Make every experiment explainable after it runs.

Make every experiment explainable after it runs.

Reproducibility fails in a specific way: the configuration existed, and then it stopped existing. A parameter lived in a notebook cell that was re-run, a seed was never printed, an asset was overwritten. SynthGraph writes those things down at the moment they are still true.

Recording provenance does not make an experiment reproducible on its own. It makes the gap between what you have and what you would need explicit — which is the part that is usually invisible.

A result is only as legible as its chain.

An evaluation number in a paper table stands on a training run, which stands on dataset versions, which stand on a generation with parameters, assets and a seed. Lineage keeps that chain intact so a result can be defended, questioned or revisited.

Compare runs on what actually differed.

Two sweeps that look identical in a folder listing may differ in one parameter and one dataset version. Comparison aligns their recorded fields so the difference is visible instead of remembered.

The dataset you trained on, not the one that exists now.

Datasets are regenerated, filtered and extended over the life of a project. Recording the DatasetVersion a run consumed keeps months-old results meaningful after the dataset has moved on.

Scene content is part of the experiment.

In synthetic data the assets are inputs. A generation that referenced vehicle-model:v2 is a different experiment from one that referenced v7, even when every other parameter matches.

Material for a reproducibility supplement.

A reproduction manifest collects the recorded configuration, references and known gaps in one place — the same information a supplement, an appendix or a reviewer’s question tends to ask for.

Experiments outlive the people who ran them.

Students graduate, postdocs move on, machines are reimaged. A lab that records provenance keeps the reasoning behind its experiments after the person who held it in their head has left.