data-science — parity with Microsoft Fabric Data Science / Azure ML¶

Source UI:

Microsoft Fabric Data Science experience — ML Model, ML Experiment, ML Job item types + MLflow + AI Functions + Semantic Link (https://learn.microsoft.com/fabric/data-science/).
Azure ML Studio — Models (https://ml.azure.com/model/list), Jobs / Experiments, and the MLflow-compatible tracking server (https://learn.microsoft.com/azure/machine-learning/how-to-track-experiments-mlflow).

Backend (Azure-native default — no Microsoft Fabric / Power BI workspace required):

Model registry + online endpoints: Microsoft.MachineLearningServices/workspaces/{ws}/models[/{name}/versions] and /onlineEndpoints via ARM REST (api-version 2024-10-01) — lib/azure/foundry-client.ts.
Jobs / experiments: Microsoft.MachineLearningServices/workspaces/{ws}/jobs — foundry-client.ts.
MLflow experiment tracking + per-step metrics: AML's MLflow-compatible tracking server https://{region}.api.azureml.ms/mlflow/v1.0/.../workspaces/{ws}/api/2.0/mlflow/* — lib/azure/mlflow-client.ts.
AI Functions: Azure OpenAI chat completions (LOOM_AOAI_ENDPOINT + LOOM_AOAI_DEPLOYMENT) — lib/azure/ai-functions-client.ts.
ML Model install provisioner: Databricks workspace/import + jobs/runs/submit training run that logs + registers the model in Unity Catalog / MLflow — lib/install/provisioners/ml-model.ts.

This experience runs entirely on Azure-native backends. Fabric MLflow / a Fabric Data Science workspace is never required; LOOM_DEFAULT_FABRIC_WORKSPACE stays unset on the default path (.claude/rules/no-fabric-dependency.md).

Fabric / Azure ML feature inventory → Loom coverage¶

Capability (Fabric Data Science / Azure ML Studio)	Loom coverage	Backend (real REST / data-plane)
ML Model — pick the AML workspace the model lives in	✅ built — bind picker workspace Dropdown	`listMlWorkspaces()` → `GET .../resourceGroups/{rg}/.../workspaces`
ML Model — browse registered models in that workspace	✅ built — bind picker model Dropdown (latest version)	`listModels(ws)` → `GET .../workspaces/{ws}/models`
ML Model — bind the Loom item to a model	✅ built — Bind button persists to Cosmos `state`	`POST /api/items/ml-model/[id]/bind` → `persistModelBinding`
ML Model — overview (name, description, latest version)	✅ built — Detail tab	`GET /api/items/ml-model/[id]` → `getModel(name, ws)`
ML Model — version list (type, created, URI)	✅ built — Versions tab + left-panel tree	`listModelVersions(name, ws)`
ML Model — MLflow flavors / signature / tags / lineage	✅ built — Detail tab renders `properties.flavors` / `tags` / run badges	version `properties`
ML Model — register a new model version	✅ built — Register-version dialog (URI + version + type)	`POST /api/items/ml-model/[id]/register` → `registerModelVersion` (PUT `.../models/{name}/versions/{ver}`)
ML Model — deploy managed online (real-time) endpoint	✅ built — Deploy tab (VM size + Deploy)	`POST /api/items/ml-model/[id]/endpoint` → `createOnlineEndpoint` + `createOnlineDeployment`
ML Model — list existing online endpoints	✅ built — Deploy tab endpoints table	`GET /api/items/ml-model/[id]/endpoint` → `listOnlineEndpoints(ws)`
ML Experiment — list experiments / job runs	✅ built — experiment grouping list (rollup by `experimentName`)	`GET /api/items/ml-experiment` → `listJobs()`
ML Experiment — experiment / job detail	✅ built — Overview tab	`GET /api/items/ml-experiment/[id]` → `getJob(id)` or filtered `listJobs()`
ML Experiment — MLflow runs + per-step metric history	✅ built — "Runs & metrics" tab (run table + metric charts)	`GET /api/items/ml-experiment/[id]/runs` → `searchRuns()`; `.../runs/[runId]/metrics` → `getMetricHistory()`
ML Experiment — submit a new job run	✅ built — Submit dialog	`POST /api/items/ml-experiment/submit` → `foundry-client.ts` job create
ML Experiment — register a run's output as a model	✅ built — Register action	`POST /api/items/ml-experiment/[id]/register`
AI Functions — `summarize`	✅ built	`POST /api/ai-functions` (`fn=summarize`) → AOAI chat completions
AI Functions — `classify`	✅ built	`POST /api/ai-functions` (`fn=classify`, `options.labels`)
AI Functions — `sentiment`	✅ built	`POST /api/ai-functions` (`fn=sentiment`)
AI Functions — `extract`	✅ built	`POST /api/ai-functions` (`fn=extract`, `options.fields`)
AI Functions — `translate`	✅ built	`POST /api/ai-functions` (`fn=translate`, `options.targetLang`)
"Prep for AI" — semantic-model AI annotations consumed by Data Agents	✅ built — Semantic Model designer (per-table / per-column annotations)	`GET/POST /api/items/semantic-model/[id]` (Cosmos)
ML Model install — train + register a model from a use-case app	✅ built — install provisioner imports a Databricks training notebook + submits a run that registers in UC/MLflow	`lib/install/provisioners/ml-model.ts` → `databricks-client.ts`
ML Model — infra gate when AML not provisioned / no RBAC	⚠️ honest-gate — bind picker `workspacesError` / `modelsError` MessageBar names `LOOM_SUBSCRIPTION_ID` + `LOOM_FOUNDRY_RG` + the AzureML Data Scientist role; full UI still renders	n/a
ML Experiment — MLflow infra gate (workspace/region unresolvable)	⚠️ honest-gate — "Runs & metrics" MessageBar names `LOOM_AML_WORKSPACE` + `LOOM_AML_REGION` (falls back to `LOOM_FOUNDRY_NAME` / `LOOM_FOUNDRY_REGION`) + AzureML Data Scientist (`MlflowNotConfiguredError`)	n/a
AI Functions — gate when no AOAI model deployed	⚠️ honest-gate — `POST /api/ai-functions` returns `501 {code:'not_configured', missing:'LOOM_AOAI_DEPLOYMENT'}` with deploy hint	`NoAoaiDeploymentError`
SynapseML (GPT/Cognitive transforms on Spark)	⚠️ honest-gate — no dedicated pane; `pip install synapseml` works in any Databricks / Synapse Spark notebook (documented, no SaaS gap)	n/a
Semantic Link / `semantic-link-labs` (read Power BI models from a notebook)	⚠️ honest-gate — no dedicated pane; `pip install semantic-link-labs` reads models via XMLA from any notebook (documented)	n/a
MLflow tracking host in sovereign clouds (GCC-High / IL5)	⚠️ honest-gate — `mlflow-client.ts` host is `*.api.azureml.ms`; the `.us` sovereign suffix is not yet parameterized (tracked: `LOOM_AML_HOST_SUFFIX`). Surfaces as the same `MlflowNotConfiguredError` MessageBar	n/a

Zero ❌. Zero stub banners. Every row is built (✅) or an honest infra-gate (⚠️, allowed per .claude/rules/no-vaporware.md).

Backend per control¶

Model bind/list: GET|POST /api/items/ml-model/[id]/bind → listMlWorkspaces() + listModels(ws) + persistModelBinding.
Model + versions: GET /api/items/ml-model/[id] → resolveModelBinding → getModel(name, ws) + listModelVersions(name, ws).
Register version: POST /api/items/ml-model/[id]/register → registerModelVersion(name, {modelUri, version?, modelType, workspaceName}).
Deploy / list endpoints: GET|POST /api/items/ml-model/[id]/endpoint → listOnlineEndpoints(ws) / createOnlineEndpoint(ws) + createOnlineDeployment(ws).
Experiments / jobs: GET /api/items/ml-experiment → listJobs(); GET /api/items/ml-experiment/[id] → getJob(id).
MLflow runs + metrics: GET /api/items/ml-experiment/[id]/runs → mlflow-client.ts searchRuns(); GET /api/items/ml-experiment/[id]/runs/[runId]/metrics → getMetricHistory().
Submit / register run: POST /api/items/ml-experiment/submit / POST /api/items/ml-experiment/[id]/register.
AI Functions: POST /api/ai-functions → callAiFn() → AOAI chat completions.

Env / RBAC¶

Env var	Backs	Default / fallback
`LOOM_SUBSCRIPTION_ID`	subscription holding the AML workspaces (required)	—
`LOOM_FOUNDRY_RG`	RG scanned for AML workspaces	`rg-csa-loom-admin-eastus2`
`LOOM_FOUNDRY_NAME`	hub workspace used when a binding has no `workspaceName`; MLflow workspace fallback	`aifoundry-csa-loom-<region>`
`LOOM_FOUNDRY_REGION`	endpoint/deployment region; MLflow region fallback	`eastus2`
`LOOM_AML_WORKSPACE`	MLflow tracking workspace ("Runs & metrics" tab)	falls back to `LOOM_FOUNDRY_NAME`
`LOOM_AML_RG`	RG of the MLflow workspace	falls back to `LOOM_FOUNDRY_RG`
`LOOM_AOAI_ENDPOINT` + `LOOM_AOAI_DEPLOYMENT`	AI Functions chat model	empty → 501 honest gate

RBAC: the Console UAMI (LOOM_UAMI_CLIENT_ID) must hold AzureML Data Scientist (f6c7c914-8db3-469d-8ca1-694a8f32e121) on the AML workspace — greenfield deploys get it via ai-foundry.bicep (hubConsoleDataScientist); BYO / deploy-planner workspaces grant it per v3-tenant-bootstrap §AzureML Data Scientist.

Bicep sync¶

AzureML Data Scientist grant (Console UAMI → Foundry hub workspace): platform/fiab/bicep/modules/admin-plane/ai-foundry.bicep (hubConsoleDataScientist).
AOAI model + LOOM_AOAI_* for AI Functions: platform/fiab/bicep/modules/ai/foundry-project.bicep (gated by agentFoundryEnabled; set param agentFoundryEnabled = true in commercial-full.bicepparam).
LOOM_AML_WORKSPACE / LOOM_AML_RG env wiring + agentFoundryEnabled threading: platform/fiab/bicep/modules/admin-plane/main.bicep + platform/fiab/bicep/main.bicep.

Per-boundary behavior¶

Boundary	ml-model	ml-experiment / MLflow	AI Functions
Commercial	✅ Foundry hub (AML Hub workspace, ARM REST)	✅ AML MLflow `eastus2.api.azureml.ms`	✅ AOAI via `agentFoundryEnabled`
GCC	✅ same ARM/AML REST	✅ same as Commercial	✅ AOAI via `agentFoundryEnabled`
GCC-High	✅ classic AML Hub (`kind=Default`, Foundry portal off)	⚠️ MLflow host suffix gap (`.us` not yet parameterized — `LOOM_AML_HOST_SUFFIX` tracked)	⚠️ AOAI in `usgov*` regions only; else 501 gate
IL5	✅ classic AML Hub	⚠️ same MLflow host-suffix gap; OSS MLflow on AKS is the alt server	⚠️ AOAI not yet IL5-authorized → 501 gate (documented)

Validation¶

Backend contract tests (Vitest): lib/azure/__tests__/aml-model-rest-shapes.test.ts (ARM URL shapes per named workspace), lib/editors/__tests__/ml-model-bff-routes.test.ts (all 5 model routes wired), lib/azure/__tests__/ai-functions-client.test.ts (5 fn prompts + 501 gate), lib/azure/__tests__/data-science-parity.test.ts (parity doc has zero ❌ / stub rows).
Acceptance: clean teardown + az deployment sub create -f platform/fiab/bicep/main.bicep -p params/commercial-full.bicepparam + the post-deploy bootstrap yields a working ml-model + ml-experiment + AI Functions experience with LOOM_DEFAULT_FABRIC_WORKSPACE unset — see data-science workload.