Model Evaluation Service  ·  View 06 of 21  ·  Structure

Layered Architecture

Six layers, and the one rule about what may cross between the evidence layer and the decision layer.

Editable source SVG draw.io All views
Consumers Engineer CLI Verdict Page Rater Console Deployment Pipeline Decision Gate Engine Significance Waiver Register Evidence Deterministic Scorers Judge Service Calibration Monitor Human Review Execution Run Planner Harness Workers Tool Sandbox Corpus Dataset Store Artefact Store Score Store Label Store Platform GKE Pub/Sub Workload Identity Cloud Observability Model Evaluation Service — Layered Architecture Application we own External / third party Decision point Security / platform Interface / broker Data store Queue / topic The frame — dataset, scorer, judge and harness version — is resolved once at the execution layer and travels with every score to the decision layer. Nothing above the evidence band may compare across frames. v 1.0 · owner Data & AI Global Practice · date 2026-09

Decisions

  • Evidence and decision are separate layers. A gate policy change never requires a re-run; a re-run never silently changes a policy.
  • The frame is resolved in the execution layer and travels upward with every score. Nothing above the evidence band may compare across frames.

Why not fewer layers

  • Folding the decision layer into scoring is the common shortcut, and it is what makes thresholds impossible to change without re-measuring everything.

Note

  • This is the only view in the set that uses stacked bands. Every other view has the shape its own kind of question deserves.