Trust levels

How Cailos classifies and enforces data handling guarantees across providers.

Every endpoint in Cailos is assigned a trust level from 0 (lowest) to 3 (highest). The level says what the provider does with what you send it: whether it may train on it, and how long it keeps it. Trust levels are a hard routing constraint: when you request a minimum trust level, your data is never sent to endpoints below that threshold.

The four levels

0 Low

The provider may use your prompts and outputs to train or improve its models, and may keep them for as long as it likes. Suitable for public or non-sensitive data only.

Typical providers: free tiers and consumer terms that allow training.

1 Standard

The provider's terms are ambiguous or incomplete. They don't say clearly enough how long data is kept or whether it can be used for training to rule either out. Treat it as unsuitable for confidential data until the terms are clarified.

Typical providers: smaller hosts and resellers without a published data policy.

2 High

The provider commits in writing to never training on your data, and deletes it within 30 days. It keeps it only for abuse and safety monitoring. Suitable for general and confidential business data.

Typical providers: major commercial APIs under their standard business terms.

3 Maximum

Zero data retention (ZDR). Prompts and outputs are processed and then discarded. Nothing is stored once the response is returned: not for abuse monitoring, not in logs. Nothing is ever used for training. Suitable for the most sensitive data, including health, financial and regulated data.

Typical providers: providers with a zero data retention agreement in place, and secure-enclave inference such as Tinfoil.

How routing works

Trust is enforced as a hard filter in the routing pipeline. When you include trust_level in your request, Cailos eliminates all endpoints below that threshold before any ranking or scoring occurs. This means:

Requesting trust_level: 2 routes only to endpoints with trust 2 or 3.
If no endpoint meets the threshold, the request fails with a clear error rather than falling back to a less trusted provider.
Trust filtering happens before cost, speed, or quality optimisation — it is never overridden by routing preferences.

Usage

Pass the trust_level hint in the cailos object of your request:

{
  "model": "auto",
  "messages": [{"role": "user", "content": "..."}],
  "cailos": {
    "trust_level": 2,
    "optimise": "quality"
  }
}

Tier restrictions

Free-tier teams are restricted to trust level 0 regardless of the requested level. Paid teams can request up to trust level 3. This is enforced server-side and cannot be bypassed by the client.

Assignment criteria

Trust levels are assigned by Cailos administrators based on provider documentation, certifications, and deployment model. The assignment considers:

Data retention policy. Does the provider store inputs/outputs? For how long?
Training policy. Are customer inputs used to improve models?
Certifications. SOC 2, ISO 27001, GDPR compliance, FedRAMP, etc.
Deployment model. Shared cloud, dedicated instance, on-premise, or secure enclave.
Jurisdiction. Where is data processed and stored?

Routing hints in the API docs Subprocessors