Research Flow and artifact lineage
Lineage answers two related questions: how did the research progress, and exactly which inputs and decisions produced a result?
Research Flow and Full Provenance
- Research Flow is the default semantic view, emphasizing experiments, data, models, evidence, claims, hypotheses, verification, and reports. Multiple technical outputs of one experiment are collapsed under its experiment node.
- Full Provenance displays the complete artifact graph, including source, logs, and intermediate outputs.
- Show Technical Provenance adds technical nodes to Research Flow without switching views.
Right-click an Experiment marked N artifacts to expand or collapse it; the toolbar can toggle all
experiment groups. Large graphs automatically use a faster layered layout and may hide edge labels
or the minimap to remain responsive.
Navigation and selection
The canvas supports pan, wheel zoom, Fit View, a minimap, and clearing selection by clicking the background. Nodes cannot be dragged or manually connected because layout and relationships come from Runtime's authoritative projection.
When a node is selected:
- purple paths highlight every upstream source;
- green paths highlight every downstream result;
- unrelated paths are muted;
- Lineage relationships lists direct source and downstream nodes for stepwise navigation.
Semantic node types include Configuration/Physics Input, Dataset, Source Code, Model, Predictions,
Evaluation, Analysis, Scientific Report, Evidence, Claim, Hypothesis, Verification Plan/Result,
Experiment, and Decision. Relations include PRODUCED, DERIVED_FROM, INFORMED_BY, and
GOVERNED_BY.
Inspect an artifact
For a node with an Artifact ID, Runtime verifies metadata and SHA-256 before displaying:
- full ID, type, digest, creation time, creator, and size;
- parents, children, related jobs, and metrics;
- sanitized JSON metadata;
- visible immutable content.
Python, JSON, YAML, text, logs, and Markdown open as text/code. CSV/TSV gets a table preview of up to 50 rows. Models, checkpoints, and unknown binary types show metadata and can export immutable bytes through the content link. The lightweight CSV preview does not fully parse quoted embedded commas; download the artifact for formal data processing.
Evaluator-private artifacts never enter the browser, Ask, timeline, or lineage. Guessing an Artifact ID cannot bypass that visibility boundary.
Decision nodes
A Decision is not a file. Its inspector shows the event-backed summary and timestamp. This lets a model or report trace back to the research action that influenced it without pretending the decision is file content.
Export the graph
The active view can be exported as:
- JSON — machine-readable nodes and edges;
- DOT — for Graphviz and compatible graph tooling;
- SVG — scalable output matching current collapse, selection, and theme state;
- PNG — convenient for documents and presentations.
Export reflects the current Research Flow or Full Provenance view, including collapsed groups. Desktop opens a save dialog; Web mode uses a browser download.
A practical audit sequence
- Identify the target report, metric, or final model in Control Room.
- Open Research Flow and select its report, evaluation, or model node.
- Follow purple upstream paths through data, source, experiments, and decisions.
- Check SHA-256 and parent IDs for every critical artifact.
- Follow green downstream paths to evaluations, claims, and reports that used it.
- Export SVG for archival viewing or JSON for programmatic analysis.
The graph is a read model, not a second scientific database. It is rebuilt from Runtime events and ArtifactStore when reopened.