AI Model Providers

How Guard picks a model from a capability tier, what disabling a vendor actually does, and the flagship-provider rule.

AI Model Providers

Guard features do not pick a concrete model ID. They ask for a capability tier. The platform then chooses a model in that tier from the vendors this tenant allows.

The substitute for a disabled vendor is same-tier, not whatever is cheapest and not the same model from another host.

Tiers

Tier	Role
Frontier	Highest-capability reasoning (interactive agents, compaction, campaign drafting)
Standard	Default agent workhorse
Fast	Light, low-latency classification
Lite	High-volume, low-stakes classification
Embedding	Vector embeddings. Not substituted; changing the embedding model invalidates stored vectors.

Experimental models exist for evaluation only and are not a tenant default.

Default policy

If this tenant has never saved a preference, Guard enables Anthropic, OpenAI, and Google, in that platform order. That is not every vendor.

The summary chips always show Praetorian first. Praetorian is platform routing, not a third-party vendor you can turn off. The saved policy never includes praetorian.

Vendors you can enable or disable: Anthropic, OpenAI, Meta, Google, Mistral, Amazon, DeepSeek, Zhipu, Moonshot, Qwen.

What happens when you disable a vendor

A feature requests a tier (or a specific model that belongs to a vendor).
If that vendor is still enabled, Guard uses it.
If that vendor is disabled, Guard walks the remaining enabled vendors in platform priority order and takes the first model that exists in the same tier.
If no enabled vendor offers that tier, the request fails.

Untiered allowlisted models (some Constantine pipeline models) have no tier substitute. Disabling their vendor means those IDs cannot be used.

Embedding requests ignore this opt-out chain and keep the embedding catalog.

Rules the editor will not let you break

Enable at least one vendor.
Enable at least one flagship vendor: Anthropic or OpenAI. A policy with neither is rejected.

You choose membership, not order. Saved lists are rewritten into platform order.

Reset to default

Restore defaults deletes the stored preference instead of writing the current default list. A deleted policy tracks new vendors Guard adds later. A frozen full list would not.

Who can edit

A super-admin or a user with manage settings. Others see Edit locked.

What this does not control

It does not turn AI Inference, Agents, or Autonomy on or off. It does not change monthly or per-scan spend. It applies to inference that goes through the model-tier registry once those features are allowed to run.