How it works

A clear path from request to evidence

Kyverance coordinates connect → govern → route → observe → evaluate as an enterprise control plane. Claims below stay within confirmed MVP capability.

  1. Connect applications

    Inputs

    • Application identity and scoped API credentials
    • Request context (project, environment, caller metadata)
    • Organization-owned provider credential references

    What Kyverance coordinates

    • Tenant-isolated application onboarding
    • Credential issuance through approved flows
    • OpenAI-compatible request entry into the control plane

    Outputs

    • Authenticated service traffic into Kyverance
    • Consistent request context for downstream policy and routing

    Why it matters. Platform teams need a single governed entry point instead of ad-hoc provider keys in every service.

  2. Govern requests

    Inputs

    • Organization and project policy
    • Allowlists, quotas, budgets, and rate limits
    • Role and membership constraints

    What Kyverance coordinates

    • Policy evaluation before provider execution
    • Tenant isolation checks on every controlled record
    • Audit-relevant authorization decisions

    Outputs

    • Allow, constrain, or deny with an explainable reason
    • Budget and rate-limit enforcement signals

    Why it matters. Governance and security teams need control and evidence before models run—not after the fact.

  3. Route and execute

    Inputs

    • Approved provider and model inventory
    • Transparent static routing configuration
    • Manual override when operators need a deliberate path

    What Kyverance coordinates

    • Deterministic route selection (MVP)
    • Retries and failover within configured policy
    • Provider-independent adapter execution

    Outputs

    • Provider/model execution on an approved path
    • Routing explanation retained with the request

    Why it matters. Operators must see why a path was chosen. The MVP does not claim autonomous adaptive routing or invisible model switching.

  4. Observe

    Inputs

    • Request and response metadata (redacted of secrets)
    • Routing explanations and outcome events
    • Operational health signals

    What Kyverance coordinates

    • Tenant-scoped request telemetry
    • Analytics with honest live / empty / unavailable / sample labels
    • Audit records for security-relevant actions

    Outputs

    • Organization-visible usage, cost, and quality signals
    • Evidence suitable for review and incident response

    Why it matters. Without observability, routing and governance cannot be trusted or improved deliberately.

  5. Evaluate and optimize

    Inputs

    • Outcome events and evaluation runs
    • Human review of routing and quality
    • Policy and configuration changes

    What Kyverance coordinates

    • Structured evaluation workflows
    • Promotion and release gates where configured
    • Deliberate optimization based on evidence

    Outputs

    • Improved prompts, policies, and routing configuration
    • Clear change history for operators

    Why it matters. Optimization follows evidence. Reinforcement learning and broad autonomous adaptation are outside MVP scope.

MVP boundaries

Included: transparent static routing, manual override, auditability, tenant isolation, provider independence, quotas, budgets, rate limits, retries, failover, request telemetry, routing explanations, and outcome events. Not included: autonomous adaptive routing, reinforcement learning, invisible model switching, or broad compliance certification claims.