> ## Documentation Index
> Fetch the complete documentation index at: https://docs.neuraltrust.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Model resolution

> How the model field in a request selects a registry — short pass-through, provider-qualified, or pool reference — and how model policies constrain it.

A client never names a provider URL — it names a **model**, and TrustGate resolves which
[registry](/trustgate/concepts/registries) to use. The `model` field accepts three forms.

## Model reference syntax

| Form                     | Example          | Meaning                                                                                                                                             |
| ------------------------ | ---------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Short** (pass-through) | `gpt-4o-mini`    | Sent as-is to the selected registry. The [load balancer](/trustgate/routing/load-balancing) picks among the consumer's registries.                  |
| **Provider-qualified**   | `@openai/gpt-4o` | Restrict routing to registries of provider `openai`, then pick.                                                                                     |
| **Pool reference**       | `pool:my-pool`   | Route via the consumer's enabled load-balancing pool whose `pool_alias` matches (case-insensitive), across the pool's members. Inline routing only. |

## Model policies

[Consumers](/trustgate/concepts/consumers) and [roles](/trustgate/concepts/roles) can
constrain which models are reachable per registry with `model_policies`:

```json theme={null}
"model_policies": {
  "<registry_id>": {
    "allowed": ["gpt-4o", "gpt-4o-mini"],
    "default": "gpt-4o-mini"
  }
}
```

* **`allowed`** — the allow-list of models for that registry. Empty/omitted means all
  models are permitted.
* **`default`** — used when the request names no model; it must be a member of `allowed`.

A request for a model outside the allow-list is rejected before it reaches the provider.

## Resolution order

1. Parse the `model` reference (short / qualified / pool).
2. Narrow the candidate registries (the consumer's `registry_ids` or the pool members,
   filtered by provider for qualified refs).
3. Apply `model_policies` — reject disallowed models; fill in the `default` when none was
   given.
4. Hand the candidates to the [load balancer](/trustgate/routing/load-balancing) to pick
   one, with [fallback](/trustgate/routing/fallback) on failure.