Model Evaluation Service  ·  View 08 of 21  ·  Structure

Integration Architecture

Everything that calls in, everything called out, and which of them the platform must answer synchronously.

Editable source SVG draw.io All views
Inbound Engineer CLI Deployment Pipeline Policy Repository Production Telemetry Rater Console Platform Evaluation API 9 endpoints Run + Gate Core Event Spine 6 topics Outbound Vertex AI Provider APIs submit verdict webhook traces labels candidates judges Model Evaluation Service — Integration Architecture Application we own External / third party Interface / broker Queue / topic synchronous event / async batch The verdict query is the only synchronous integration the platform must honour during a release window; everything else degrades to slower. Slice reporting is a read of the score store, and chat and pager delivery is a transport on the event spine — neither is drawn as an integration. v 1.0 · owner Data & AI Global Practice · date 2026-09

Decisions

  • The verdict query is the only synchronous obligation. Everything else may degrade to slower without stopping a release.
  • Rater labels arrive inbound and asynchronously; human capacity is never on a run's critical path.

Contract

  • Verdict API returns PASS, BLOCK or INSUFFICIENT_EVIDENCE. The pipeline treats the third as a stop, not a retry-until-pass.

Deliberate omission

  • Slice reporting is a read of the score store, and chat and pager delivery is a transport on the event spine — neither is an integration.