Create and test compatibility policies

Convert product requirements into versioned rules so model selection is reviewable, repeatable, and safer than an ad hoc shortlist.

Why use policies

A policy separates eligibility from preference. Regions, capabilities, context, lifecycle, and maximum cost become explicit constraints; only candidates that satisfy them can be ranked.

Create a policy version

1. Give the policy a stable name, such as EU production.
2. Enter valid JSON requirements.
3. Select Create policy version. Reusing the same name creates the next version instead of overwriting history.

{
  "providers": ["openai", "anthropic"],
  "regions": ["eu"],
  "capabilities": ["tool_calling", "structured_output"],
  "lifecycle_statuses": ["available"],
  "min_context_tokens": 128000,
  "max_total_usd": 0.02
}

Test the policy

1. Copy the policy UUID from Live data.
2. Enter a representative input and output token profile.
3. Select Test stored policy.
4. Review the eligible deployments and their calculated costs in Latest result.

Interpret results

Results contain only deployments that satisfy every hard constraint, have pricing, and stay below max_total_usd when set. Candidates are ordered by total quoted cost. An empty array is a meaningful failure: loosen a requirement deliberately or add eligible deployments—do not bypass the policy silently.

Governance pattern

• Use stable names and treat every saved version as immutable decision history.
• Test with several workload profiles, not only the default token counts.
• Reference the approved policy when creating a lockfile.
• Re-evaluate after lifecycle, pricing, or capability evidence changes.