Routing Policies
policy-2026-07-14Hard placement constraints applied before any candidate is scored: regions, providers, egress trust tier, ZDR and data classification ceilings.
Route policies
A candidate endpoint must satisfy every column to survive hard filtering — rejections are recorded per candidate in the route plan.
| Policy | Allowed regions | Allowed providers | Min trust tier | ZDR | Data class ≤ | Profiles |
|---|---|---|---|---|---|---|
Default enterprise routing rp_default | ap-southeast-1us-east-1eu-west-1 | openaianthropicamazon-bedrockgoogle-vertexself-hosted-vllm | Tier C | optional | internal | 7 |
Standard + ZDR rp_standard_zdr | ap-southeast-1us-east-1eu-west-1 | openaianthropicamazon-bedrockgoogle-vertex | Tier C | required | confidential | 1 |
Confidential (SG, Tier A/B, ZDR) rp_confidential | ap-southeast-1 | anthropicamazon-bedrockself-hosted-vllm | Tier B | required | restricted | 1 |
Production code paths rp_code | ap-southeast-1us-east-1 | anthropicself-hosted-vllmopenai | Tier B | required | confidential | 1 |
Media generation rp_media | us-east-1eu-west-1 | openaigoogle-vertex | Tier C | optional | internal | 2 |
Effective policy per profile
What a request actually runs under once its service profile resolves route, guardrail, cache and resilience policy together.
@company/general-fastlatencyv4
@company/general-balancedbalancedv7
@company/general-bestqualityv3
@company/code-productionqualityv9
@company/confidentialbalancedv6
@company/visionbalancedv2
@company/embeddingscostv3
@company/voicelatencyv1
@company/imagebalancedv2
@company/videobalancedv1
@company/autobalancedv5
fabric/autobalancedv5
How enforcement layers compose
A request's effective constraints are the intersectionof its route policy, guardrail-issued route restrictions (e.g. a confidential-data finding forcing SG-region + Tier A/B + ZDR), the project's data-class ceiling and the profile's cost cap. Hard filters run before scoring, so a candidate that violates any layer never competes — it appears in the route plan under hardFilterRejections with the specific reason.