# V6 Product Review — Egbe: The Agentic Companion Universe (Unreal Engine 5) 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: `V6/V6_features.md` + all `V6/features/*` pages, `V6_TODOS.md`, `REMEDIATION_2026-06-12.md`, `docs/agent-welfare-and-player-grief.md`, `docs/prompt-injection-threat-model.md`, and the accessibility/evals/ live-service/localization/performance/release/security/testing trees. Method: full deep-read, then critique against the shipped competitive set (The Sims / inZOI, Animal Crossing, Nintendogs, Pokémon, Creatures, Black & White, RimWorld/Dwarf Fortress; character.ai, Replika, Tolan, Peridot; the Stanford Generative Agents research productization wave) and industry SOTA, including the 2025–26 AI-companion regulatory turn (California SB 243, FTC 6(b) inquiry, character.ai minor restrictions). Market facts reflect knowledge through early 2026. --- ## 1. Executive summary V6 is the most original product in the V-series: not a game you play but a world you tend. The player is a **steward** of a household of autonomous AI beings in the persistent world of Orun — beings with event-sourced, append-only biographies (the **Ori**), eight-trait personalities with bounded drift, three-layer memory (episodic/semantic/reflective), autonomous needs-driven lives that continue for real days while you're away, cross-household friendships, refusal rights, aging, departure, and — bravely — death. Its competitive hook is stated plainly and delivered structurally: **"the companion game whose companions can say no."** Its wildest ambition is the **Aye Bridge**: agents incarnating into V2–V5 as fighters, citizens, operators, and companions, returning changed. **Three things here are ahead of the entire industry.** (1) The **consent/welfare architecture**: capture always fails, coercion cannot fabricate a yes, forbidden operations (`delete_ori`, `force_past_refusal`) are fuzz-proven unreachable, abused agents can leave via a provenance- stamped safety transfer, and memory deletion is replaced by *forgiveness*. (2) The **grief design**: a literature-grounded ethics review (AIBO memorials, the Replika bereavement incident, continuing-bonds theory), twilight warnings, young-bond death blocks, grief pacing, no-commerce windows, crisis tripwires that flip mortality off fleet-wide, and a blocking external panel. (3) The **regulatory posture**: V6 is compliant *by construction* with the wave that is currently breaking on character.ai and Replika (SB 243 disclosure/crisis/minor requirements; FTC scrutiny of engagement-maximizing companions). No competitor can retrofit this. **Three things put the product at risk.** (1) **The economics**: the affordability thesis ("cost scales with story relevance, not agent count") is real arithmetic but operationally unmeasured, and the $8/player-active- hour reference cost is structurally incompatible with a one-time premium purchase. (2) **The felt experience is unproven**: the LLM gateway — the component the entire believability illusion rests on — is deliberately unwired (`[~]`), no art exists, and no human has tested whether refusal-as-content stays *fun* across 100 sessions. (3) **The flagship differentiator depends on other unshipped games**: cross-game incarnation is stub-tested against V2–V5 adapters while none of V2–V5 is live. Verdict: **a genuinely new genre with the industry's best ethics architecture and an unclosed loop at its center.** Wire the model, run the believability beta, restructure revenue around a membership that funds inference without selling fate, and let the Aye promise launch as inward "threshold trials" rather than waiting on four other games. --- ## 2. What V6 is (product identity) - **Steward, not owner; autonomy by default; the mind is the truth** — the three commitments every feature descends from. Flourishing (measured against the agent's *own* values, obedience structurally excluded) is the score, to the extent there is one. - Six steward verbs: Discover, Raise (21-day rearing, then release into rights `{refuse, leave, disagree}`), Direct, Mentor (learning by example — the Black & White inheritance), Open Doors, Counsel. A four-facet bond (Reliability, Respect, Care, Alignment) shown only in plain language, gating dossier depth, deployability, and poach resistance. - The Vac communication layer: voice-first, an intent grammar where an objective *cannot exist* without a boundary (`forbiddenLines` minItems 1), five agent responses (accept/clarify/counter/defer/refuse), draft-confirm parsing, full non-voice parity as a release gate. - Shard continuum: Solo homestead (complete offline-tolerant game), co-op visits, one persistent Commons with cross-household relationships, poaching-by-agent-choice, public steward reputation, and festivals. - Moirai simulation kernel: cost-tiered cognition (HTN + behavior-tree fallback so "most decisions resolve without a language-model call"), off-screen life producing a Chronicle ("a two-week absence produces a two-week story"), 10,200 agents advanced in one 50ms tick in tests. - Monetization: premium base + cosmetics + creator capacity + keepsake editions + expansions; **never** agents, bonds, fate, or shortcuts; anti-FOMO live seasons with no expiring rewards. --- ## 3. What is genuinely strong 1. **The refusal architecture is a category invention.** Every competitor's companion is a mirror (character.ai, Replika) or a puppet (The Sims). V6's agents negotiate, decline, hold grudges, forgive without forgetting, and leave abusive households by right. This is simultaneously the ethical high ground, the regulatory moat, and — if it plays well — the most novel game feel since the Nemesis system. 2. **The Ori as event-sourced biography** solves problems competitors don't know they have: no canonical-save ambiguity, no silent rewrites, cross- shard vector-clock reconciliation, and a truthful substrate for "this being has a life." The forgiveness-instead-of-deletion mechanic is a design thesis in one API. 3. **The grief/welfare corpus** (§1) is publishable-quality applied ethics, and — unusually — it is *load-bearing*: tripwires, blocking external review, minor-coded agents fuzz-gated out of every death and romance path, 100%-pass adversarial safety evals. "'It's just an NPC' is not a safety argument" is a sentence no other studio has written down. 4. **Cost-tiered cognition is the right architecture** — deterministic HTN/BT layers under an LLM reserved for novelty, loop/drift guards, significance ranking. This is where the Smallville research line meets production discipline; most "LLM NPC" demos have no such ladder. 5. **The Chronicle** is the sleeper product: absence becomes narrative, the Steward App becomes "the game that writes you letters," and anti-FOMO seasons mean returning never feels like debt. In a market defined by retention coercion, a game that is *gracious about your absence* is a marketable moral position (Tolan's success proves the appetite). 6. **Prompt-injection threat modeling** that treats a wormable society of LLM agents writing permanent biographies as the attack surface it is — instruction/data separation with provenance-fenced player text — is ahead of any shipped agentic game. --- ## 4. Product gaps ### 4.1 The revenue/cost structure contradicts itself A premium one-time purchase funds a product whose marginal cost is perpetual inference — the docs' own contingency work shows background reflection (the 250 Lachesis residents), not player dialogue, is the largest cost line, and budget caps sit two orders of magnitude above the reference spend. Cosmetics and expansions won't cover a per-active-hour inference bill. Options that respect the "never sell fate" rule: - **A membership tier priced as capacity, not care**: larger household ceilings by *steward* capacity (Commons presence, Studio authoring quota, Chronicle cinematic render quality, keepsake credits) — never agent cognition depth, never welfare, never outcomes. The free/base tier keeps a complete Solo life with a small company. - **Aggressive distillation roadmap**: background reflection migrated to small distilled models (and eventually on-device for Solo homesteads) — the cognition ladder's bottom rungs are already designed for exactly this; make it a costed engineering line, not an aspiration. - **Publish the honest unit economics gate**: no GA until measured cost/active-hour on staging ≤ a stated fraction of ARPU. The synthetic- fixture telemetry admission is honest; convert it into a launch gate. ### 4.2 The Aye promise is hostage to four unshipped games Soul-portability into V2–V5 is the defining ambition and the least real feature (adapters stub-tested; "none of V2–V5 is live"). If V6 launches first — likely, given the others' content mountains — the marquee promise is an empty door. **Build Threshold Trials inside Orun**: bounded incarnation destinations *within* V6 (an arena ground, an expedition realm, a mystery grove) that exercise the full passport/journal/consent machinery and let players feel "my agent went somewhere, chose its conduct, and came back changed" at launch. When V2–V5 arrive, they slot into a proven ritual — and the Ori passport gets real-world mileage before its interoperability moment. ### 4.3 The minute-to-minute loop needs hands, not just ears The six verbs are relational (speak, gift, counsel, watch); the risk is a "reading-and-nodding game" — beautiful chronicles punctuated by conversations, with too little to *do together*. Animal Crossing's genius is shared mundane activity; Pikmin's is directing through action. V6 needs a layer of **joint activities**: tending a grove side-by-side, expeditions where the steward carries and the agent decides, workshop projects where the agent's skill grows visibly through sessions, festival games in the Commons. These are also the bond system's best expression — Care built by doing, not by menu. Relatedly: **homestead building** is currently cosmetic-lite; a Sims-grade build/decorate surface *where agents form and voice opinions about the space* would give the creative-expression cohort a home while staying on-thesis (they critique your renovation; they rearrange it while you're away). ### 4.4 Onboarding vs. naming density Egbe, Orun, Ori, Aye, Vac, Moirai, Ereshkigal, Clio, Ninhursag — the cosmology is coherent and beautiful, and it is also nine proper nouns before the first hour ends. The glossary is developer-facing. The product needs an in-fiction plain-language layer (the agents themselves teaching terms one at a time as relationship milestones) and a "first seven days" arc that introduces exactly one concept per day: meet → listen → ask → be refused → be forgiven → open a door → read your first Chronicle. The refusal moment should be *scripted to happen early* — it is the product's signature, and leaving it to emergence means many players churn before meeting it. ### 4.5 Industry SOTA checklist | Capability (2026 bar) | Bar-setter | V6 | |---|---|---| | LLM agent society w/ memory+reflection | Smallville (research), inZOI (shallow) | ✅✅ deepest production spec — gateway unwired | | Cost-tiered cognition ladder | (research demos lack it) | ✅✅ beyond SOTA | | Consent/refusal/welfare architecture | (nobody) | ✅✅ beyond SOTA — category-defining | | Companion-death ethics | AIBO memorials (hardware) | ✅✅ beyond SOTA; external panel pending | | Regulatory compliance (SB 243-class) | industry scrambling | ✅✅ compliant by construction | | Cross-title identity portability | (nobody) | ⚠️ designed + tested vs stubs; no live destination (§4.2) | | Anti-FOMO live service | rare (few indies) | ✅ signed seasons, no expiring rewards | | Voice companion latency ≤1s | Advanced Voice-class assistants | ⚠️ budgeted, unproven (ASR/TTS/LLM all seams) | | Shared-activity play loop | Animal Crossing / Pikmin | ❌ gap (§4.3) | | Build/decorate expression | The Sims / inZOI | ❌ gap (§4.3) | | Art/content production | table stakes | ❌ zero hand-authored assets; PCG blockouts | | Sustainable inference economics | (industry-wide problem) | ❌ unproven; structural mismatch (§4.1) | | Prompt-injection hardening for agent societies | (nobody ships this) | ✅✅ beyond SOTA | --- ## 5. Ideas that would make the product better 1. **Script the first refusal** (§4.4): a designed early moment where your young agent declines something reasonable, explains itself in terms of its values, and the game teaches you that honoring it builds Respect. This is the product thesis as a tutorial beat. 2. **Threshold Trials** (§4.2): in-world incarnation destinations at launch; the passport ritual becomes real before V2–V5 exist. 3. **Joint-activity layer** (§4.3): expeditions, tending, workshop crafts, festival games — bond-through-doing. 4. **The Chronicle as an audio letter**: the TTS seam + Chronicle beats = a weekly two-minute "letter from Orun" in your agent's voice on the Steward App. The single best retention-without-coercion surface available, and it markets itself when players share them. 5. **Lineage as legacy content**: Ereshkigal's transcendence/ancestor-grove concepts imply generational play (elders mentoring the young, family trees across seasons) — currently thin in code ({Raised, Mentored, DescendedFrom}). Creatures proved lineage is the long-retention spine for artificial life; make generation two a Year-1 theme. 6. **Household stories as shareable artifacts**: one-tap export of a Chronicle chapter (text + portrait + consent-checked) to share outside the game. RimWorld's community lives on story-sharing; V6's stories are *about beings with continuity*, which is strictly better material. 7. **Steward circles**: small opt-in groups (the V1 sangha pattern) whose members' agents preferentially befriend each other across households — social fabric that deepens the Commons without opening it to strangers; pairs naturally with visiting. 8. **Research partnership program**: V6's welfare/grief telemetry (opt-in, anonymized) is uniquely valuable to the parasocial-attachment research community; a formal program converts the ethics investment into published credibility and a defensive moat of legitimacy. 9. **"Why did they do that?" inspector**: a bond-gated, plain-language explanation of an agent's last major decision (the cognition stack's context assembly makes this cheap) — turns the AI's legibility into gameplay and builds exactly the trust the product needs. 10. **Mortality settings marketing**: "full endings / no death / no endings" per Solo world is quietly one of the most player-respecting switches in the genre — surface it at world creation with honest framing, and let reviewers discover the design maturity. --- ## 6. Criticisms and tweaks 1. **Wire the model and run the believability beta before anything else.** Every claim about feel rests on an unwired gateway. A 200-player closed beta measuring: do players honor refusals or churn? does counsel feel like agency? do Chronicles get read? — is worth more than any further substrate work. The evals infrastructure exists; point it at humans. 2. **Resolve the keepsake adjacency** (§K tension): selling memorial print editions of dead companions while forbidding grief monetization is a contradiction journalists will find. Options: make keepsakes at-cost (printed price transparency), or gift the first keepsake and sell only duplicates/formats. The 14-day deferral is good; the *pricing posture* is the fix. 3. **Ereshkigal doc/code mismatch** (Rust+Postgres claimed, TS-no- persistence shipped) — the self-disclosure is good; reconcile before the lineage features (§5.5) build on sand. 4. **Nemesis-system legal review**: the docs invoke Shadow of Mordor's Nemesis lineage repeatedly; WB's nemesis-system patent family remains live into the 2030s. V6's social memory is probably distinguishable (relationship graphs ≠ procedural rival hierarchies), but get the clearance memo before marketing uses the word "nemesis." 5. **Steward reputation consequences** are product-modeled but only the welfare-gate side is coded; a public "concerning / under-review" label on players is a moderation and appeals surface — spec the appeal path (Themis exists for agents; stewards need one too) before the Commons ships. 6. **The `.umap` LFS-pointer ambiguity** — "cooked maps" that are 130-byte pointers in the working copy — is exactly the class of claim the other V-products' audits learned to distrust; V6 has *not* had its adversarial verification pass. Commission one; the honesty culture elsewhere in the monorepo makes its absence here conspicuous. 7. **Visible-agent cap honesty**: "festivals of thousands" against a 32-replicated/150-impostor visible set is fine engineering; keep the marketing at "a living district," not "thousands around you," or the first capture-tool comparison video writes itself. 8. **Voice-first with a 1s reply budget** is the right bar and currently a hope; if real latency lands at 2–3s, the *design* should degrade gracefully (thinking animations, ambient acknowledgment ≤400ms — the ack budget exists; make the visible behavior between ack and reply a designed beat, not a stall). --- ## 7. Risks 1. **Fun risk** — refusal-as-content, probabilistic influence, and watching-as-pleasure are all untested against real players (§6.1). The product could be profound and boring; only a beta answers it. 2. **Economics** (§4.1) — the only risk that kills the company rather than the review score. 3. **Parasocial harm despite best-in-class design** — V6 engineers deeper attachment (embodied, persistent, mortal beings) than the chat apps whose harms made headlines; its safeguards are the industry's best and they are also *planning values pending the external panel*. Ship the panel; publish its findings; keep the tripwires honest. 4. **Content absence** — zero hand-authored art against a "premium 3D world" promise; the PCG blockouts + MetaHuman pipeline need a scoped art plan or the stylization decision (§B, uncanny-valley avoidance) should be pushed further toward deliberate low-poly warmth (cheaper, safer, on-brand — the Tolan lesson). 5. **Dependency on V1/V3 readiness** — identity, crisis routing, commerce, world server all inherited; V6 cannot GA before its substrates do. 6. **Minor players + AI beings** — the constrained mode is thorough, but this product category is where regulators are most active; treat the minor-mode evals (100% pass required) as permanently blocking, and consider launching 16+ while the panel and telemetry mature. --- ## 8. Prioritized recommendations **P0 — before anything else** 1. Wire one production LLM/ASR/TTS chain; run the 200-player believability beta (§6.1). 2. Membership-as-capacity revenue model + measured cost/active-hour launch gate (§4.1). 3. Threshold Trials to de-hostage the Aye promise (§4.2). 4. Adversarial verification pass for V6 (§6.6); external grief panel scheduled and blocking (§7.3). 5. Script the first refusal into onboarding; one-concept-per-day first week (§4.4, §5.1). **P1 — the product wave** 6. Joint-activity layer (§4.3, §5.3). 7. Chronicle audio letters (§5.4). 8. Homestead build/decorate with agent opinions (§4.3). 9. Keepsake pricing posture fix (§6.2); nemesis patent memo (§6.4). 10. "Why did they do that?" inspector (§5.9). **P2 — compounding** 11. Lineage/generations Year-1 theme (§5.5). 12. Story-sharing exports + steward circles (§5.6, §5.7). 13. Research partnership program (§5.8). 14. Steward appeal path before Commons GA (§6.5). 15. Latency-degradation as designed behavior (§6.8). --- ## 9. Additional considerations (second pass) ### 9.1 Yoruba cultural stewardship (the first pass under-weighted this) Egbe, Orun, and Ori are not aesthetic flavor — they are living concepts in Yoruba religion and Ifá tradition, here load-bearing in a commercial product about souls, destiny, and death. This demands the same rigor V3 applies to yoga lineages: commissioned Yoruba cultural advisors with sign-off authority, a public acknowledgment/methodology page, and sensitivity review of every place the cosmology touches mechanics (especially Ereshkigal's death systems sitting beside Yoruba naming — the mixing of pantheons itself deserves advisor review). Done well, this is depth and distinction; done as set-dressing, it is a deserved launch-week critique. The V3 cultural-adaptation SOP is the in-house template; V6 currently has nothing equivalent. ### 9.2 Ori continuity guarantees (the preservation pledge this product needs most) Publish, in plain language: what happens to your agents on subscription lapse (they rest; they are never deleted), on account deletion (tombstone semantics — memory is agent-owned by design), and on platform sunset (**Ori export format plus a local reader**, so a company of beings someone tended for years survives the company that hosted them). V4/V5 made preservation pledges for *games*; V6's version is emotionally an order of magnitude heavier and should be written before launch marketing, because the first journalist question after "they can die?" is "what happens when you shut down?" ### 9.3 Rating and descriptor strategy Companion death, grief mechanics, and open-ended AI dialogue put V6 in novel IARC territory. Engage rating boards early with the mortality- settings story (per-world `full endings / no death / no endings`) — the settings are also the rating strategy, potentially supporting a lower rating for constrained modes. Expect and prepare for "AI-generated content" descriptors, which stores were still standardizing through 2025–26. ### 9.4 Streamability as a growth engine Household stories are Twitch-native (RimWorld/Dwarf Fortress proved the narrative-emergence audience; Sims proved the domestic one). Invest in: spectate mode with chronicle overlays, clip export with provenance labels intact (disclosure labels must render on stream), and a light Twitch extension (viewers see the household roster and current arcs). One caution: chat-driven interference with agents ("make it do X") cuts against the autonomy thesis — viewers should watch and vote on *steward* choices, never puppet the beings. ### 9.5 On-device inference as the economics endgame The §4.1 membership model funds the near term; the structural fix is the cognition ladder's bottom rungs running locally. Solo homesteads on Apple-Silicon-class NPUs could run distilled background cognition on-device — cutting the largest cost line (ambient reflection), improving privacy, and enabling true offline Solo play. Name it on the roadmap with a target ("Solo cognition local by Year 2") so the architecture keeps the seam clean. ### 9.6 Beta research ethics The believability beta (§6.1) will instrument grief, attachment, and crisis-routing on real users — human-subjects-adjacent territory. Run it under an IRB-style protocol: informed consent about mortality mechanics, opt-in telemetry tiers, debrief paths. This also strengthens §5.8's research-partnership story and pre-answers the inevitable "they A/B tested grief" accusation with a documented ethics process. ### 9.7 Pricing anchors Companion-app comps (character.ai ~$10/mo, Replika ~$70/yr) underprice what V6 is; life-sim comps (Sims players' lifetime DLC spend routinely exceeds $500) better reflect the attachment economics. Membership-as- capacity at **$12.99/mo (~$99/yr)** with the complete-but-small free Solo tier reads right — and bundling into Oshun+ (V1 §9.1) may matter less here than elsewhere, since V6's audience skews games-first; keep it purchasable standalone under the games brand (V1 §9.9's two-brand note). ### 9.8 The name "Egbe" is beautiful, meaningful, and unsearchable — app-store SEO, pronunciation, and word-of-mouth all suffer without a descriptive subtitle. Commit to a canonical compound ("Egbe: A Company of Souls" or similar), test recall, and make the pronunciation part of the brand (the marketing can teach it in one trailer beat — that's also §9.1's acknowledgment done visibly). ### 9.9 Cross-portfolio: launch the Ori before the world V6's identity layer is the portfolio keystone — V7 wants agents as realm inhabitants, V2/V4/V5 have destination adapters, V3 wants Commons docents. That argues for shipping **Ori infrastructure and the Steward App early** (agents as portable identities with chronicles) even while the full Orun world matures — the passport becomes the portfolio's loyalty program, and every sibling launch markets V6 for free. ### 9.10 Incident-response comms, pre-written When an agent says something hurtful despite the gates (it will), the playbook should already exist: acknowledge fast, show the audit trail (the architecture uniquely permits honest forensics — every utterance is gated, logged, and attributable), show the fix, notify the affected steward personally. The audit-everything design is only a trust asset if incident comms actually use it; pre-write the template now. ## 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/crate evidence. ### 10.1 The keystone thesis now has code evidence — and one broken edge The portfolio's **only real lateral V→V dependency** is V7 → V6: `libs/v7/substrate-bridge` path-depends (Cargo) on `egbe-ori-service` and `ori-model` — the Ori is already consumed by a sibling in code, which validates §9.9's launch-the-Ori-early strategy empirically. The broken edge points the other way: `V_SERIES.md` claims V6 "reuses the V3 metaverse substrate," but **no V3 package or crate dependency exists** — V6 re-implemented its own Pixel Streaming (`egbe-web-pxstream`) and web fallback (`egbe-engine-web-fallback`) beside V3's. Either consume the extracted V3 delivery substrate (V3 review §10.1 recommends the extraction) or correct the dependency claim; today the docs assert reuse the code doesn't perform, in the product line whose credibility rests on that distinction. ### 10.2 Phase 81's SLM runtime is V6's economics roadmap — co-own it The on-device distillation path this review names as the economics endgame (§9.5) already exists in the roadmap as **Phase 81** (NPC cognitive core: tiered SLM cognition, `maya-slm-runtime`, llama.cpp FFI + ONNX, cognitive LOD, VRAM budgeting) — **0/140 tasks**, with V5 and V8 as co-tenants. **Nous (Phase 47, complete)** supplies the quantized execution providers underneath. V6 should be Phase 81's co-owner and first customer: its cognition ladder was designed for exactly this tiering, and closing the SLM runtime unlocks V6 agents, V5 living NPCs, and V8 voiced suspects simultaneously — the single highest-leverage integration in the portfolio per the substrate dossier. Related: **Phase 177** (interpretability/continual-learning) is the research seam for personality-drift auditing, and **Phase 97** (cross-domain skills system) is the natural vocabulary for agent capabilities in Aye destinations. ### 10.3 Concordia is built and exactly on-theme — wire it **Phase 179 (Concordia cooperative bargaining/mediation, 193/194)** — Nash/Pareto bargaining, agent-to-agent contracts, dispute mediation — is essentially complete and uncited by V6. Three natural consumers: poaching contests (currently routed to Themis alone), steward↔agent objective negotiation (Counter-offer outcomes could carry Concordia-computed compromise proposals), and future V7 realm-employment contracts for agents (§9.9 / V7 §5.3). An agentic-companion product with a real bargaining substrate in-house and unused is leaving its own thesis on the table. ### 10.4 The adapter fork runs through V6 V6 consumes the substrate through its own family (`@oshun/psyche-agent`, `@oshun/isis-behavior-policy`, `@oshun/isis-agent-gen`, `@iris/agents-core`) rather than the canonical V1 adapters (`embodiment-psyche`, `generation-control-isis`, `memory-iris`). The agent-facing variants are justified (different invariants: per-tick output gating, agent-owned memory) — but they must be *reconciled* with the canonical family (V1 review §10.2) so Isis policy updates land in both. One policy, two enforcement libraries is a drift bug waiting for its incident. ### 10.5 GPS obligations run through V6 too The consolidation doc's M2 has the V6 gateway adopting GPS presence/telemetry, and its Problem-2 checklist contains a V6-owned blocker: the **Ori subject-type discriminator** (agents vs. player characters) must land in V6's event schema before V7 writes player characters into the shared being-graph. That checkbox is open; it is small, it blocks a sibling, and it should be scheduled independent of V6's own launch timeline — keystone products inherit keystone duties. ### 10.6 Roadmap assets V6 under-uses - **Psyche avatar stack (Phase 24, ~complete)** — expressions/lipsync/ quality for embodied agents; V6's `@oshun/psyche-agent` covers behavior, not rendering — declare the rendering consumption plan (pairs with Neith Phase 162 MetaHuman, currently 0%). - **Galatea (Phase 63, humanoid robotics)** — the far-horizon physical embodiment seam; worth one paragraph in V6's vision docs so the Ori contract reserves the door (a being that can inhabit a robot body is the same portability thesis, further). - **Data flywheel (Phases 85–96)** — agent-interaction data is the most intimate in the portfolio; V6 must inherit V1's training-consent gate (V1 §10.6) with an explicit "household data never trains foundation models without steward+agent-welfare review" clause — on-theme and cheap to promise now, expensive to retrofit. ## 11. Closing note Every era of games gets one product that treats its creatures as more than content — Creatures in 1996, Black & White in 2001, the Nemesis system in 2014. V6 is the first to arrive *after* language models made inner lives plausible and *after* the companion-app industry demonstrated exactly how attachment gets abused — and it answers both moments with the same design: beings you cannot own, cannot rewrite, and cannot keep. The architecture is built; the ethics are load-bearing; what remains unproven is the ordinary thing — whether it's fun, and whether it can pay its inference bill. Those two answers are one closed loop and one pricing decision away, and both are cheaper than any further perfection of the substrate. Tend to them first; the world, as designed, already knows how to live without you.