V3 Product Review — Lilith Metaverse (Tara Studio, Saraswati Stage, Lilith Commons)

Status: independent product review — revised and expanded in a second meticulous pass on 2026-07-07 (§9, Additional considerations) and a third pass on 2026-07-07 covering monorepo integration, shared functionality, and TODOS/ roadmap alignment (§10) Reviewer: Claude (product-perspective deep review) Date: 2026-07-06 (first pass) · 2026-07-07 (second and third passes) Scope: V3/V3_features.md + all V3/features/* pages, V3_TODOS.md, VERIFICATION_AUDIT_2026-05-31.md, REMEDIATION_2026-06-12.md, the root validation/ops corpus (Pixel Streaming, FPS, voice latency, concerts, launch readiness), and the concerts/creator-docs/regions/legal subtrees. Method: full deep-read, then critique against the shipped competitive set (Supernatural, FitXR, Alo Moves (+XR), Apple Fitness+, Peloton, TRIPP, Headspace XR, Down Dog; VRChat, Rec Room, Horizon Worlds; Fortnite/Roblox concerts, Wave, AmazeVR; Hatsune Miku / K/DA / virtual-artist lineage) and industry SOTA. Market facts reflect knowledge through early 2026.

1. Executive summary#

V3 turns Oshun V1 into an embodied, multi-user surface on UE 5.5: Tara Studio (live and on-demand yoga/breathwork/meditation with verified human instructors, on-device pose coaching via Aja, and supervised AI personas), Saraswati Stage (six AI virtual artists with generated catalogs and Sequencer-driven live concerts up to 4,096 attendees), and Lilith Commons (eight contemplative social venues re-rendering V1 domains), reachable natively on desktop/mobile/VR/console, via Pixel Streaming in a browser, and through a three.js Tier-2 fallback.

What is genuinely excellent: the consent-and-rights regime (voice-clone contributor pools with withdrawal rights, recording consent with retroactive redaction, privacy seats enforced by the world server, C2PA on every export, public-figure voice prohibition) is years ahead of industry practice in AI music and social VR; the trauma-informed class-safety design (invitational language, hands-off defaults, solitary cells, "step back") is best-in-class; and the instructor verification/economics (80–90% revenue share, lineage attestation, background checks gated to minors' classes) form a credible supply-side proposition.

What the review must weigh heavily: the project's own adversarial audit found the validation corpus substantially synthetic — every performance "measurement" (FPS, RTT, session-start success, voice latency, POP failover) is a target typed as an observation, GA supply (12 instructors, 6 artists) is factory-generated fiction, and the launch gate is honestly red post- remediation. V3 is a well-engineered core (real physics, real HRTF audio, real C2PA signing, 100%-recall crisis classifier) with zero real-world operational proof and a per-session GPU economics problem it visibly knows about.

The product-strategy question the docs never quite ask: is embodiment the right primary surface for yoga? The mass market practices on a mat in front of a phone or TV; the genuinely novel assets here (Aja's on-device pose coaching, "practice with the recording," humane class culture) do not require a metaverse — while the metaverse's costs (GPU workers, moderation, VR physical-safety risk) are all incurred up front. The strongest version of V3 leads with the camera-and-companion practice loop everywhere, uses embodiment as the community and event layer, and treats VR as an option with a passthrough-first safety posture — which currently isn't specified at all.

Verdict: industry-leading trust architecture and class design, wrapped around an unvalidated and expensive embodiment bet; re-center the practice loop, prove one real cohort and one real concert, and fix the physical-safety and economics blind spots before GA.

2. What V3 is (product identity)#

• One UE5.5 project driving Tier-1 native surfaces (Windows/macOS/Linux, Steam Deck, iOS/Android, Quest 3/Pro, Vision Pro, PSVR2, SteamVR), web via Pixel Streaming (H.264/AV1 WebRTC), and a three.js/WebGPU Tier-2 fallback; a tier router picks per-device/bandwidth/capacity.
• World model Realm→District→Venue→Room→Cell; capacity tiers Intimate 12 / Class 64 / Hall 256 / Stadium 4,096 (banded: native front-256, Pixel Streaming mezzanine 768, LL-HLS crowd 3,072).
• One account (the V1 account); V1 trust substrate (crisis behavior, persona policy, provenance, residency, DSAR) governs every embodied action.
• Commerce: tickets (anti-scalp, face-value resale only), tips, layered subscriptions, on-chain signed editions with royalty waterfalls (Aje), and fiat via Stripe/IAP.

3. What is genuinely strong#

1. The consent architecture is the best I've seen in this category. Named voice contributors with a 12% pro-rata pool, 90-day withdrawal that destroys trained weights, anti-impersonation similarity checks against a public-figure corpus, no web-scraped corpora, recording consent with observer-mode silhouettes and retroactive redaction, minors auto-assigned privacy seats, C2PA manifests on every export. After the FN Meka and "Heart on My Sleeve" fiascos and the label-vs-Suno/Udio litigation era, this is precisely the regime the industry is being dragged toward; V3 arrives with it built. This is the marketing story for Saraswati — not "AI artists," but "the first rights-clean virtual artists."
2. Trauma-informed embodiment. Invitational language enforced by persona policy rather than instructor discretion; hands-off and eyes-open defaults; student-reversible pose overlays; solitary cells in every venue; one-tap "step back" (spatial mute + distance) — social VR's harassment problem addressed structurally, ahead of VRChat/Horizon norms.
3. Aja's privacy posture: camera frames never leave the device; only derived joint angles cross the wire. This matches the strictest smart-gym-hardware standard (Peloton Guide) while running on the phone the user already owns.
4. "Practice with the recording" — joining a shared room with a recorded instructor's avatar while Aja scores your live pose against theirs at each asana marker — is a genuinely novel format ("ghost class") no competitor ships; it converts a video library into an embodied practice partner.
5. Instructor supply-side economics: 80/90% splits, 10% privates fee, donation classes with waived fees, an optional Lineage Fund — dramatically more attractive than Alo Moves' salaried content model or YouTube's ad economics, with verification depth (credentials, lineage, insurance, auditions) that platforms like Insight Timer lack.
6. Honest self-audit culture. The verification audit calling out its own fabricated metrics, the de-circularized launch gate now honestly red, and an on-box UE build catching a real compile bug — as with V1/V2, the process integrity is a product asset (when it's allowed to gate claims — see §6.1).

4. Product gaps#

4.1 The surface-market mismatch (strategic)#

Yoga's mass market is mat + phone/TV: Alo Moves, Peloton, Apple Fitness+, Down Dog, and free YouTube own the practice loop with zero-friction video. VR fitness (Supernatural, FitXR) succeeds in cardio/rhythm formats where occlusion and gamification help; yoga is the worst VR fitness fit — floor work, inversions, balance poses, and eyes-closed rest are all hostile to a headset. V3's most defensible assets (Aja coaching, humane class culture, "practice with the recording," instructor economics) are surface-independent, yet the product is framed metaverse-first.

Recommendation: invert the framing. The flagship Tara experience is the phone/TV/web class with Aja (the tier router already reaches these surfaces); the embodied rooms are the community layer (cohorts, circles, events) and the event layer (Saraswati, Observatory nights). This costs nothing technically — it is a positioning and onboarding change — and it puts V3 in the market where the users are, instead of asking yoga practitioners to adopt VR before they can discover what's special here.

4.2 VR physical safety is unaddressed (blocking gap for the VR surface)#

Nothing in the corpus covers physical safety of practicing yoga in an HMD: balance poses with occluded vision, floor transitions, prop use, sweat/hygiene guidance, or guardian-boundary sizing for flow sequences. Industry practice (Supernatural's play-space guidance, Quest's boundary system) covers boxing, not tree pose. Passthrough/MR should be the default for standing and balance sequences on Quest 3 / Vision Pro — both platforms are passthrough-first devices now, and MR is also where headset yoga is actually plausible (you see your mat and your walls; the instructor avatar joins your room). The docs never mention passthrough at all. This is both a safety obligation and the strongest VR product idea available to V3.

4.3 Supply-side cold start (both marketplaces)#

GA promises ≥12 verified instructors, ≥6 artists with catalogs, ≥8 venues — and the audit found the GA inventory is factory-generated fiction; zero real instructors are onboarded and zero editorial-signed concerts exist. The verification pipeline (5-day SLA, 4 operators) and the 2–3-week concert authoring pipeline are plausible machinery, but there is no recruitment strategy, no anchor-teacher partnerships, no launch-artist creative slate in evidence. A marketplace with world-class rails and no trains. The single most important pre-GA milestone is boring: run one real 6-week cohort with a real teacher and real students, and stage one real concert end-to-end, and let those artifacts calibrate everything (SLAs, tooling, economics, retention).

4.4 Saraswati's demand question — and the missing co-creation loop#

The proven demand for virtual artists (Hatsune Miku, K/DA, Gorillaz) runs on fan co-creation: Miku is a fandom because fans write the songs. V3's Saraswati is editorial-supplied — six personas, staff-authored catalogs, pre-generated banter, no live Q&A at GA — with fan participation limited to emoji, tips, and a request-from-list. Remix rights exist per-track but are buried as a licensing flag. Without a co-creation loop, AI artists are a novelty concert series competing against Fortnite's licensed megastars; the Velvet Sundown backlash (2025) shows audiences punish AI music that arrives at them rather than from them.

Recommendation: make co-creation the Saraswati thesis. Fan lyric/stem submissions through the same Isis/rights pipeline with revenue share; fan-voted setlists and arrangement variants; remix competitions whose winners enter the canonical catalog with on-chain credit; club-pass members voting persona lore forward. The rights regime — V3's crown jewel — is exactly what makes safe fan co-creation possible at scale, and no competitor can copy that quickly.

4.5 Economics of the streamed metaverse#

One UE worker per browser session is the Stadia cost model, and the abuse posture (120 free minutes/day, per-IP caps, 85%-utilization shunting to Tier-2) shows the team knows it. Concerts are the worst case: mezzanine (768 Pixel-Streaming sessions per instance) is a per-viewer GPU bill for what is functionally a shared cinematic. Pre-render the mezzanine: the master-capture worker already records every concert full-fidelity; a low-latency composited broadcast (per-band camera mixes over LL-HLS with the existing 200ms-batched emoji overlay) would collapse mezzanine cost to CDN economics while reserving Pixel Streaming for interactive non-concert use. Cross-tier sync is already only ambient (≤300–350ms bands), so nothing real is lost. Relatedly: market the crowd band honestly as "watch together," not as being "in" a 4,096-person room (§6.4).

4.6 Missing table stakes for the fitness/wellness category#

• Biometrics: no heart-rate/HRV integration (watchOS/Wear OS companions exist in the monorepo; Supernatural/Apple Fitness+ set the expectation). Breathwork is the deepest fit — HRV-guided paced breathing with the existing cadence taxonomies would be a differentiator, not just parity.
• Progress and programs: Arete continuity writes practice logs, but there's no visible "program" spine (30-day foundations, prenatal series beyond cohorts) or milestone system a Peloton/Alo user expects.
• Offline/downloadable classes for the on-demand library (mobile surfaces; travel/retreat use is core to this audience).
• Web accessibility: the audit found no axe-core anywhere in V3; for a wellness product claiming 18 locales and inclusive design, WCAG coverage on the web player/studio is a launch gate, not a nice-to-have.
• Chair/adaptive practice as a first-class category: the 4-step modification ladder and contraindication model are built for this; surfacing "adaptive-first" classes would serve an underserved market and match the product's stated values.

4.7 Industry SOTA checklist#

5. Ideas that would make the product better#

1. Passthrough-first "Mat Mode" (§4.2): MR default on Quest 3/Vision Pro — instructor avatar on your floor, Aja cues anchored to your mat, full VR reserved for savasana/meditation environments. Safety fix and killer demo in one.
2. Fan co-creation for Saraswati (§4.4) — the Miku playbook on rights-clean rails.
3. Hybrid IRL-streamed classes: let a verified instructor teach their physical studio class while streaming into a V3 room (avatar driven by phone camera capture). Solves supply cold-start with classes instructors are already teaching, and gives studios a global second room with an 80/20 split better than ClassPass economics.
4. HRV-guided breathwork (§4.6) with the watch companions; "coherence score" trends into Arete. The breath-cadence taxonomy (box, 4-7-8, coherent) is already modeled.
5. Nyx event nights as the Commons growth engine: the Observatory is Stadium-capable and fed by live Nyx data — scheduled eclipse/meteor watch-parties are free, repeatable, shareable events that introduce the platform without a subscription. Market them as the "free concert" of the Commons.
6. Corporate/studio B2B tier: Tara studio passes sold to employers and physical studios (the tenant machinery and cohort format already fit); wellness B2B is where meditation apps found their margin (Calm Business, Headspace for Work).
7. Persona meet-and-greets as a club perk: 1:1 or small-room Saraswati persona conversations under the strict persona-policy + adversarial-suite regime (green-room infra exists). This is the Inworld-style feature fans expect from "AI artists," delivered with V3's safety differentiator — pre-approved topical envelope, visible AI labeling, operator handoff.
8. "Ghost class" as a marquee brand (§3.4): name it, market it, and add asynchronous cohort ghosts — practice "alongside" your cohort's recorded sessions during the week between live meetings.
9. Retreat mode: multi-day programs combining scheduled live sessions, solitary-cell reservations, Lantern Hall evening rituals, and device downtime nudges (Lilith attention policy) — premium-priced, deeply on-brand, and impossible for video-only competitors to copy.
10. Instructor analytics that respect the tone: drop-off points and asana-coverage exist; give teachers a humane dashboard ("students found the second vinyasa fast" rather than churn-red charts), which becomes a recruitment selling point.

6. Criticisms and tweaks#

1. Purge the synthetic "observations." Validation docs still print fabricated precision ("Observed sustained 71.8 fps," "99.7966% success") that the audit proved was typed, not measured. Every such number should be relabeled "budget/target" until a real run exists — the honesty culture that governs the TODOS must govern the validation corpus inline, or an external reviewer/partner will (rightly) treat all claims as suspect.
2. Subscription stack is fatigue-inducing: Tara studio passes + per-artist Club passes + multi-artist passes + venue memberships + signed editions + tips. Consolidate to one Oshun+ membership (all-tenant base) with à-la-carte artist clubs on top; keep everything else transactional.
3. Concert cadence promise (≥12/month across 6 artists) is over-committed for an editorial pipeline that needs 2–3 weeks per concert with 2 dress rehearsals and 4 Saraswati operators; either grow the bench, build a reusable-show variant system, or halve the promise before it becomes a missed-schedule story.
4. "4,096 together" framing oversells the crowd band (LL-HLS spectators with 1Hz aggregated emoji). Market bands as named experiences (Front Row / Mezzanine / Watch Party) with prices and expectations matched — the anti-scalp face-value resale is a genuinely consumer-friendly story to pair with it.
5. AI-persona instructor cap (1 per 4 humans/region) is good policy — also publish it. It's a trust differentiator hiding in an internal doc.
6. Moderation staffing (16 across 3 shifts, contracted 4h-SLA for non-primary locales) vs. a 24×7, 15-region, 18-locale embodied platform: either narrow the launch surface (fewer regions/venues live-moderated at GA) or staff up; harassment SLAs (30-min programmed) are promises with payroll implications.
7. Cultural-adaptation ledger is a pending skeleton — for a product monetizing lineage-attributed practice (Iyengar disclosures, devotional genre cells, Vedic passages), the named-reviewer sign-offs must exist before those items ship; the verifier currently enforces only structure.
8. Quest 2 deferral is right; Vision Pro native at GA is questionable — the smallest possible audience with the highest per-platform cost. Ship Vision Pro when Mat Mode (§5.1) exists to justify it.
9. Voice-similarity FPR is synthetic-corpus only (§41.4 [~]): the anti-impersonation gate is a headline claim; buy/collect a real evaluation corpus before marketing it.
10. Tier-2 Aja cues as HUD text — the fallback tier delivers coaching as flat overlay; fine, but listings should disclose per-tier coaching fidelity so a Tier-2 user doesn't review-bomb a "feedback" promise their tier can't render.

7. Risks#

1. Economics — per-session GPU streaming + lean subscriptions + niche VR hardware; the cost curve is Stadia-shaped unless mezzanine pre-rendering (§4.5) and Tier-2-first defaults hold the line.
2. Two-sided cold start (§4.3) — rails without trains; fictional GA inventory is the audit's most product-lethal finding.
3. AI-music reception — post-Velvet-Sundown, "AI band" is a hostile headline by default; the rights-clean + co-creation framing (§4.4) is the only durable counter-position. Never let a Saraswati track ship without its provenance story attached.
4. Physical injury in VR yoga (§4.2) — one viral injury story defines the product; passthrough defaults and physical-safety UX are cheap insurance.
5. Moderation incident in a contemplative space — the brand promise raises the damage of any failure; staffing (§6.6) is the exposure.
6. V1 coupling — V3 cannot ship independent of V1 readiness (identity, DSAR, crisis routing), and V1's own review flags P0 residuals; sequence the launches deliberately.
7. Music-rights fragmentation — Saraswati's narrower regional launch is correctly modeled; keep generated-catalog PRS registration and per-region licensing ahead of any user-generated remix expansion.

8. Prioritized recommendations#

P0 — before any GA date

1. Run one real cohort + one real concert end-to-end; onboard ≥12 real instructors and retire the fictional inventory (§4.3).
2. Passthrough Mat Mode + VR physical-safety posture (§4.2, §5.1).
3. Relabel all synthetic metrics; gate marketing numbers on real device/load runs (§6.1).
4. Mezzanine pre-rendering decision to fix concert economics (§4.5).
5. axe-core/WCAG coverage on web player + Studio (§4.6).
6. Subscription consolidation into Oshun+ (§6.2).

P1 — the differentiation wave 7. Fan co-creation pipeline for Saraswati (§4.4) + persona meet-and-greets (§5.7). 8. HRV/heart-rate integration via watch companions (§5.4). 9. Hybrid IRL-streamed classes (§5.3). 10. Nyx Observatory event nights as free acquisition (§5.5). 11. Publish the trust regime (AI-persona caps, consent architecture, anti-scalp resale) as public policy pages (§6.5).

P2 — compounding 12. Retreat mode (§5.9); adaptive-first class category (§4.6). 13. Corporate/studio B2B tier (§5.6). 14. Ghost-class branding + async cohort ghosts (§5.8). 15. Offline downloads for on-demand library (§4.6). 16. Real voice-similarity evaluation corpus (§6.9).

9. Additional considerations (second pass)#

9.1 Store policy vs. the commerce stack#

Meta's Quest store and Apple's App Store restrict or prohibit in-app crypto purchases and alternative payment rails for digital goods — the signed-edition/ticket/Aje stack must route web-side on those platforms (same IAP posture recommended for V1), and the PSVR2 path runs through PSN Wallet only (already modeled). Decide the per-store checkout matrix before certification, not during; it also affects how Club passes can be merchandised inside headset builds.

9.2 Live-class economics across time zones#

"Live-service from day one" with ≥12 instructors means the live schedule is really three regional schedules (Americas/EMEA/APAC prime times), or the "live" promise collapses to one region's evening. Peloton's lesson applies: on-demand-first with scheduled regional live windows, and the ghost-class format (§3.4) as the bridge that makes off-hours feel live. Instructor recruitment should be time-zone-distributed from the first cohort, not retrofitted.

9.3 Music licensing in classes is a litigation surface#

"Royalty-cleared music stream" hides the industry's most expensive recurring lesson: Peloton's music-licensing suit exposed it to nine-figure claims for uncleared class music. V3 needs a cleared-catalog program (instructors pick from pre-licensed/original tracks; no uploads), with per-territory public-performance coverage — and Saraswati's generated catalog is actually a structural advantage here: the platform can supply unlimited rights-clean class music from its own artists. Make that the default and market it to instructors as "never think about music rights."

9.4 Body inclusivity as visible product values#

A yoga platform's avatars make a statement whether designed or not. Commission an inclusive VRM body set (plus-size, pregnant, older, limb- different, seated-practice representation) and let modification-ladder depth pair with adaptive avatars — this is both on-values and a recruitment signal to the adaptive-yoga instructor community (§4.6's adaptive-first category needs teachers who see themselves representable).

9.5 Health-claims and injury-liability discipline#

Prenatal, recovery, and grief-adjacent classes carry elevated duty-of-care expectations. The credential gates exist; add per-listing contraindication surfacing (the taxonomy already models contraindications), region-aware disclaimers, and an incident-reporting path for practice injuries distinct from conduct reports. Also negotiate a platform group liability policy for instructors — the >25-seat insurance requirement becomes a recruiting lever if the platform brokers the coverage.

9.6 Vestibular comfort, not just photosensitivity#

The PSE/strobe governance is strong; VR comfort needs the vestibular dimension too — comfort ratings per experience (Commons locomotion, concert camera moves), teleport/vignette defaults, and seated-mode parity for every Commons venue. Rate every experience the way Quest store comfort labels do, before Meta's cert asks.

9.7 Creator payout operations#

Stripe Connect + Aje payouts to instructors and voice contributors across 15 regions triggers DAC7/1099-K reporting, KYC, and withholding obligations at GA scale. The commerce contracts model the splits; the tax and compliance ops need a named owner and a launch-gate checklist — V7's papers cover this ground for realms and should be inherited here.

9.8 Cross-portfolio notes#

V1's Tara library is the on-demand seed content for Tier-2 surfaces (reuse, don't re-record); V6 Egbe agents could staff Commons ambience (greeters, observatory docents) through the Aye Bridge — the same cold-start remedy recommended for V7 realms; and the Observatory's live-Nyx feed is a shared asset with V9 (a V9 astronomy lesson should deep-link to tonight's Commons sky event, and vice versa). One caution mirrors V1's brand note: Saraswati's pop-concert energy and Tara's trauma-informed intimacy share one moderation pool and one brand — keep tenant tone policies (already separate) visible in marketing too, so a concert clip never sets expectations for a grief circle.

10. Monorepo integration & shared functionality (third pass)#

Audited against DOMAINS/, the 185-phase TODOS/ roadmap, V_SERIES.md, V_SERIES_PLATFORM_CONSOLIDATION.md, and package-dependency evidence.

10.1 V3 is the best-integrated V-app — say so, and finish the job#

Dependency evidence shows V3 with real edges into the shared substrate that its siblings mostly lack: it is the only V-product reusing the canonical @oshun/memory-iris and @oshun/persona-policy-lilith adapters (via memory-iris-spatial, lilith-body-policy, isis-music), plus @oshun/content-signing, the shared release-gates/quality-judge, and the Tara live-class booking package. The embodiment-without-compromising- V1-trust promise has code behind it. The unfinished part: V3's Pixel Streaming player, Tier-2 web fallback, and tier router are not extracted as shared packages — and V6 has already re-implemented its own (egbe-web-pxstream, egbe-engine-web-fallback) rather than reusing V3's, with V7's web tier next in line. Extract the metaverse delivery substrate now (pxstream player, fallback engine, tier router, POP/abuse posture) into neutral shared libs before a third copy appears; V3 is the donor because V3's is the most complete.

10.2 Tara Studio sits on three unclaimed fitness assets#

The roadmap already built most of Tara Studio's hardest layer under other names: Phase 15 "Nike" (732/732 tasks — a fitness-instruction engine with 65-joint skeletal/IK work that physically lives in libs/aja/fitness-animation because Nike is a ghost domain with no package of its own), Phase 16 Aja (pose estimation, retargeting, video-to-motion — V3's stated pose layer), and Phase 31 Shakti (1760/1760 — yoga/martial-arts form analysis and biomechanics, the same library whose fighting bridge feeds V2's rulesets). V3's Aja coaching should be explicitly specified as a consumer of Shakti biomechanics + the Phase-15 instruction engine, which both de-risks the "cue clarity" acceptance gate and creates the portfolio's best cross-product story: the same movement-science substrate teaches your yoga and referees a fighting game.

10.3 Saraswati Stage is a productization of three completed domains#

The concert experience is specified as new V3 build, but the roadmap has already marked complete: Calliope (Phase 66) — "autonomous AI musical personas: voice, visual identity, catalog, fanbase," which is Saraswati Stage's exact persona spec; Euterpe (Phase 38 + Phases 132/133/148) — composition→mastering→royalty lifecycle, sovereign audio runtime, DAW, and DJ workstation; and Uzume (Phase 62) — live performance and stagecraft. The Saraswati docs cite Isis/Yemaya but not these three. Re-basing the Stage on Calliope personas + Euterpe audio + Uzume stagecraft converts a greenfield build into productization — and Themis Music Shield (Phase 74) supplies the originality/IP screening the voice-clone regime needs at track intake. (Also note the naming collision flagged portfolio-wide: the Saraswati domain in DOMAINS/ is EV/battery/hardware technology, not music — V3's venue name will misroute newcomers and search; consider renaming the tenant or documenting the collision prominently.)

Resolved for cross-product integration, 2026-07-17: the V10 channel and shared technical identity are V3 Stage / v3.stage. Historical @oshun/tenant-saraswati-stage, Prisma, contract, and Unreal identifiers remain in place; @saraswati/* continues to mean the separate industrial-technology domain. The V3 Stage capability descriptor and architecture/feature docs now cite @calliope/stage as a specification source without falsely claiming a runtime dependency. The detailed source and render-gap audit is V10/RA8_V3_STAGE_CHANNEL_INTEGRATION.md.

10.4 The GPS tier and the concert economics#

V3's per-session GPU cost problem (§4.5) intersects Phase 140 (sovereign cloud-gaming platform) — a planned GeForce-NOW-class streaming stack. If Phase 140 proceeds, V3's Pixel Streaming POP fleet is its first tenant and the mezzanine pre-rendering recommendation becomes a shared-platform feature; if it doesn't, V3's POP/abuse machinery is the de-facto portfolio streaming layer and should be named as such in V_SERIES.md. Either way, one streaming fleet — V3, V6, and V7 each running their own PxStream relays (three SFU/voice stacks already exist: V3, V6's egbe-pxstream-relay, V7's mawu-gateway) is the duplication census's voice/streaming finding in miniature.

10.5 Roadmap assets V3 under-uses#

• Psyche avatar stack (Phase 24, ~759/761) — libs/psyche/avatar-* (core, expressions, lipsync, TAA, quality, Isis integration) is the ready embodiment layer for instructor and Saraswati avatars; V3's avatar pipeline should declare which pieces it consumes vs. rebuilds.
• Aglaea (Phase 34, fashion/beauty) — wardrobe/cosmetics content pipeline for the avatar wardrobe economy.
• Concordia (Phase 179) — royalty and dispute bargaining for the instructor/voice-contributor economy (currently Themis-only).
• Data flywheel (Phases 85–96) — class-movement data is intimate; V3 must inherit the V1 training-consent gate (V1 review §10.6) before any Aja-derived data reaches the flywheel.

11. Closing note#

V3's engineering core is real, its safety and rights architecture is the best in its class, and its self-audit culture caught its own fabrications — all rare. But the product as framed bets its cost structure on embodiment that yoga's market hasn't asked for, its supply on inventory that doesn't exist, and its numbers on measurements that were never taken. The path through is already latent in the design: the practice loop (Aja + humane classes) on the surfaces people own, embodiment as community and spectacle, co-creation on rights-clean rails, and passthrough before poses. Prove one cohort, one concert, one real number — then scale the cathedral.