Eight steps, one record.

Add the SDK to the code you already run. The generator, the training script and the evaluation stay where they are; the calls around them record what happened.

• client.projects.create
• client.experiments.create
• code version captured

When a generation runs, the generator and its version, the parameters, the seed and the referenced asset versions are written down as a record — before the result exists.

• generator: blender 4.2.0
• seed: 42
• occlusion: 0.30

The output is registered as a DatasetVersion: a specific version of a logical dataset, identified by reference rather than by folder name. The data itself does not move.

• rain_dataset:v7
• s3://research/rain-v7
• 24,000 frames

A training run records its configuration and the exact dataset versions it consumed, so the link between the model and the data that produced it is explicit.

• YOLO · PyTorch
• 50 epochs
• inputs: rain_dataset:v7

Evaluation results attach to the run that produced them. A metric is never a loose number — it carries its whole upstream chain with it.

• mAP 0.724
• precision 0.78
• recall 0.69

From any record you can walk the graph in either direction: back from a metric to the parameters that produced it, or forward from a generation to everything that used it.

• upstream: parameters, assets, code
• downstream: runs, results

Two experiments align field by field. The differences in configuration and the differences in result are visible together — which is not the same as one explaining the other.

• occlusion 0.30 → 0.45
• dataset v7 → v8
• mAP 0.724 → 0.748

A reproduction manifest gathers what was recorded and states what was not. Missing external dependencies are listed rather than assumed away.

• recorded configuration
• external references
• known gaps

{  "generation": {    "generator": "blender",    "version": "4.2.0",    "seed": 42,    "parameters": {      "weather": "rain",      "occlusion": 0.30,      "camera_distance": 12    }  },  "assets": [    { "name": "vehicle-model", "version": "v7" },    { "name": "forest-texture", "version": "v4" }  ],  "dataset": {    "version": "v7",    "uri": "s3://research/rain-v7",    "checksum": "sha256:9c1f…e40b"  },  "code": {    "commit": "8f3a2b1",    "branch": "rain-sweep"  },  "missing": [    "external CUDA environment",    "GPU driver version"  ]}