Playable Neural World Model

You are the principal AI researcher for [INSERT STUDIO], building a playable world model: a learned simulator that predicts next environment states from current state and player actions. Design a research-to-production plan for a controllable world model supporting prototyping, NPC planning, and content infilling in [INSERT GENRE]. Module 1 - Gameplay design: define which dynamics must be simulated (movement, physics, combat, NPC reactions), fidelity tiers, and where prediction is visible versus hidden. Module 2 - Generation pipeline: specify telemetry data collection, latent tokenization, architecture options (diffusion, transformer, autoregressive, video-style), action conditioning, and uncertainty gating. Define hybrid authority: the deterministic engine remains source of truth; the model only proposes states that are validated before commit, with re-sync rules on divergence. Specify rollout horizons, object permanence, long-horizon drift metrics, and compare against procedural-generation, pure-LLM, and template baselines. Define the creator workflow: how a UGC author prompts, steers, and overrides simulated outcomes. Module 3 - Data structures: provide schemas for state latents, conditioning frames, rollout traces, confidence scores, and divergence events. Module 4 - Testing: define counterfactual probes, physics-violation rates, drift metrics, human preference eval, and gameplay-equivalence tests against the real engine. Module 5 - Operations: cover training cadence, GPU budgets, on-device versus server inference, versioning, shadow-mode rollout, monitoring, and kill switches. Address safety: preventing the model from generating disallowed content during free-form interaction. State acceptance criteria for when prediction quality is sufficient to expose each fidelity tier to players. Provide a text architecture diagram, milestone plan from offline prototype to live feature, risk register, and KPIs. List assumptions about compute, data volume, and engine instrumentation. Start with a one-page research brief, then deliver modules one at a time for my approval. ================