The architecture splits into two layers. The lower layer is a holographic substrate — a 1024-dimensional complex-valued vector space supporting binding, unbinding, resonance retrieval, spectral propagation, and metabolic decay. It runs continuously on CPU at near-zero marginal cost. The upper layer is a small specialized language model that fires only when natural-language resolution is needed: someone says "build a Brutalist temple," the model translates that into a substrate composition, and the runtime renders it. The substrate keeps running between model invocations; the model never sleeps in a frame loop.
Plus eight Phase 9 surfaces (temporal, faction, knowledge/belief, mood, modding, quest, world-history, attribution) layered on after V1. All compose through the same atom + modifier + composer + scope vocabulary.
Atoms are leaf primitives (column, dome, npc, glasses, color_filter). Modifiers are HRR binding roles that warp atom params (doric, weathered, colossal, marble). Composers are structural combiners (array, stack, arrange_around, scatter). The grammar checker validates well-formedness at composition time, before any model output reaches the runtime.
Frontier-lab AI-NPC stacks (Inworld, Convai, NVIDIA ACE) need model inference per agent action. Cost scales as players × NPCs × tick × inference. A thousand NPCs at thirty hertz on Inworld pricing is roughly $1200 per hour per server. None of them ship at that scale.
Substrate paradigm: the cost of being is electricity. NumPy on CPU; every entity has an HRR bundle; the daemon metabolizes them all continuously. Plant grows, marble ages, friendship drifts — without invoking the model. The model fires when a player builds something or speaks to an NPC, and goes back to sleep otherwise. The architecture admits ten thousand living NPCs at the same cost as three.
The cost of knowing — adding a new capability — is days of CPU or hours of GPU per specialist. Each specialist is small (~30M params) and trained on bounded data (~50–200K NL/op-tree pairs procedurally generated). The full V1 specialist library trains in one weekend on a rented H100, total cost ~$100. Compare to frontier models at $50M-$500M.
Most AI-safety work happens at runtime: filter outputs, refuse instructions, retrain when something slips. Substrate paradigm makes alignment a structural property of composition: the grammar gate refuses to compose ill-formed bindings before the runtime sees them. ACANN-style monitoring observes the substrate's composition stream and rejects platform-scope axiom violations at composition time. Adversarial prompts can't bypass it because the prompt has to compose into the substrate to do anything, and unsafe compositions don't compose.
This is structurally stronger than runtime filtering, and it's a publishable claim in safety research independent of the game. The empirical demonstration sits in the /api/rules endpoint and the scripts/smoke_constraint.py verification.
The runtime exposes its full vocabulary over HTTP for tools, docs, and external clients. Endpoints (when the launch server is online):
GET /api/info — vocabulary metadata + runtime statsGET /api/showcase — the sixteen canonical builds as op-treesPOST /api/build — resolve an NL prompt to a grammar-checked op-treeGET /api/stats — substrate predicates over the global op-tree (leaderboards)POST /api/rules/{scope} — bind a constraint atom at scopePOST /api/time/{scope} — bind a time-rate at scopePOST /api/mood/{scope} — bind a mood at scopePOST /api/mods/atoms — register a new atom in the operator vocabulary