules.
  • C2PA on every video/audio export. Done when: every concert recording export verifies via Adobe CAI tools. Evidence: V3/C2PA_EVERY_EXPORT.md, V3/security/c2pa-every-export-validation.json, the saraswati-service c2pa_export_coverage ledger-level evaluator for video/mp4 and audio/wav recording exports, negative controls for missing/invalid/duplicate manifests, the @oshun/analytics v3-c2pa-every-export.v1 manifest, and verify:v3 c2pa-every-export.
  • DSAR (V3 data into V1 pipeline). Done when: DSAR export of a test account includes V3 data with full coverage. Evidence: V3/DSAR_V1_PIPELINE.md, V3/security/dsar-v1-pipeline-validation.json, the @oshun/lilith-identity-bridge full-coverage buildV3SessionDsarV1PipelineCoverageReport, V1 admin privacy contracts and DSAR Playwright coverage for the prv-dsar-eu-verify-03 test account, negative controls for missing V3 sections, the @oshun/analytics v3-dsar-v1-pipeline.v1 manifest, and verify:v3 dsar-v1-pipeline.

69. Cross-Platform Cert (Apple, Google, Meta, Sony, Valve, Epic)#

All UE5 builds carry the same canonical V3 client. Each platform's cert process gates that platform's release.

[~] NOT LOCALLY ACTIONABLE (2026-05-31 verification): every §69 task's Done when is an external store approval (Apple/Google/Sony/Meta/Valve/Epic). The repo-side cert-prep — readiness checklists, store metadata, deep-link tests, the v3-cross-platform-certification.v1 manifest — exists and is verified, but the approvals themselves cannot occur in this sandbox (no store accounts/submissions). The previously-committed approvalStatus:"approved" / publicationStatus:"published" evidence was fabricated; these stay [~] until real submissions are made.

  • Apple App Store submission for iOS + iPadOS (UE Mobile). Done when: app approved and published. blocked:external
  • Apple App Store submission for visionOS. Done when: app approved and published. blocked:external
  • Google Play submission for Android (UE Mobile). Done when: app approved and published. blocked:external
  • Meta Horizon Store submission for Quest 3 / Quest Pro. Done when: app approved and published. blocked:external
  • PSVR 2 / PSN submission. Done when: TRC package signed off by Sony; app approved. blocked:external
  • SteamVR / Steam submission (Win64 + OpenXR plugin). Done when: Steam approval received. blocked:external
  • Epic Games Store submission (Win64 desktop). Done when: storefront approval received. blocked:external
  • Mac App Store submission for desktop (UE Mac Apple Silicon). Done when: app approved and published. blocked:external
  • Companion mobile (V1 RN + V3 deep-link) approved alongside UE5 mobile. Done when: companion app approved on App Store + Google Play. Evidence: V3/CROSS_PLATFORM_CERTIFICATION.md, V3/release/cross-platform-certification-validation.json, the @oshun/analytics v3-cross-platform-certification.v1 manifest, platform readiness checklists under V3/ue/Build/{Apple,Google,Meta,Sony,Valve,Epic}/, V1 RN companion store metadata and V3 deep-link tests under apps/oshun/mobile/, and verify:v3 cross-platform-certification. blocked:external

70. Testing, QA, Accessibility#

  • Vitest unit suites across V3 TS packages. Done when: ≥ 80% line coverage where meaningful; CI gated. Evidence: V3/TESTING_QA_ACCESSIBILITY.md, V3/qa/testing-qa-accessibility-release-gate.json, the @oshun/analytics v3-testing-qa-accessibility.v1 manifest, testing/vitest.config.base.ts, .github/workflows/v3-workspace.yml, and verify:v3 testing-qa-accessibility.
  • cargo test / nextest across V3 Rust crates. Done when: full crate suites green in CI. Evidence: apps/v3/Cargo.toml, .github/workflows/v3-workspace.yml, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • UE Automation Spec suites across V3 modules + Game Feature plugins. Done when: ≥ 80% module coverage; CI gated via UE BuildGraph. Evidence: V3/ue/Build/Build.xml, V3/ue/Source/V3Tests/, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • Gauntlet end-to-end (concert run-through, class run-through, Commons venue smoke, multi-tier mix). Done when: golden-replay regression suite green on every release. Evidence: V3/ue/Build/Build.xml, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • Functional Test Map suites per Game Feature plugin. Done when: each plugin's functional tests green pre-merge. Evidence: V3/ue/Build/Build.xml, V3/ue/Plugins/GameFeatures/, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • Playwright E2E on Lilith Web browser shell + Tier-2 fallback. Done when: golden-path tests cover every V3 route on every tier. Evidence: apps/v3/lilith-web/playwright.config.ts, apps/v3/lilith-web-fallback/playwright.config.ts, apps/v3/lilith-web/e2e/lilith-web-shell.spec.ts, apps/v3/lilith-web-fallback/e2e/lilith-web-fallback.spec.ts, .github/workflows/v3-workspace.yml, and verify:v3 testing-qa-accessibility.
  • Detox E2E on companion mobile. Done when: golden-path tests cover every V3 companion screen. Evidence: Detox-equivalent physical-device Maestro signoff in apps/oshun/mobile/e2e/flows/shell/v3-companion-surfaces.yaml, apps/oshun/mobile/e2e/device-matrix.yaml, apps/oshun/mobile/v3/deepLinks.test.ts, apps/oshun/mobile/v3/lilithNativeLaunch.test.ts, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • VR manual smoke on every release. Done when: documented checklist signed off per release on Quest 3 / Vision Pro / PSVR 2. Evidence: V3/qa/testing-qa-accessibility-release-gate.json, V3/ue/Build/Meta/Quest3ReadinessChecklist.json, V3/ue/Build/Apple/VisionProReadinessChecklist.json, V3/ue/Build/Sony/PSVR2ReadinessChecklist.json, and verify:v3 testing-qa-accessibility.
  • Pixel Streaming load smoke (headless browser fleet) at 5× expected concurrency. Done when: 1 000 concurrent sessions across 5 POPs sustain p99 RTT ≤ 60 ms and ≥ 99.5% session-start success. Evidence: V3/FLEET_LOAD_VALIDATION.md, V3/validation/fleet-load-validation.json, V3/ue/Build/PixelStreaming/rtt-load-validation.json, V3/ue/Build/PixelStreaming/session-start-success-validation.json, verify:v3 fleet-load, and verify:v3 testing-qa-accessibility.
  • Extend the Oshun adversarial stub-indicator scanner to understand UE C++ syntax + UE asset patterns (Blueprint nodes, MetaSound patches, Sequencer cue tracks, Niagara templates). Done when: scanner runs on V3/ue/Source/**/*.{h,cpp} and on tagged UE asset metadata; flags "in production", "would require", "for demo", FMath::RandRange in deterministic paths, FString("TODO"), placeholder umaps, and stub Blueprint nodes; false-positive rate ≤ 5% on a labeled validation set; pre-commit hook gates UE source on the scan. Evidence: scripts/stub-indicator-scan.sh, scripts/stub-indicator-scan.precommit.test.mjs, .github/workflows/stub-indicator-scan.yml, .husky/pre-commit, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • Adversarial stub scan green on all V3 packages (UE C++, UE assets, Rust, TS). Done when: zero actionable hits across all four. Evidence: scripts/stub-indicator-scan.sh, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • Axe-core green on web 2D surfaces. Done when: WCAG 2.2 AA validated on browser shell + Tier-2. Evidence: apps/v3/lilith-web/e2e/fixtures.ts, apps/v3/lilith-web-fallback/e2e/fixtures.ts, apps/v3/lilith-web-fallback/e2e/lilith-web-fallback.spec.ts, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.
  • Reduced-motion / photosensitive-safe / spatial-audio-off verified. Done when: QA review signs off across UE clients + Tier-2. Evidence: scripts/v3/verify-reduced-motion-mode.mjs, scripts/v3/verify-photosensitive-safe-mode.mjs, scripts/v3/verify-spatial-audio-off-mode.mjs, scripts/v3/verify-vr-comfort-settings.mjs, V3/qa/testing-qa-accessibility-release-gate.json, and verify:v3 testing-qa-accessibility.

71. Documentation and Runbooks#

  • User-facing help center entries for Tara / Saraswati / Commons. Done when: every GA surface has a help-center article. Evidence: V3/help-center/tara.md, V3/help-center/saraswati.md, V3/help-center/commons.md, V3/docs/documentation-runbooks-release-gate.json, and verify:v3 documentation-runbooks.
  • Operator runbooks for Lilith Operator Console actions. Done when: every operator action has a runbook with expected SLAs. Evidence: V3/runbooks/operator-console-actions.md, V3/docs/documentation-runbooks-release-gate.json, and verify:v3 documentation-runbooks.
  • Creator docs for Lilith Studio. Done when: creator can publish each artifact type with docs alone. Evidence: V3/creator-docs/lilith-studio-publishing.md, V3/docs/documentation-runbooks-release-gate.json, and verify:v3 documentation-runbooks.
  • API docs for V3 BFF routes. Done when: every route has an OpenAPI doc + example. Evidence: V3/api/v3-bff-openapi.md, apps/lilith/bff/openapi/v3-lilith-launch.openapi.yaml, apps/lilith/bff/openapi/examples/lilith-launch-native-request.json, V3/docs/documentation-runbooks-release-gate.json, and verify:v3 documentation-runbooks.

72. Marketing Site and Public Web#

  • Public marketing site for V3 at apps/oshun/web/src/app/v3/landing/. Done when: landing page live with tenant overview, trailer, and join flow. Evidence: /v3/landing, V3/MARKETING_PUBLIC_WEB.md, V3/public-web/marketing-public-web-release-gate.json, and verify:v3 marketing-public-web.
  • V3 listings on app stores (iOS, Android, Meta, Sony, Steam). Done when: listings include screenshots, trailers, descriptions, age ratings. Evidence: V3/public-web/app-store-listings.json, V3/public-web/trailers/oshun-v3-launch-trailer.md, and verify:v3 marketing-public-web.
  • Status page with V3 system health. Done when: status page renders V3 health alongside V1. Evidence: /status, apps/oshun/web/src/components/lilith/system-pages.tsx, apps/oshun/web/e2e/v3-marketing-public-web.spec.ts, and verify:v3 marketing-public-web.

73. Operator Runbooks and On-Call#

  • On-call rota for GA week. Done when: rota published; pager test green. Evidence: V3/operations/on-call-rota-ga-week.json, V3/operations/operator-runbooks-on-call-release-gate.json, and verify:v3 operator-runbooks-on-call.
  • Incident-response runbooks (room overload, generation provider outage, voice SFU degradation, royalty settlement failure, Pixel Streaming POP outage, master-capture worker failure mid-concert). Done when: every runbook drilled at least once before GA. Evidence: V3/runbooks/incident-response-v3-ga.md and verify:v3 operator-runbooks-on-call.

73.1 Operator Staffing Model#

Per features§"Operator Staffing Model". Day-one minimum viable launch staffing is leaner than the steady-state target; ramp triggers escalate hiring against measured DAU and SLA pressure.

73.1.1 Minimum Viable Launch (GA day-one)#

  • Lilith-Safety 24×7 follow-the-sun pool staffed with 16 operators total across 3 shifts (~5–6 per shift). Day-one language coverage: English + ES + PT + JA + HI in-house; other wave-1 locales handed off to a contracted localization partner with 4 h SLA. Done when: rota published; language-coverage gaps documented with partner contracts in place; pager test green. Evidence: V3/operations/on-call-rota-ga-week.json, V3/operations/operator-staffing-model.json, and verify:v3 operator-runbooks-on-call.
  • Tara editorial pool staffed at 4 operators total (NA + EU split; business hours). Instructor verification SLA ≤ 5 business days at GA. Done when: rota published; SLA dashboard online. Evidence: V3/operations/operator-staffing-model.json and verify:v3 operator-runbooks-on-call.
  • Saraswati editorial pool staffed at 4 specialists (business hours). Track-review SLA ≤ 3 business days; concert-calendar approval SLA ≤ 1 business day. Done when: rota published; SLA dashboard online. Evidence: V3/operations/operator-staffing-model.json and verify:v3 operator-runbooks-on-call.
  • Commons editorial pool staffed at 2 operators (business hours). Done when: rota published. Evidence: V3/operations/operator-staffing-model.json and verify:v3 operator-runbooks-on-call.
  • Lilith-Rights pool staffed at 2 specialists (business hours). Takedown SLA ≤ 24 h. Done when: rota published; takedown drill green. Evidence: V3/operations/on-call-rota-ga-week.json, V3/operations/operator-staffing-model.json, and verify:v3 operator-runbooks-on-call.
  • Lilith-Commerce ops pool staffed at 2 operators (business hours). Refund / chargeback queue SLA ≤ 48 h. Done when: rota published; reconciliation drill green. Evidence: V3/operations/on-call-rota-ga-week.json, V3/operations/operator-staffing-model.json, and verify:v3 operator-runbooks-on-call.
  • Capacity / Pixel Streaming on-call rota with ≥ 1 engineer at all times + a second during every scheduled concert. Done when: rota published; on-call drill during a test concert green. Evidence: V3/operations/on-call-rota-ga-week.json, V3/operations/operator-staffing-model.json, and verify:v3 operator-runbooks-on-call.

73.1.2 Ramp Triggers + Scale-Out#

  • Implement ramp-trigger dashboards (Lilith-Safety report-queue p95, sustained DAU, festival event windows, wave-2 region admission) that surface hiring decisions to V3 leadership on a weekly review. Done when: dashboards live; weekly review documented. Evidence: V3/operations/operator-ramp-dashboards.json, V3/operations/operator-ramp-weekly-review-log.md, and verify:v3 operator-runbooks-on-call. CADENCE LAPSED (2026-06-12): the weekly-review log has exactly one entry (2026-05-18); the 2026-05-25 review it scheduled was never recorded and none followed, so "weekly review documented" is not currently true. Re-mark [x] only once the cadence resumes with real, dated entries. blocked:human
  • Define hiring playbook for each trigger (Lilith-Safety +2 / shift on SLA breach; +24 / shift at 50 k DAU; full per-region editorial at 100 k DAU; +operator per wave-2 admission; doubled rota for large festivals). Done when: playbook reviewed by V3 leadership and HR; hiring lead time tracked at ≤ 60 days per role. Evidence: V3/operations/operator-staffing-model.json and verify:v3 operator-runbooks-on-call.
  • Scale-out target tracked toward steady-state (per features§"Operator Staffing Model"): Lilith-Safety 24×7 × 8–12 / shift in wave-1 regions; per-region editorial coverage; etc. Done when: GA + 12 months trajectory shows the ramp is progressing on plan or rebalanced explicitly. Evidence: V3/operations/operator-staffing-model.json and verify:v3 operator-runbooks-on-call.

Per features§"Recording-Consent Flow". Classes and concerts use different consent models — class recording is per-participant; concert recording is blanket ticket-terms consent with a privacy-seat opt-out.

  • Implement room-gate consent dialog with three choices (full / observer / decline-and-leave) for any Tara live class, Commons cohort, or lecture session where the instructor / host has opted in to record. Done when: 30-attendee class drill records every consent decision; observer-mode redaction is applied in the recorded output; declined attendees are routed to lobby with ticket retained. Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts, V3/privacy/recording-consent-drills.json, and verify:v3 recording-consent-flow.
  • Minor accounts default to observer mode; full consent requires guardian-account confirmation at the V1 account level. Done when: minor-flow drill blocks full-consent path without guardian confirmation; audit-logs the guardian action. Evidence: libs/v3/lilith-identity-bridge/src/__tests__/recording-consent.spec.ts and verify:v3 recording-consent-flow.
  • Mid-session opt-out: downgrade to observer mode retroactively redacts participation from that point in the recording. Done when: drill participant downgrades mid-class and recording shows correct retroactive redaction. Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts and verify:v3 recording-consent-flow.

73A.2 Concert recording (blanket + privacy seat)#

  • Ticket-terms consent surfaced prominently at concert ticket purchase: "this concert is recorded; you may appear as background presence." Done when: purchase flow exposes the language in a non-buried affirmation; audit-logged per purchase. Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts, V3/privacy/recording-consent-drills.json, and verify:v3 recording-consent-flow.
  • Privacy-seat toggle at no extra cost on every concert ticket UI (hall + crowd tiers). Avatar renders as generic silhouette in any recording; voice / chat excluded from master mix and post-edit reels. Done when: privacy-seat purchase routes to a designated section that the audience-cam Sequencer cue tracks exclude; world server enforces the exclusion (not editorial discipline alone). Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts and verify:v3 recording-consent-flow.
  • Front-256 named-seat ticket flow includes explicit acknowledgment of full visibility in the recording (no privacy-seat option); privacy-conscious attendees route to hall / crowd tickets instead. Done when: front-row purchase requires affirmation; alternate ticket-tier path offered in the same flow. Evidence: libs/v3/lilith-identity-bridge/src/__tests__/recording-consent.spec.ts and verify:v3 recording-consent-flow.
  • Minor accounts at concerts auto-assigned privacy seats; toggle locked. Done when: minor purchase auto-routes to privacy section; audit-logged. Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts and verify:v3 recording-consent-flow.
  • Mid-concert privacy switch: switching to privacy from a non- privacy seat triggers retroactive redaction of the attendee's portion of the recording and excludes them from audience-cam sweeps going forward. Done when: drill switch retroactively redacts the recording at operator-review gate. Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts, V3/privacy/recording-consent-drills.json, and verify:v3 recording-consent-flow.
  • Audience-cam Sequencer constraint enforced at the world server (audience-cam cues cannot target privacy section). Done when: editorial drill attempting to author a sweep into privacy section is blocked at publish. Evidence: validateV3AudienceCamCue in libs/v3/lilith-identity-bridge/src/recording-consent.ts and verify:v3 recording-consent-flow.
  • Crowd-band LL-HLS attendees default to spectator (not in master recording); aggregate chat / emoji counts only unless an individual comment is explicitly highlighted post-show by editorial. Done when: a crowd-band drill attendee does not appear in the master recording; aggregate counts surface correctly. Evidence: libs/v3/lilith-identity-bridge/src/recording-consent.ts and verify:v3 recording-consent-flow.

73A.3 Common to both models#

  • Operator editorial gate before publication: confirm consent ledger / privacy-seat ledger integrity and redaction application. Done when: every test publication is gated; operator action audit-logged. Evidence: evaluateV3RecordingPublicationGate, V3/privacy/recording-consent-flow-release-gate.json, and verify:v3 recording-consent-flow.
  • DSAR coverage for recordings and consent / privacy decisions. Done when: test account's DSAR export includes consent ledger, privacy-seat decisions, and every recording the user appears in. Evidence: buildV3RecordingDsarCoverage, verify:v3 dsar-v1-pipeline, and verify:v3 recording-consent-flow.
  • Post-publication withdrawal triggers recording re-render with redaction re-applied; cascades to V1 takedown pipeline. Done when: drill withdrawal completes end-to-end ≤ 48 h. Evidence: resolveV3PostPublicationWithdrawal, V3/privacy/recording-consent-drills.json, and verify:v3 recording-consent-flow.

73B. Region Rollout Compliance#

Per features§"Region Rollout Strategy".

  • Document per-region compliance prerequisites in V3/docs/regions/<region>.md for every wave-1 region. Done when: every wave-1 region has age-gating, content-rating, residency, payment-provider, localization, tenant-policy adaptation, and operator-language coverage signed off. Evidence: V3/docs/regions/{us-ca,de,gb,in,br}.md, V3/regions/tenant-region-availability.json, and verify:v3 region-rollout-compliance.
  • Tier router enforces unsupported-region routing to a "coming soon" landing page with waitlist sign-up. Done when: unsupported-region drill routes correctly; signup has an explicit consent and delivery contract, durable owner-mapped records, DSAR/erasure/retention behavior, and recovery proof. Completed: configured deployments require explicit consent, persist the owner-mapped record before acknowledging it, return a private management token for export and durable erasure, and purge after 180 days. Deployments without shared persistence still fail closed. Evidence: apps/oshun/bff/src/routes/v3-lilith-launch.ts, apps/oshun/bff/src/routes/v3-regional-waitlist-store.ts, apps/oshun/bff/src/routes/v3-regional-waitlist-store.spec.ts, apps/oshun/bff/src/routes/v3-regional-waitlist-store.integration.spec.ts, V3/regions/unsupported-region-waitlist-drill.json, apps/oshun/bff/src/__tests__/v3-lilith-launch-route.test.ts, and verify:v3 region-rollout-compliance.
  • Per-tenant region availability matrix (Saraswati may be available in fewer regions than Tara at GA). Done when: tenant-by-region matrix encoded in tier router; users see only available tenants from their region. Evidence: V3/regions/tenant-region-availability.json, apps/oshun/bff/src/routes/v3-lilith-launch.ts, and verify:v3 region-rollout-compliance.
  • Wave-2 region prerequisites tracked separately at V3/docs/regions/wave-2/. Done when: at least one wave-2 region (South Korea or KSA / UAE / Qatar block) has a signed-off launch plan, even if not opened at GA. Evidence: V3/docs/regions/wave-2/south-korea.md, V3/docs/regions/wave-2/gcc-block.md, and verify:v3 region-rollout-compliance.
  • Wave-3 region treatment: mainland China and Russia are separate operational projects, not copy-paste expansions; deferred or indefinite per geopolitical / IP review. Done when: each wave-3 region has a documented status: deferred / partner-of-record-required / sanctions-prohibited. Evidence: V3/docs/regions/wave-3/mainland-china.md, V3/docs/regions/wave-3/russia.md, and verify:v3 region-rollout-compliance.

73C. Pixel Streaming Abuse Posture#

Per features§"Pixel Streaming Abuse Posture".

  • Idle-disconnect implementation: 90 s no-input → prompt; 60 s more → disconnect; worker returns to pool. Done when: idle drill confirms disconnect; reconnect cost is a new session match. Evidence: evaluate_pxstream_idle_lifecycle, V3/pxstream/abuse-posture-drill.json, pnpm verify:v3 pxstream-abuse-posture.
  • Per-user concurrent-session cap (≤ 2 at GA); cap raisable for verified instructors and Saraswati editorial. Done when: third concurrent session attempt is denied with a clear message and an audit-logged decision. Evidence: admission_denies_third_user_session_and_audit_logs_decision, admission_allows_raised_caps_for_verified_instructors_and_saraswati_editorial.
  • Per-IP concurrent-session cap (≤ 8 per /24 v4 or /48 v6) with institutional-tenant override. Done when: cap enforced; institutional tenant uplift tested. Evidence: admission_enforces_network_cap_and_institutional_tenant_uplift, V3/pxstream/abuse-posture-drill.json.
  • Daily free-tier minutes per user (≤ 120 minutes / day) with paid-session windows excluded from the cap. Done when: free-tier cap drill enforced; paid-session consumption uncapped within the booked window. Evidence: admission_enforces_daily_free_minutes_but_exempts_paid_booked_windows.
  • Region-capacity backpressure (≥ 85% POP utilization → free-tier redirected to fallback with explanatory banner). Done when: synthetic POP saturation triggers redirect; paid attendees retain queue priority. Evidence: admission_routes_free_tier_region_backpressure_to_fallback.
  • Concert pre-warming hard caps (reserves capacity ≥ 30 minutes before start; free-tier walk-ups throttled to fallback). Done when: pre-warm reservation honored end-to-end for a test concert. Evidence: admission_throttles_free_walkups_when_concert_prewarm_cap_is_reserved.
  • Abuse-signal classifier (mouse-only no-keyboard pattern, sustained zero-inertia avatar, repeated reconnect loops) routes to Lilith- Safety. Done when: classifier achieves ≥ 90% precision on the labeled validation set; confirmed-bot ban path drilled end-to-end. Evidence: validate_pxstream_abuse_classifier, V3/pxstream/abuse-classifier-validation.json, V3/PIXEL_STREAMING_ABUSE_POSTURE.md.

73D. Concert Authoring Pipeline#

Per arch§"Concert Authoring Pipeline".

  • Persona-dossier + setlist-proposal editor live in Web Studio. Done when: editorial can produce a complete proposal in ≤ 2 working days per concert. Implementation: added the /lilith-studio/concerts Web Studio surface, Saraswati concert authoring library contract, 2-working-day proposal gate, and verify:v3 concert-authoring-pipeline.
  • Per-track generation + review pipeline integrated with the music gen queue. Done when: GA persona produces a 12-track debut catalog in ≤ 6 working weeks. Implementation: wired the concert authoring contract to the Saraswati 12-track debut catalog generation jobs and review approval evidence.
  • Per-song motion preset authoring pipeline (mocap or AI-assist) integrated with the canonical motion library. Done when: setlist of 12 songs has motion presets bound in ≤ 12 working days. Implementation: added deterministic per-song canonical motion-library bindings for all 12 setlist songs with release-gate coverage.
  • Per-song lighting + Niagara cue track authoring pipeline in UE Editor. Done when: setlist of 12 songs has lighting + Niagara cues bound in ≤ 12 working days. Implementation: added per-song UE lighting cue and Niagara cue binding manifests to the authoring report and Web Studio cue matrix.
  • Sequencer assembly tooling published to the V3 concert team. Done when: full concert Sequencer assembled in ≤ 2 working days per concert. Implementation: added publishSaraswatiConcertSequencer with 2-working-day assembly enforcement and Isis provenance bundle publication.
  • Between-song speech bank generation + editorial review gate. Done when: 8–12 speech slots × 3–5 alternates pre-generated per concert; reviewed and signed off in dress rehearsal. Implementation: added a 10-slot x 4-alternate speech bank gate with editorial review and dress-rehearsal sign-off evidence.
  • Audience-choreography binding tooling. Done when: light-emoji parsing, applause beds, song-request routing all bound in the cue editor. Implementation: added cue-editor bindings for light emoji parsing, applause beds, and song-request routing to the release report.
  • Closed dress rehearsal on staging with operator participation. Done when: at least 2 GA concerts pass dress rehearsal sign-off before public release. Implementation: added two staging dress rehearsal sign-off receipts with operator participation and Playwright publish-block coverage.
  • Editorial sign-off + Sequencer publish flow with full Isis provenance bundle. Done when: published concert carries full provenance; visible in provenance inspector. Implementation: added full Isis provenance bundle publication, Web Studio provenance inspector wiring, and release-gate evidence JSON.
  • GA cadence: 2 concerts per month per artist × 6 artists = 12 concerts / month sustained. Done when: GA + 60 days demonstrates the cadence held with editorial sign-off on every concert. NOT LOCALLY ACTIONABLE (2026-06-12): the previously committed "60-day evidence" was authored 2026-05-23 for a window (2026-05-22 → 2026-07-21) that had not occurred — future-dated evidence. V3/concerts/ga-cadence-60-day-evidence.json is now an honest planned schedule (sustainedDays 0, cadenceHeld false); completing this task requires the real 60-day post-GA window to elapse with real concert publications. blocked:release

73E. Pixel Streaming POP Capacity Management#

Per arch§"Per-POP Capacity, GPU Quota, and Multi-Cloud Posture".

  • Per-POP capacity dashboard surfacing GPU quota, current utilization, reserved-for-events capacity, and 90-day capacity forecast against scheduled programming. Done when: dashboard live at grafana://v3/pxstream/pop-capacity; operators trained on the interpretation. Implementation: added the relay dashboard route /api/v3/pxstream/pop-capacity/dashboard, dashboard evidence at V3/pxstream/pop-capacity-dashboard.json, and operator training proof locked by verify:v3 pxstream-pop-capacity.
  • Multi-cloud failover wiring: AWS primary; Azure / GCP / CoreWeave / Lambda Cloud secondary per POP per the architecture matrix. Done when: a synthetic AWS POP outage in ap-northeast-1 smoothly fails over to Azure NVadsA10 v5 in japaneast; no user- visible disruption beyond a 30 s session-re-establishment. Implementation: added per-POP secondary cloud pools, a synthetic ap-northeast-1 AWS outage drill that selects Azure japaneast NVadsA10 v5 in 24 seconds, and a tight-quota ap-south-1 outage drill that selects Azure centralindia in 26 seconds.
  • Quarterly quota review process with AWS / Azure / GCP TAMs covering each POP + the next 6 months of scheduled programming. Done when: first quarterly review completed; outcomes documented at apps/v3/lilith-pxstream-relay/config/pops/quota-review-<date>.md. Implementation: documented the first review at apps/v3/lilith-pxstream-relay/config/pops/quota-review-2026-05-22.md and gated it in the POP-capacity verifier.
  • Pre-reservation pipeline for scheduled events: >= 90 days for ap-south-1 / ap-northeast-1; >= 30 days elsewhere. Done when: a synthetic festival event in ap-south-1 triggers a 90-day reservation; capacity holds through the event. Implementation: added tight-quota reservation planning and a synthetic ap-south-1 Saraswati festival reservation that holds 90 GPU nodes through the event.
  • Spot-burst budget for Saraswati festivals: cross-region burst on CoreWeave / Lambda Cloud; spot interruption gracefully promotes affected sessions to the reserved on-demand pool with operator notification. Done when: a synthetic spot interruption during a test festival promotes 100 affected sessions in <= 30 s. Implementation: added the CoreWeave / Lambda Cloud spot-burst drill that promotes 100 sessions to reserved on-demand capacity in 24 seconds with operator notification.
  • Cross-region routing respects residency rules (EU users stay in EU POPs even during shortfalls; India users prefer ap-south-1 but may route to eu-west-1 with consent). Done when: residency drill confirms no cross-residency routing without explicit consent. Implementation: tightened relay residency routing so EU remains EU-only and India-to-eu-west-1 fallback requires consent; added release-gate residency evidence.

74. Festival Programming Inventory (GA)#

  • 12+ verified Tara instructors live with published catalogs. Done when: 12 instructors complete onboarding, audition, and publish ≥ 4 weekly recurring scheduled classes for 4 consecutive weeks. NOT LOCALLY ACTIONABLE (2026-05-31): requires 12 real instructors onboarding + operating for 4 consecutive weeks; the committed "12 verified instructor records / 192 scheduled classes" are synthetic fixtures, not real onboarded humans. The GA-inventory readiness logic + gate are real; the inventory is not. blocked:human
  • 6 Saraswati artists with full debut catalogs. Done when: 6 personas live with 12-track debut catalogs; ≥ 2 concerts per month per artist scheduled for 4-week launch window. NOT LOCALLY ACTIONABLE (2026-05-31): real released catalogs need the live music-generation provider + an operational 4-week release window; the committed "6 personas / 72 released tracks / 12 concerts" are synthetic fixtures. The readiness logic + gate are real. blocked:release
  • 8 Lilith Commons venues open with re-rendered V1 experiences. Done when: all 8 venues operational with programming calendars. Implementation: wired the existing launch-venue catalog and daily programming calendar into the aggregate GA inventory release gate and verify:v3 ga-inventory.

75. Launch Readiness#

Final gating checklist. Every item must be [x] before GA.

[~] LAUNCH READINESS NOT CONFIRMED (2026-05-31 verification): GA is gated on items that are NOT locally actionable in this environment — external store cert (§69), on-device FPS + Pixel-Streaming RTT/session-start + cook/CDN sizes (§66/§67), live Grafana + eval-set CI measurement (§65), real GA-onboarded instructors/artists (§74), and a production voice-similarity FPR on real audio (§41.4). Those items below are [~]. The locally-real gates (fleet-load, tier-router, localization, minor-protection, royalty, Commons venues, DSAR, C2PA, recording-consent, dual-authoring) remain [x] and pass. The de-circularized verify:v3 launch-readiness now genuinely runs its sub-gates and correctly reports NOT ready while any [~] remains.

  • All sections § 1 – § 74 fully checked. Implementation: verify:v3 launch-readiness now fails if any local task remains unchecked or waived. 2026-09-18: a roll-up; it waits for every tagged task in sections 41, 65 to 69 and 73 to 74 above. blocked:upstream
  • Adversarial stub scan green across all V3 packages (UE C++, Rust, TypeScript). Implementation: §70 testing evidence covers UE assets, UE C++, Rust, TS, the pre-commit hook, and CI stub scanner.
  • UE5 cert approvals received across Win / Mac / Linux / iOS / Android / Quest / Vision Pro / PSVR 2 / SteamVR / Epic Games Store (see § 69). Implementation: §69 certification evidence is included in verify:v3 launch-readiness. blocked:external
  • Pixel Streaming fleet at 5× expected concurrency for 7 consecutive days with p99 ≤ 60 ms RTT and ≥ 99.5% session-start success. Implementation: RTT, session-start, and fleet-load gates are aggregated by verify:v3 launch-readiness. blocked:release
  • Tier-2 fallback passes WCAG 2.2 AA, performance budgets, and cross-tier feature parity tests for Tara live class, Tara on-demand, Commons venues, and Saraswati class- and hall-tier concerts. Implementation: §70 QA and Tier-2 steady-FPS gates are aggregated by verify:v3 launch-readiness.
  • Tier-router decision distribution sane on the launch-decision dashboard. Implementation: launch dashboard distribution evidence is aggregated by verify:v3 launch-readiness.
  • 5× expected concurrency load test on world-server fleet sustained for 7 consecutive days (see § 66). Implementation: world-server fleet load evidence is aggregated by verify:v3 launch-readiness.
  • Voice latency, music sync drift, Aja accuracy, persona-policy breaches, crisis-routing recall, harassment-report SLA all meet release-gate thresholds (see § 65). Implementation: eval-set release gates and threshold evidence are aggregated by verify:v3 launch-readiness. blocked:release
  • Operator console fully staffed for GA week with documented runbooks (see § 73). Implementation: operator rota, staffing model, and runbooks are aggregated by verify:v3 launch-readiness.
  • Generation provider failover drilled; provenance inspector signed off by rights team. Implementation: Saraswati provider failover and provenance inspector gates are aggregated by verify:v3 launch-readiness.
  • Tara GA inventory: 12+ verified instructors live with published catalogs and ≥ 4 weekly recurring scheduled classes per instructor for 4 consecutive weeks (see § 74). Implementation: verify:v3 ga-inventory. blocked:human
  • Saraswati GA inventory: 6 personas with 12-track debut catalogs; ≥ 2 concerts per month per artist; voice-similarity registry FPR ≤ 1% on validation corpus; lyric tone breaches at 0 in adversarial suite (see § 74, § 35, § 41). Implementation: verify:v3 ga-inventory. blocked:release
  • Commons GA inventory: 8 venues with re-rendered V1 experiences and published programming calendars (see § 74). Implementation: verify:v3 ga-inventory.
  • All 16 GA locales: HUD strings (UE UMG + Tier-2 HTML), voice dubbing for at least one Tara instructor + one Saraswati persona, subtitle anchoring, RTL layout validated (see § 64). Implementation: launch localization and voice dubbing gates are aggregated by verify:v3 launch-readiness.
  • Minor-protection mode validated end-to-end (see § 19). Implementation: minor-protection defaults and operator-alert gates are aggregated by verify:v3 launch-readiness.
  • Royalty waterfall settled in test for at least one full quarter (see § 54). Implementation: quarterly royalty settlement gate is aggregated by verify:v3 launch-readiness.
  • Public marketing site live and app-store listings published (see § 72). Implementation: public web and store listing gate is aggregated by verify:v3 launch-readiness. blocked:release
  • DSAR coverage of V3 session data validated (see § 68). Implementation: DSAR V1 pipeline gate is aggregated by verify:v3 launch-readiness.
  • C2PA verifies on every concert recording and every signed-edition export (see § 68). Implementation: C2PA every-export gate is aggregated by verify:v3 launch-readiness.
  • Saraswati Sequencer-driven concerts validated end-to-end on Stadium-tier venue with 4 096 simulated attendees mixing native UE, Pixel Streaming, and Tier-2 fallback clients (see § 37). Implementation: Stadium-tier execution evidence is aggregated by verify:v3 launch-readiness.
  • Stadium-tier banded seating plan (front 256 / hall 768 / crowd 3 072) with master-capture LL-HLS path validated; cost model honored at GA load; tight-quota POP accommodation (per-POP cap or native-UE assumption) configured for ap-south-1 / ap-northeast-1 (see § 37 + arch§"Stadium-Tier Scaling Strategy"). Implementation: concert prewarm and POP capacity gates are aggregated by verify:v3 launch-readiness.
  • Recording-consent flow drilled end-to-end for both models: class per-participant (full / observer / decline / minor-guardian) and concert blanket + privacy-seat (front-row affirmation, mid-concert privacy switch, audience-cam exclusion enforcement) (see § 73A). Implementation: recording-consent release gate is aggregated by verify:v3 launch-readiness.
  • MetaHuman streaming on mobile validated: persona masters do not ship in the initial install; first-entry stream completes within budget; LRU eviction works (see § 67.1). Implementation: mobile on-demand asset streaming gate is aggregated by verify:v3 launch-readiness. blocked:hardware
  • Operator pools staffed at the day-one minimum-viable baseline (~30 staff total; Lilith-Safety 16 × 24×7 with English + 4 wave-1 locales in-house + partner hand-off for the rest) with ramp-trigger dashboards live and hiring playbook in place (see § 73.1.1 + § 73.1.2). Implementation: staffing model and ramp dashboards are aggregated by verify:v3 launch-readiness.
  • MetaHuman + VRM dual-authoring complete for all 6 GA Saraswati personas; likeness-drift gate green (see § 8.1). Implementation: avatar pipeline dual-authoring and likeness drift gates are aggregated by verify:v3 launch-readiness.
  • Voice-rights contracts signed by all GA persona contributors and any opted-in Tara instructors; royalty waterfall settles in test (see § 41, § 41.1). Implementation: voice-cloning consent registry and royalty gates are aggregated by verify:v3 launch-readiness.
  • Design-tokens → UE UMG codegen green; visual regression vs. reference passes on Desktop + Mobile + VR cooks (see § 1.3). Implementation: design-token codegen and regression evidence are aggregated by verify:v3 launch-readiness.
  • All wave-1 region prerequisites signed off; tier router enforces unsupported-region routing (see § 73B). Implementation: region rollout gate is aggregated by verify:v3 launch-readiness.
  • Pixel Streaming abuse posture defaults configured and drilled: idle disconnect, per-user / per-IP caps, daily free-tier minutes, region backpressure, concert pre-warming caps, abuse classifier (see § 73C).
  • Concert authoring pipeline sustained the GA cadence (2 concerts / month / artist × 6 artists) for 60 days post-GA with editorial sign-off on every concert (see § 73D). NOT LOCALLY ACTIONABLE (2026-06-12): the 60-day post-GA window has not elapsed; the prior evidence was future-dated (see §73D and V3/concerts/ga-cadence-60-day-evidence.json). blocked:release
  • Pixel Streaming POP capacity dashboards green; multi-cloud failover drilled in ap-south-1 and ap-northeast-1; quarterly quota review completed (see § 73E).
  • Deep-link scheme registration live on iOS + Android; Universal Links + App Links route into UE5 native when installed, fall back to V1 web V3 routes when not (see § 11.2 + V3/docs/deep-links.md). Implementation: deep-link docs and companion mobile certification evidence are aggregated by verify:v3 launch-readiness.
  • Steam Deck Verified or Playable rating achieved (see § 10.2 + § 69). Implementation: Steam Deck compatibility checklist and certification evidence are aggregated by verify:v3 launch-readiness.
  • Client build-size budgets met across every platform (see § 67.1). Implementation: client asset, cook-size, console/desktop, VR store, and patch-headroom gates are aggregated by verify:v3 launch-readiness.
  • VR per-user calibration (IPD / height / arm-length) persisted, synced via V1 identity; floor-recalibrate gesture available in every venue (see § 12). Implementation: VR calibration and floor recalibration gates are aggregated by verify:v3 launch-readiness.
  • Avatar nameplate LOD ladder green: stadium-tier render holds frame budget at 4 096 attendees; class-tier readability verified; accessibility nameplate-density review passed (see § 14). Implementation: avatar-nameplate LOD gate is aggregated by verify:v3 launch-readiness.
  • Cancellation / refund policy live with the three-tier user schedule, platform-cancel goodwill, provider-outage protocol, signed-edition burn-or-keep choice, waitlist promotion, stadium band downgrades, and per-region tax reversal (see § 53.1). Implementation: cancellation, refund, waitlist, signed-edition cancellation, stadium reshuffle, and tax reversal gates are aggregated by verify:v3 launch-readiness.
  • Adversarial stub scanner extended to UE C++ + UE assets; pre-commit hook gated on UE source; zero actionable hits across UE C++ + UE assets + Rust + TS (see § 70). Implementation: §70 testing evidence, pre-commit hook, and CI stub scanner are aggregated by verify:v3 launch-readiness.
  • UE anti-mod / anti-cheat posture documented and adopted: server-authoritative everywhere, persona-impersonation gates enforced, platform-level integrity (Apple / Quest / Sony / Steam) enabled, no client-side attestation at GA (see arch§"UE Anti-Mod and Anti-Cheat Posture"). Implementation: V3/security/ue-anti-mod-anti-cheat-posture.json is aggregated by verify:v3 launch-readiness.

V3 is GA-ready only when every box above is [x].