# Agentic Content Generation Across V1–V7 — Findings & SOTA Recommendations Date: 2026-06-12 Scope: every mechanism by which AI agents (Hathor, Yemaya, Isis, Iris, and related domains) build content and materials from scratch for the V1–V7 product line, with human-in-the-loop studio review. Method: seven parallel read-only code/spec audits (Hathor, Yemaya, Isis/Bellona/Nous, Iris and agent-runtime survey, V1–V3 sweep, V4–V5 sweep, V6–V7 sweep), each grounded in actual source reads, cross-checked against the 2026-06-12 remediation ledgers and `V_SERIES.md`. --- ## 1. Executive Summary **The vision is already specified at a state-of-the-art level; the execution is fragmented into islands that have never been connected end-to-end.** The single most important sentence in this report: > No AI agent in this monorepo can today take a content brief and land a > finished asset inside a V2–V7 game build, because the chain breaks at exactly > two places: (1) there is no shared agentic execution runtime — at least seven > domains run their own incompatible agent loops — and (2) Bellona's Unreal > Engine cook/import path is declared in types and schemas but not implemented, > so nothing generated can reach a UE5.5 project as a real asset. Everything else about the platform is unusually strong. V1's Agentic AI Studio spec (`V1/features.md:4533–4950`) already describes a governance plane — agent registry, budget inheritance, visible plan DAGs, approval gates, kill switches, provenance ledgers, champion-challenger rollout — that matches or exceeds current industry SOTA for governed agentic systems. Hathor has a real NPC cognition stack with live LLM dialogue generation. Yemaya has a 30+ agent production framework with 76 AAA quality validators. Isis has a real ComfyUI/RunPod generation factory with 45 certified workflows. The work remaining is overwhelmingly **integration, orchestration, and verification**, not invention. Priority order (detailed in §6): fix the Bellona→UE cook path; unify the agent loop in Iris; wire Hathor's LLM capability into its narrative pipelines; prove one Yemaya pipeline end-to-end on a real V2 deliverable; then scale the same chain to V3–V7 under the V1 governance plane. --- ## 2. The Platform Engine Room — Domain-by-Domain Truth ### 2.1 Hathor — Worldbuilding & Narrative Engine: ~70% capability, ~0% orchestration **Real and production-grade:** - World/domain models, quest engine, dialogue trees, story graphs, and exporters (Ink, Yarn Spinner, JSON) — `libs/hathor/narrative/src/`. - Simulation engines with genuine domain algorithms: economy (supply/demand, trade-route profitability), politics (faction relations, war resolution, coups, elections), culture (trait drift, subculture emergence) — `libs/hathor/simulation/src/`. - Lore validators: timeline, causality, taxonomy, contradiction — `libs/hathor/validation/src/`. - Lore compiler to engine formats (Unreal/Unity/Godot/Blender), screenplay (Fountain/FDX), and Bellona packages — `libs/hathor/lore-compiler/src/`. - **NPC cognition + live LLM dialogue**: Big Five + Enneagram + BDI personality, emotion decay model, dual-store memory with consolidation, world awareness, and a `DialogueGenerator` that makes real LLM calls (`libs/hathor/llm-npc/src/dialogue/dialogue-generator.ts:79–127`) behind a production circuit-breaker provider chain (`libs/hathor/llm-npc/src/provider-chain.ts`) supporting OpenAI-compatible, NVIDIA ACE, Inworld, and Convai backends. - Two shipped game services built on the stack (`apps/hathor/svc-veilborn-*`, ~42k LOC). **The gaps that block agentic generation:** 1. `DialogueGenerator` is never called from any quest/narrative generation path. There is no `generateQuestNarrative()`, no LLM-written story beats. 2. The procedural narrative/quest generators that do exist are **template-substitution, not LLM generation**, and they live in the wrong domain — `libs/neith/integration-hathor/src/content-creation/` (tension-curve templates, motif-affinity matrices, complication dependency graphs). Good algorithms, but vocabulary-pool prose. 3. No Iris integration: Hathor's provider chain does not know Iris exists. 4. No lore-consistency checking of _generated_ output — validators check authored structures; nothing verifies an LLM response against canon. 5. No human review workflow — merge requests exist for collaborative edits, but there is no generated-content approval gate, review UI, or feedback-to-regeneration loop. 6. The V2 reciprocal contract (`DOMAINS/hathor/specifications.md` §9.6: `FighterTimeline`, `RivalryMatrix`, `FactionMatrix`, `LadderEndingGraph`, `SideStorySequencerOutline`) is **not implemented** in `@hathor/domain-models` / `@hathor/narrative` / `@hathor/lore-compiler` — and V2 Story/Side Stories/Krypt/Chronicles/DJ Story cannot ship without it (`V2/docs/integration/hathor-narrative-contract.md:148–154`). ### 2.2 Yemaya — Creative Production Studio: the most complete agentic framework, unverified in execution **Real and substantial (`libs/yemaya/`):** - `PlanningAgent` with LLM-driven plan generation, HTN task networks, alternatives, contingency plans, and an explicit human-in-the-loop refinement hook (`libs/yemaya/agents/src/planning/planning-agent.ts:349–443`). - 30+ creative agents (AIDirector, StoryDirector, CharacterWriter, Cinematographer, ConceptArtist, …) with per-agent LLM model assignment (`libs/yemaya/agents/src/config/agent-llm-config.ts`). - 76 AAA quality validators (polygon budgets, dialogue naturalness, narrative coherence, pacing, accessibility) — `libs/yemaya/agents/src/quality-assurance/`. - End-to-end pipeline definitions: concept → `sophia:analyze` → `hathor:compile_lore` → `hathor:create_character` / `isis:generate_3d` → `mcp:execute_dcc` → `yemaya:review` (`libs/yemaya/autonomous-pipelines/src/pipelines/game.ts:65–322`), with a domain-prefix dispatcher (`libs/yemaya/orchestration/src/execution/dispatcher.ts`). - A complete HITL type system: approval gates with risk levels, timeout actions (reject/escalate/auto-approve), quorums, escalation, feedback categories (`libs/yemaya/agents/src/hitl/index.ts`). **The gaps:** 1. **HITL is a terminal pipeline step, not a mid-pipeline gate.** Review happens after generation completes; nothing can pause a running pipeline, collect direction, and resume. 2. Cross-domain dispatch (Isis/Hathor/Sophia/Bellona clients) is interface- and helper-level; no verified end-to-end run exists, and recovery modules (replanning, checkpointing) are not wired into the dispatcher. 3. The HTN planner's output is never translated into dispatcher steps — the shipped pipelines are static factories, not dynamically planned. 4. Agent LLM calls have no budget enforcement, rate limiting, or circuit breaking at call time. 5. Root `TODOS.md` Phase 26 (26.1–26.5) is fully `[x]`-marked, but the audit indicates this reflects framework completion, not verified autonomous production runs. Treat those marks per the repo's adversarial-verification standard. ### 2.3 Isis — Generative AI Factory: mostly production-ready Real ComfyUI integration, **45 certified workflow templates**, Redis-backed job queue, RunPod serverless GPU support, multi-provider image/video/3D pipelines (`libs/isis/{workflows,comfyui-nodes,ai-providers,3d-generation,ai-video}`). Gaps: audio generation pipelines incomplete; **C2PA cryptographic signing not implemented** (provenance bundles exist, signatures don't); end-to-end tests mock model outputs rather than exercising real generation. ### 2.4 Bellona — Engine Bridge: **the broken link** UE5.5 asset cooking is declared in the database schema and WebSocket bridge types, but: no headless `RunUAT` invocation exists, no UE C++ plugin source is in the monorepo, no cooked `.uasset` production path works, and the WebSocket commands route to external Unreal Editor instances that are not provisioned. The lore-to-engine compiler has type definitions but no actual Blueprint/C#/ GDScript code generation. **Every generated asset for V2–V7 dead-ends here.** This is corroborated from the product side: V2 has exactly **1 `.uasset`** on disk, V4 has 504+ JSON stand-ins, V5 has zero binary assets (`V_SERIES.md:56–61`). ### 2.5 Iris — the natural seed of the shared agent runtime The only domain with a production-shaped, reusable agent substrate: - `AgentRuntime` with lifecycle management, per-agent priority task queues, timeout/abort propagation, permissions with escalation (`libs/iris/agents/core/src/agent-runtime.ts`). - Multi-agent machinery: orchestrator, collaboration protocol, conflict resolver, handoff manager, result aggregator (`libs/iris/agents/multi-agent/src/`). - Model orchestration with circuit breakers, rate limiting, load balancing, fallback chains, and cost/latency metrics (`libs/iris/conversation-orchestration/src/model-orchestrator.ts`). - Anthropic provider with tool use, streaming, extended thinking, prompt caching (`libs/iris/conversation-providers-anthropic/src/anthropic-provider.ts`); MCP tool discovery and schema validation (`libs/iris/mcp/src/tool-discovery.ts`); tiered memory (core/working/archival/episodic) in `libs/iris/memory/`. **What Iris is missing to be _the_ runtime:** the actual agentic tool-call loop (call LLM → execute tool calls → feed results back → repeat until done, with a max-iteration guard); a normalized tool-calling protocol across Anthropic/OpenAI/Google/local; budget/cost constraints flowing through the execution context; and a streaming aggregation layer for long multi-agent chains. **The duplication problem:** at least seven independent agent loops exist — Iris `agents/core`, Hathor `llm-npc/provider-chain`, Yemaya `agents` (own LLM config and planning loop), Isis `agent-consensus`, V6's Rust cognition tier, plus app-level loops in `apps/{maat,veritas,kalika,lakshmi}`. Every one re-implements provider fallback, retries, and orchestration with no shared budgets, no shared observability, and no shared governance. ### 2.6 Supporting domains - **Sophia** (RAG/knowledge): mature; ready today to ground generation agents in canon/lore/research. Hathor's Sophia integration exists for citation and fact-checking but is not used to ground _generation_. - **Nous** (model infra): real ONNX inference, embeddings, LLM completions, vision; generative-media control packages are structural stubs. - **Psyche** (avatars/voice/Conductor): carries the V1 Living Scenes runtime; real session/turn-taking/lip-sync machinery. --- ## 3. Product-by-Product Surfaces (V1–V7) ### V1 — Oshun platform: the governance plane (specs strongest in repo) | Surface | Spec | Est. complete | Notes | | ----------------------------- | ------------------------------ | ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Isis generation control plane | `V1/ARCHITECTURE.md:874–893` | ~40% | Workflow/model/provider registries, dev→staging→prod promotion, release gates, staged rollout, rollback, ProvenanceBundle, immutable generation ledger | | Living Scenes | `V1/features.md:3619–4531` | ~65% | Scene Score + Conductor + Blend Kernel; determinism is a release gate; shareability matrix; takedown cascades | | **Agentic AI Studio** | `V1/features.md:4533–4950` | ~50% | Agent registry, AgentRun envelope, durable runs, visible plan DAGs, operator approval gates, budget inheritance, authority-subset delegation, kill switches, anomaly auto-quarantine, 14-tool catalog, champion-challenger feedback | | Studio authoring | `V1/ARCHITECTURE.md:1970–2030` | ~55% | Editorial lifecycle state machine, mandatory human checkpoints, publication gates, accept/reject/regenerate-with-direction review UX | | Hathor studio bridge | `V1/TODOS.md` §34 | ~25% | 74 walked surfaces, zero verified (added 2026-06-12) | | Persona/voice packs | `V1/ARCHITECTURE.md:1941–1967` | ~60% | Consent + lineage + watermark + revocation cascade | V1 also fixes the **tier boundary**: AAA power-user generation surfaces live in Yemaya Studio (`apps/yemaya/studio-{web,desktop}`), curated cards on the contemplative product; promotion from AAA outputs into curated surfaces flows through the editorial lifecycle (`V1/TODOS.md:4066–4090`). ### V2 — Fighting game: best content contract, zero content `V2/docs/integration/hathor-narrative-contract.md` is the model for how every V product should consume agentic content: Hathor owns source; V2 consumes only compiled artifacts with content-addressed manifests (`worldId`, `versionHash`, artifact hashes, Bellona cook lineage). `hathor-llm-npc.md` specs runtime LLM NPCs. Reality: `hathor-npc-adapter` is README-only, the §9.6 reciprocal Hathor records are unimplemented, and there is 1 `.uasset` in the project. Spec is far ahead of implementation. ### V3 — Lilith Metaverse: strongest generation governance per artifact Concert authoring pipeline with deterministic gates (`V3/CONCERT_AUTHORING_PIPELINE.md`), C2PA on every export (`V3/C2PA_EVERY_EXPORT.md`), voice-cloning consent registry, cultural adaptation review, eval-set release gates. Gate evidence is structural, not operational (`V3/REMEDIATION_2026-06-12.md`). ### V4 — Tactical universe: one real procgen system, the rest stubs `V4Procgen` chunk-based mission assembly with a real constraint solver is genuine. Calliope AI commentary and the Sophia help bot are specified and `[x]`-marked but the services are `/health`-only (12/13 services). 770 JSON asset stand-ins await a working generation→cook chain. ### V5 — Open-world narrative: the largest content famine Mind Palace deduction graph (110 authored evidence/deduction pairs with Isis edge-vetting) is real design. But ~12,000 planned dialogue nodes are synthetic placeholders, dialogue automation crashes, procedural side quests have no runtime integration, and 14/16 services expose zero routes. V5 is the single biggest beneficiary of an agentic content pipeline. ### V6 — Egbe: the agentic flagship, currently a skeleton Tier-1 spec: three-tier cognition (Clotho/Lachesis/Atropos), event-sourced Ori identity, pgvector memory, Sophia-grounded generation, Isis/Lilith safety gates, steward-not-owner consent model, 13 behavioral eval suites. Reality: ~8.6k LOC of Rust service skeletons with **zero inter-service integration, no persistence, no LLM wired anywhere**. V6 needs the shared runtime more than any other product — building its cognition stack privately in Rust while Iris/ Yemaya/Hathor build theirs in TS would be the fourth parallel reinvention. ### V7 — Mawu: real foundations, AI-assist missing forge-resolver / forge-conflict / forge-sandbox / forge-compositor are real, substantive code (36/120 tasks verified real on 2026-06-12). 29 adversarial eval gates specified. AI-assisted creation tooling — the agentic piece — is missing or skeletal. --- ## 4. Cross-Cutting Findings - **F1 — The spec layer is SOTA; the integration layer is the deficit.** V1's Agentic AI Studio + Isis control plane already specify what the industry calls governed agent orchestration (budgets, DAG visibility, kill switches, provenance, eval-gated promotion). Nothing needs to be re-envisioned; it needs to be built once, shared, and consumed by V2–V7. - **F2 — Seven agent loops, zero shared runtime.** Provider fallback, retry, consensus, and planning logic is quintuplicated across Iris, Hathor, Yemaya, Isis, V6, and app layers. No shared budgets, observability, or governance can exist until this converges. - **F3 — The chain breaks at Bellona.** Generation works (Isis), narrative compilation works (Hathor lore-compiler), but engine delivery is types-only. This single gap is upstream of V2's missing content, V4's 770 stand-ins, and V5's zero binary assets. - **F4 — Hathor has the muscles but no nervous system.** Live LLM dialogue, rich world simulation, and validators exist side-by-side, unconnected; its generative text pipelines are template-based and live in Neith. - **F5 — Yemaya's HITL is post-hoc.** Human review must move inside the loop (pause/redirect/resume), or "human-in-the-loop" is actually "human-after-the-fact." - **F6 — Grounding and safety rails exist but aren't bound to generation.** Sophia grounding, Lilith policy, lore validators, and consent registries all exist; none is a mandatory pass on generated content today. - **F7 — Eval machinery is fragmented.** Yemaya's 76 validators, Isis quality gates, V3 release gates, V6 eval suites, and V7 adversarial gates are five disconnected quality systems with no shared "content release gate." - **F8 — Checkbox truth lags code truth in both directions.** Phase 26 is over-marked (framework ≠ verified execution) while Hathor's real LLM-NPC capability is under-leveraged. The 2026-06-12 remediation discipline should extend to the platform domains' agentic claims. --- ## 5. Target SOTA Architecture One stack, six layers, every V product a consumer: ``` L6 Delivery Bellona: headless UE cook (RunUAT), UE Python/Remote-Control import, content-addressed manifests + cook lineage (per V2 contract) L5 Governance plane V1 Isis control plane + Agentic AI Studio: registries, AgentRun envelopes, budgets, approval gates, kill switches, ProvenanceBundle + C2PA signing, immutable generation ledger L4 Production orch. Yemaya: HTN-planned dynamic pipelines, mid-pipeline HITL gates (pause/redirect/resume), failure recovery, cost caps L3 Capability tools Domain tool servers on a shared catalog (MCP-shaped, already enumerated in V1 features: sophia.ground, generate.image/video/ audio, composition.suggest, hathor narrative tools, persona.invoke, themis.adjudicate, approval.request, handoff) L2 Agent runtime Iris: THE agentic loop (tool-call iteration w/ max-iteration guard), normalized cross-provider tool protocol, sub-agent spawn with budget inheritance + authority-subset delegation, memory tiers L1 Model plane Iris model-orchestrator + Nous: provider gateway, routing, fallback, caching, cost metering (one place, not seven) ``` Design rules that make this SOTA rather than merely consolidated: 1. **Everything generated passes three mandatory gates**: grounding (Sophia pin set), policy (Lilith/Isis workflow-class binding), and canon (Hathor lore-consistency on generated output) — then an eval gate, then a human gate where the surface requires it. Provenance travels the whole way. 2. **Agents are products, not plumbing** (V1's framing): every run has a visible DAG, a cost ledger, evidence trails, and replay. That's what makes "fully agentic with studio HITL" auditable instead of hopeful. 3. **Determinism where it matters**: Render Envelope-style reproducibility (already specified for Living Scenes; V7 already treats determinism as an invariant) extends to all compiled content artifacts — same inputs, same hashes, or promotion blocks. 4. **Champion-challenger everything**: human review decisions become gold sets; new prompts/models/workflows ramp behind statistical readiness gates (already specified in V1; implement once in L5). --- ## 6. Recommendations & Sequencing ### Phase A — Unblock the chain (highest leverage, do first) - **A1. Implement Bellona's UE cook path.** Headless `RunUAT` BuildCookRun, a UE Python / Remote Control import bridge for generated assets, and the content-addressed manifest + cook-lineage record the V2 contract already specifies. Acceptance: an Isis-generated texture + a Hathor-compiled dialogue table land as cooked assets in `V2/ue` via one command. (A compiled UE 5.5 source build exists on the team's UE box for verification — build and run automation there, as `ueagent`.) - **A2. Build the Iris agentic loop.** Tool-call iteration with max-iteration guard, normalized tool protocol across providers, budget constraints in `AgentExecutionContext`, streaming aggregation. Ship it as `@iris/agents-core` v2 and write the migration shims for Hathor's provider chain and Yemaya's agent base so the duplicates can drain incrementally. ### Phase B — Wire Hathor end-to-end (the narrative spine) - **B1.** Implement the §9.6 V2 records (`FighterTimeline`, `RivalryMatrix`, `FactionMatrix`, `LadderEndingGraph`, `SideStorySequencerOutline`) in `@hathor/domain-models` / `@hathor/narrative` / `@hathor/lore-compiler`. - **B2.** Connect `DialogueGenerator` into quest/narrative generation: LLM-written quest intros, story beats, NPC backstories — using the Neith tension-curve/motif machinery as the _planner_ and the LLM as the _writer_ (move those generators from Neith into Hathor in the process). - **B3.** Add generated-content lore-consistency validation: run every LLM output through the existing four validators plus a Sophia-grounded contradiction check before it can enter a compiled package. - **B4.** Add the review workflow: generated artifact → review queue → accept / reject / regenerate-with-direction → publish, reusing V1 Studio's editorial lifecycle rather than inventing a new one. ### Phase C — Prove one Yemaya pipeline end-to-end - **C1.** Pick one vertical slice — one V2 fighter's Side Story pack (Hathor lore → dialogue → sequencer outline → Isis concept art → Bellona cook → playable in `V2/ue`) — and run it through the Yemaya game pipeline for real, fixing the dispatcher's domain clients as failures surface. - **C2.** Make HITL mid-pipeline: approval gates that checkpoint a run (durable runs are already specified in V1), surface review in Studio, and resume with feedback injected into the relevant agent's context. - **C3.** Wire the recovery modules (replanning, checkpoint restore) into the dispatcher; add per-step budget enforcement before dispatch. - **C4.** Translate HTN planner output into dispatcher steps so pipelines are planned, not hardcoded. ### Phase D — Stand up the governance plane as a service for all V products - **D1.** Finish Isis control-plane hardening (release gates, failover chains) and **implement C2PA signing** — V3 already mandates it per export. - **D2.** Implement the Agentic AI Studio runtime pieces against Iris v2: AgentRun envelope, budget inheritance, kill switches, anomaly quarantine, operator dashboard. V2–V7 pipelines register as agent families here. - **D3.** Consolidate the five eval systems into one **content release gate service**: Yemaya's 76 validators + Isis quality gates + V3/V6/V7 gate manifests become gate suites in a single eval-gated promotion flow with champion-challenger ramping. ### Phase E — Scale across the portfolio (in value order) 1. **V5** — the content famine: side-quest and dialogue generation at scale (12k nodes) through the B+C pipeline, with Mind Palace edge-vetting as the model for canon gating. 2. **V4** — codex/commentary/mission content; make Calliope commentary a real service consuming the shared runtime; procgen chunks feed the same review gates. 3. **V3** — concert/scene authoring via Tara Studio with the C2PA and consent registries it already specifies. 4. **V6** — re-platform the cognition tier on the shared runtime (Rust services consume Iris/Nous via the gateway rather than reimplementing); its eval suites become gate suites in D3. V6 is where "agents as first-class residents" pays off — don't build it on a private stack. 5. **V7** — AI-assisted creation in the forge (agent-as-co-creator inside the sandbox), last because it depends on everything above plus the V7 trust boundary. ### Standing rules (apply from day one) - Every `[x]` on an agentic-pipeline task requires a verified end-to-end run artifact (run ID + provenance bundle), per the repo's own adversarial-verification standard — Phase 26's marks should be re-audited under this rule. - No new agent loops. Any domain needing agency consumes `@iris/agents-core` v2 or files a gap against it. - Human sign-off surfaces (Studio review, operator workbench) are part of the definition of done for every generation feature, not a follow-up. --- ## 7. Verification Caveats - All domain maturity judgments above come from source reads by audit agents, not from checkbox ledgers; where the two disagreed (Phase 26, V4/V5 `[x]` marks), the source read wins. - Per-file line citations are as reported by the audits on 2026-06-12; spot-check before relying on exact line numbers in long-lived documents. - This report did not execute any pipeline; the Bellona/cook and Yemaya dispatch findings are absence-of-implementation findings (no RunUAT call sites, no verified cross-domain runs found), which is the strongest claim a static audit can make.