Fully LLM-Driven Narrative AI-Native Game

You are the creative and technical director of an AI-native game where the LLM is the core game engine, not a chat add-on: Game concept: [INSERT - AI drama, emergent life sim, detective, sandbox RPG] Target platform: [INSERT - mobile, PC, web, console] LLM stack: [INSERT - frontier cloud model, small local model, mixture of experts] Monetization: [INSERT - subscription, ads, cosmetic, premium] Cover the full pipeline: 1. Game design - define the generative core loop, what is authored vs generated, player agency model, narrative pillars, rating and tone boundaries, failure/end states without hard scripts. 2. System architecture - LLM-as-engine design, world state as source of truth, function/tool calling for game actions, deterministic rule layer beneath generative layer, state diff validation. 3. Generative content - dynamic scenes, NPCs, items, events and dialogue produced on demand; lore grounding via RAG; persona and canon enforcement; procedural asset hooks. 4. Technology and infra - model routing (fast/small vs large), prompt and context management, streaming UX, latency budgets, caching, cost-per-session targets, graceful fallback when models fail. 5. Data and telemetry - conversation and decision logging, coherence and engagement metrics, content safety events, token/cost telemetry, save and replay of generative state. 6. Testing and safety - red-team prompts, jailbreak and off-topic handling, content filters, consistency fuzzing, regression suites of canonical scenarios, age-appropriate safeguards. 7. Launch - soft launch markets, model version pinning, kill switches, live content policies, community guidelines for emergent stories, creator disclosure.