UX-ML-001 — Present the actual ML lifecycle
| Field | Value |
|---|---|
| Status | Open product finding; source-verified |
| Severity | Medium |
| Components | sparqd/oqullus, platform documentation, sparqd/agent-based-ml-platform, MLflow, AI Gateway |
| Associated issues | Not opened pending finding review |
Concrete observed case
Experiments and Models are read-only browsers. AutoML does not consistently surface its results in either one. The Model Serving route is disabled, while prior documentation described Models as a promotion lifecycle and AI Gateway as the place to deploy trained models. AI Gateway actually routes external GenAI provider models and is only an AutoML reasoning dependency.
Expected condition
Navigation and documentation state what is usable now: AutoML training, read-only MLflow experiment inspection, and read-only registered-model inspection. Missing transitions and serving remain clearly unavailable until a supported lifecycle exists. AI Gateway remains a separate GenAI concept.
Why it matters
Users otherwise search for actions that do not exist or may mistake an external LLM endpoint for deployment of their trained artifact. This obscures ownership, lineage, rollout, observability, and rollback requirements.
Existing controls and workarounds
The serving route is not shown and unimplemented registry controls are mostly commented out, which avoids exposing dead actions. Documentation is corrected by this assessment, but the product still lacks explicit boundary copy and links.
Alternatives considered
- Recommended now: label the pages Experiment runs and Registered models or state their read-only status, and show unavailable next steps explicitly.
- Build registration and trained-model serving as separate gated lifecycle stages with stable lineage.
- Reuse AI Gateway — rejected unless it deliberately adds an independently designed ML artifact runtime; its current identity and controls model external GenAI providers.
Component involvement
- Oqullus owns navigation, labels, empty states, links, and available actions.
- AutoML and MLflow own authoritative lifecycle status and lineage.
- Documentation describes only verified behavior.
- AI Gateway remains a separately assessed subsystem.
Reproduction and verification
Follow ML-S010.
Resolution conditions
Close when users can identify current capabilities and unavailable stages without inference, and every enabled transition has an authoritative backend.
Validation and blockers
The current UI/source boundary is confirmed. Product must decide whether and when registration and trained-model serving enter scope.