Why Scarf records provenance

A single-cell result is a chain of choices. A marker table depends on a clustering, the clustering depends on a graph, and the graph depends on exact cells, features, normalization, and reduction. When several branches share one datastore, a filename or cluster column cannot identify that chain.

Scarf therefore persists substantial results as immutable artifacts. Each artifact records the producing operation, scientific parameters, and exact upstream inputs. Matching provenance can reuse completed work; changing an input creates a distinct downstream branch.

        flowchart LR
    counts["Counts"]
    selection["Cell and feature selections"]
    norm["Normalization"]
    pca15["PCA, 15 dimensions"]
    pca30["PCA, 30 dimensions"]
    graph15["Neighbour graph"]
    graph30["Neighbour graph"]
    clusters["Clusters"]
    markers["Marker table"]
    counts --> selection --> norm
    norm --> pca15 --> graph15 --> clusters --> markers
    norm --> pca30 --> graph30
    

The branches share counts, selections, and normalization. Their reductions and downstream graphs remain distinct because those inputs differ. No branch becomes an implicit current result.

A PipelineRun adds one durable record for a complete workflow invocation. It binds named outputs and frozen cell and feature views without changing the artifact identity rules.

Provenance establishes computational relationships. It does not prove that a parameter was scientifically appropriate, decide which branch is best, or replace study records.

Use Provenance and reuse for the single executable guide to listing, inspecting, branching, invalidating, and comparing artifact lineage. Use Artifacts, lineage, and summaries API reference for the complete API contract.