Goal: Prove the migrated app behaves at least as well as the original — and catch regressions before they hit production — with reusable golden datasets, an evaluator SDK, and Azure AI Foundry result tracking.
🚧 Placeholder. Content lands here in Phase 2 (seed) and Phase 6 (reconciliation) of the consolidation rollout. See docs/internal/ for sequencing.
README.md— evaluation methodology, LLM-as-Judge patterns, Foundry result tracking.data/— 54 pre-built golden test cases spanning 7 scenarios (RAG, classification, tool calling, structured output, multi-turn, etc.).data/templates/— dataset starter templates for teams building their own.
src/evaluate/— theMigrationEvaluatorPython SDK.samples/rag_pipeline/— self-contained RAG demo with swappable model stages, dual-layer evaluation, A/B comparison, drift analysis.notebooks/—azure_openai_migration_technical.ipynb,azure_openai_evaluation_guide.ipynb.docs/—building-golden-datasets.md,evaluation-guide.md,migrating-multi-step-apps.md,cloud-eval-tracking-across-models.md.
Nightly evaluation runs land in ../.github/workflows/eval-on-schedule.yml, wired to publish results to Azure AI Foundry for trend tracking across migrations.
A green evaluation gate hands off to:
../05-rollout/— canary, observability, and rollback for the production cutover.