# ADR-0001: Separate Raw Assets From Canonical Datasets

## Status

Accepted

## Context

EmboFlow must support both structured embodied dataset formats and unstructured or semi-structured delivery-style raw assets, including:

- RLDS
- LeRobot v2/v3
- HDF5
- Rosbag
- Raw video directories
- Archive packages

If the platform treats every input as an already-standardized dataset, ingestion and delivery workflows become awkward and lossy.

## Decision

The platform will model:

- Raw assets as first-class resources
- Canonical datasets as derived semantic resources

Raw assets preserve original structure, paths, naming, and metadata layout. Canonical datasets provide normalized semantics for conversion, workflow execution, and export logic.

## Consequences

### Positive

- Supports customer delivery package workflows
- Supports embodied dataset conversion workflows
- Preserves original structure for inspection and debugging
- Avoids forcing visualization to depend on a lossy normalized format

### Negative

- Adds one more layer to the object model
- Requires readers and mappers instead of direct format-to-format conversion

## Notes

Visualization may operate on raw assets directly. Processing and export should primarily operate on canonical semantics where possible.