Unified Data Model & Traceability

Canonical entity graph

Every meaningful lifecycle object has a first-class place in the graph instead of living in component-local metadata islands.

• Datasets, episodes, sensor streams, derived assets, and annotations
• Training runs, checkpoints, promoted models, and artifacts
• Deployments, telemetry, maintenance records, and intervention events

Lineage and promotion boundaries

The graph has to explain not only what exists, but how it came to exist and when mutable work became immutable promoted state.

• Upstream and downstream relationship traversal
• Immutable dataset versions, models, and release artifacts
• Audit-friendly provenance across approvals and workflows

Governance and ownership context

The platform needs explicit ownership and scope before it can do anything trustworthy across multiple teams or deployments.

• Organization, workspace, region, and environment ancestry
• Policy and audit-linked state where required
• Actor attribution for human, service, and robot actions

Data foundation entities

Robot data is more than a file upload. It is structured collection context, synchronized sensor streams, semantic labels, and quality decisions.

• Dataset, episode, sensor stream, and derived asset
• Annotation, annotation project, curation ruleset, and quality score
• Versioned dataset outputs used downstream by model development

Model-development entities

Training lineage is only credible when every promoted result is connected to the exact data and evaluation context that produced it.

• Experiment, training run, and checkpoint
• Evaluation outputs and promotion evidence
• Promoted model identity with parent-child lineage

Release and operations entities

The same graph needs to explain what was packaged, what was deployed, what happened in the field, and how those outcomes fed future decisions.

• Artifact, deployment, deployment target, and rollout history
• Telemetry event, maintenance record, and intervention event
• Robot, fleet, site, and hardware context for field behavior

From raw session to deployed model

Robot -> dataset -> episode -> annotation and quality decisions -> dataset version -> training run -> model -> artifact -> deployment -> live telemetry

From field failure back to source evidence

Incident -> deployment -> artifact -> model -> training run -> dataset version -> selected episodes -> collection context

From intervention to next model improvement

Intervention -> correction capture -> review -> curation ruleset -> updated dataset version -> retraining -> evaluation -> promoted model

• Which field incidents trace back to dataset version v12?
• Which operator sessions influenced the current model at one site?
• Which benchmark cases were added because intervention frequency rose?