How it works

A single chat completion request travels through three checkpoints on the platform before you get a response.

  Your app  ──▶  Advania ALT API  ──▶  Guardrails  ──▶  Model  ──▶  Response + decision id

• API — authenticates the request, applies quotas, forwards to guardrails
• Guardrails — classify the request's sensitivity, route to a model that is approved for that content
• Model — runs on Advania-operated GPUs in Sweden (for sensitive content) or a supported cloud provider (for non-sensitive content)

The response body follows the OpenAI chat-completions shape. Every response also carries these headers you can log:

Header	Purpose
x-decision-id	Guardrails decision reference for support and audit
x-policy-version	The version of the guardrail policy that produced the decision
x-request-id	Unique per HTTP request

When the platform blocks a request

If the classifier decides a request cannot be served by any model available to you, the platform returns 403 Forbidden with an OpenAI-compatible error body:

{
  "error": {
    "message": "…",
    "type": "guardrail_error",
    "code": "…"
  }
}

The x-decision-id header is still present — share it with Advania if you need to investigate.

When guardrails cannot complete

If the guardrail evaluation cannot finish safely, the platform returns 503 — it does not silently forward the request. Retry later or contact Advania.