Model Layers

Model layers are LODE’s primary use case. They describe pools of AI models available at a given tier in the StackNet registry.

Schema

model-layer-v1.lode


{
  "schema": "lode",
  "version": 1,
  "name": "ModelLayer",
  "fields": [
    { "id": 1, "name": "name",         "type": "string",             "required": true },
    { "id": 2, "name": "isDefault",    "type": "bool",               "required": true },
    { "id": 3, "name": "models",       "type": "string[]",           "required": true },
    { "id": 4, "name": "defaultModel", "type": "string",             "required": true },
    { "id": 5, "name": "defaults",     "type": "map<string,string>", "required": false }
  ]
}

TypeScript Interface


interface ModelLayer {
  name: string;                      // "preview", "fast", "turbo"
  isDefault: boolean;                // Default for bare aliases
  models: string[];                  // Pool of canonical model IDs
  defaultModel: string;              // Default from the pool
  defaults?: Record<string, string>; // Per-capability overrides
}

Example Layer

preview-layer.json


{
  "name": "preview",
  "isDefault": true,
  "models": [
    "gpt-oss:20b",
    "qwen3:30b",
    "qwen3:8b",
    "qwen3-coder:30b",
    "minimax-m2.5:229b",
    "qwen3-omni:30b",
    "qwen3-vl:30b",
    "nomic-embed-text",
    "devstral:24b"
  ],
  "defaultModel": "gpt-oss:20b",
  "defaults": {
    "code": "qwen3-coder:30b",
    "embed": "nomic-embed-text",
    "vision": "qwen3-vl:30b",
    "media": "qwen3-omni:30b"
  }
}

Compilation


$ lodec compile preview-layer.json -s model-layer-v1.lode -o preview.vein
Compiled preview-layer.json → preview.vein (187 bytes)
Fingerprint: d4e8f1a2...

The equivalent JSON is ~350 bytes (pretty-printed) or ~220 bytes (compact). The Vein binary is ~187 bytes — roughly 15% smaller than compact JSON with full type safety and verifiability.

Distribution

The registry serves model layers as Vein binaries:


// Server
app.get("/v1/registry/layers/:name/vein", (req, res) => {
  const layer = registry.getLayer(req.params.name);
  const schema = loadLode("model-layer-v1.lode");
  const vein = compile(layer, schema);
  const fp = fingerprint(layer);
 
  res.set("Content-Type", "application/x-vein");
  res.set("X-Lode-Fingerprint", fp);
  res.send(Buffer.from(vein));
});

Client Verification


const res = await fetch(`${registry}/v1/registry/layers/preview/vein`);
const vein = new Uint8Array(await res.arrayBuffer());
const expected = res.headers.get("X-Lode-Fingerprint");
const schema = loadLode("model-layer-v1.lode");
 
if (verify(vein, expected, schema)) {
  const layer = decode(vein, schema);
  console.log(`Using ${layer.defaultModel} from ${layer.name} layer`);
}

Per-Capability Resolution

The defaults map allows capability-specific model resolution:


function resolveModel(layer: ModelLayer, capability: string): string {
  return layer.defaults?.[capability] ?? layer.defaultModel;
}
 
resolveModel(previewLayer, "code");   // → "qwen3-coder:30b"
resolveModel(previewLayer, "vision"); // → "qwen3-vl:30b"
resolveModel(previewLayer, "chat");   // → "gpt-oss:20b" (default)