A record of what produced every result.

The run that produced the data, recorded as a run.

A Generation is a first-class record. It names the generator and its version, the parameters it was given, the seed it used, the asset versions it referenced and the code version it ran under. It exists from the moment the run starts, not after it succeeds.

generator.name

e.g. blender

generator.version

e.g. 4.2.0

parameters

the arguments the run was given

reproducibility.seed

the seed used

status

pending, running, completed, failed

experiment_id

the experiment it belongs to

DatasetVersion, not a folder name.

A logical dataset accumulates versions. Every generation, training run and evaluation refers to a specific DatasetVersion, so a record still means something after the dataset has grown, been filtered, or been regenerated.

dataset

the logical dataset

version

the specific version referenced

uri

where the data actually lives

checksum

content identity, where available

format

image, video, annotation set

created_by

the generation that produced it

Scene assets are versioned inputs too.

A generation can reference exact AssetVersions rather than an asset name. Replacing a model or a texture on disk no longer silently changes the meaning of every run that used it.

asset

the logical asset

version

the version referenced

referenced_by

the generations that used it

A model is downstream of specific data.

A TrainingRun records its configuration and the exact DatasetVersions it consumed. That single link is what turns a pile of checkpoints into a history you can reason about.

framework

e.g. PyTorch

model

e.g. YOLO

config

hyperparameters and training setup

inputs

the DatasetVersions consumed

Metrics keep their chain of custody.

An EvaluationResult belongs to a TrainingRun and, through it, to the dataset versions, the generation and the parameters upstream. A number in a table can always be traced back to the run that produced it.

metrics

e.g. mAP, precision, recall

training_run_id

the run evaluated

dataset_version

what it was evaluated against

Query the experiment history, not the filesystem.

Experiments are searchable by project, generator, time and recorded metadata — including the parameters on the generation itself.

Field-by-field differences between two experiments.

Two experiments aligned on their recorded fields, with the differences marked. It shows what changed; it does not claim to show why the result changed.

Comparison is an inspection tool. Attributing a result to a parameter remains a research judgement, and SynthGraph does not make it for you.

A manifest that includes what is missing.

The reproduction manifest gathers the recorded configuration, seeds, asset versions, dataset references and code version, and lists the dependencies that were external or unrecorded.

SynthGraph makes dependencies and known provenance explicit. It does not guarantee exact reproduction when external dependencies were not captured.

References to data, not copies of it.

Datasets are referenced by URI and, where available, by checksum. Large generated data can stay on the researcher’s filesystem, lab storage or object storage; SynthGraph stores the pointer and the provenance around it.

Reference documentation for the SDK and the data model.

Written documentation covering the client, the record types and the lifecycle rules is being produced alongside the SDK during the pilot.