Model Evaluation Service  ·  View 10 of 21  ·  Data

Data Architecture and Ownership

Four zones with genuinely different rules — and the one place the immutability claim is deliberately broken.

Editable source SVG draw.io All views
Immutable · the reason old scores still mean something Corpus Dataset Versions GCS · indefinite Contamination Canaries never published Evidence Run Traces GCS · 400 d Judge Rationales GCS · 400 d Facts Scores BigQuery · 5 y Verdicts BigQuery · 7 y Ground truth · survives any re-scoring of the platform's own output Owned by the rater programme Human Labels Cloud SQL · indefinite Adjudications Cloud SQL Mutable · operational, rebuildable Run state Run + Checkpoint State Cloud SQL · RPO 5 min Candidate Register Cloud SQL Derived Score Cache Memorystore · disposable Report Views BigQuery · rebuildable Source of truth for policy · outside every database Reviewed as code Suites · Gates · Rubrics Git · RPO 0 Slice Taxonomy Git Model Evaluation Service — Data Architecture and Ownership Nothing in the immutable zone is ever updated. A correction is a superseding record; a dataset fix is a new version; a re-score is a new frame. Deletion requests are the one exception and are handled as a tombstone, which is why the reproducibility claim is bounded. v 1.0 · owner Data & AI Global Practice · date 2026-09

Decisions

  • Immutability is the comparability mechanism, not data hygiene. A correction is a superseding record; a dataset fix is a new version.
  • Policy lives in Git, outside every database, so the source of truth for a gate is reviewable and cannot be edited in place.
  • Human labels sit in their own project and their own zone because they are what the judge is measured against.

Retention

  • Verdicts 7 y · scores 5 y · traces 400 d (3 y for gate-blocking runs) · dataset versions and labels indefinite (stated assumptions).

Known tension

  • A user deletion request tombstones an example: content removed, identity and slice labels retained. Aggregate comparability survives; full reproducibility from artefacts does not. This is Core Architecture Question 7, and it is open.