AI operations / AI platform teams
Governed model routing for security agents: provider choice without ambient authority
How to separate model capability from evidence access, tool authority, retention, caching, and cost attribution.
AI operations
Bound every specialist to the same verified context.
Agent work stays inspectable
New models are easy to add. Keeping the same evidence, retention, tool, cost, and approval boundary is the hard product work.
A model identifier is not an authorization policy
Model routing decides which capability handles a bounded task. It should not decide what workspace data can be retrieved, which tools can run, or whether an action can change state. Those controls belong to ordinary tenant-aware software around the model.
Hyperoru uses named reasoning tiers for routing, classification, specialist review, and explicit escalation. The current production identifiers are provider-pinned through LLMGateway so the requested route is visible in telemetry.
Keep context selection deterministic
Before a model call, the platform selects bounded evidence by workspace, repository, source, freshness, relationship, finding, and review question. It masks sensitive material, records a context hash, checks budget, and allows only narrow typed tools.
This means a future model or provider can be evaluated behind the same boundary. The model changes. The tenancy, evidence contract, output schema, cache isolation, and action rules do not.
Retention and cost are part of model quality
A useful model that requires unacceptable data retention is not a valid route for a confidential review. Hyperoru sends store false for responses and uses a gateway organisation configured for metadata-only retention. The platform records final structured artifacts and usage rather than hidden reasoning or tool transcripts.
Cost attribution should identify the workspace action, model, provider request, tokens, cache state, latency, and category. That makes quality and cost comparable at the audit, remediation, recommendation, and conversation level.
“Provider choice is safe only when the policy around the provider is stronger than the convenience of the route.”
Evaluate the wider catalogue deliberately
LLMGateway exposes a broad provider and model catalogue through one interface. Catalogue availability is not the same as Hyperoru support. A candidate route must pass task capability, structured output, tool behavior, retention, latency, reliability, regional, and cost tests before it can be enabled.