How a causal run works

From a CSV file to a causal graph: upload, choose an algorithm, run inline or queued (with progress and cancel), read the graph, reason on it and export it.

The path in one picture

1. Upload a table of numbers on the Data tab.
2. Choose the dataset and an algorithm on the Discovery tab.
3. Run. A small table runs at once; a large one is queued. Either way a progress bar shows the step, and a queued run can be cancelled.
4. Read the graph: the variables are boxes, an arrow is a discovered cause-and-effect relationship with a weight and a confidence.
5. Reason on the graph: paths between two variables, what-if values, scenarios, recommendations.
6. Export the edges as a CSV file, or come back to the graph later from the Graphs tab.

The screenshots below were taken in this release with the Sachs protein-signalling data (853 measurements of 11 proteins, Sachs et al. 2005), uploaded as a CSV file.

Step 0: choose a mode

The page asks how you want to work every time you open it.

Choose Your Experience

• Executive: three steps (Data, Analyze, Insights), always the UCMT v2 Global algorithm, plainer wording.
• Technical: five tabs (Data, Discovery, Reasoning, XAI, Graphs), every algorithm and option.

The button at the top right of the page (Technical or Executive) brings the choice back. The rest of this guide follows Technical mode and notes where Executive differs.

Step 1: upload

On the Data tab choose a .csv file (a header row of variable names, then one row per observation), check the preview, name the dataset, pick a domain or General Purpose, and select Upload Dataset. Details: Data tab.

Step 2: choose an algorithm

On the Discovery tab pick the dataset under Select Dataset and an algorithm under Select Causal Discovery Algorithm. UCMT V3 is marked experimental: its internal benchmark figures have not been independently reproduced. Details: Discovery tab.

Step 3: run, follow, cancel

Select Run Causal Inference (Run Analysis in Executive mode). The button changes to Running Inference... and a progress bar names the step: loading the dataset, discovering the causal structure on the engine, checking the result against known ground truth, updating your saved graphs.

Inline or queued. A table of up to 50,000 cells (rows × variables) runs at once: the page waits for the answer, usually a few seconds.

An inline run in progress

A larger table is sent to the queue: a worker picks it up, and the page follows it until it ends. While it waits the bar says Waiting for a free engine worker, and a Cancel run button appears.

A queued run with its Cancel button (5,118 rows × 11 variables)

Cancel run stops the worker. The page then says The run was cancelled.

After cancelling

You can leave the page while a queued run works; it keeps running on the server and its graph appears under Graphs and Discovery runs when it is done. Your organization can have a limited number of runs queued or running at once; see Messages and errors for what the page says when that limit, a time budget or your plan stops a run.

Step 4: read the graph

When the run ends, the right-hand side shows the summary tiles, the graph and the relationship lists. The graph is saved at once. What each part means: Reading the result.

Step 5: reason on it

Open the Reasoning tab: it works on the graph you just ran (or one you load from Graphs). See Reasoning tab.

Step 6: export

• Export Edges (CSV) under Key Discovered Edges downloads the edge table: number, source, target, weight, confidence, time lag and effect type.
• The Discovery runs page exports the list of runs as a CSV file.
• Reasoning results can be saved in this browser with Save Reasoning; they are not stored on the server.

What the page does not show in this release

• The UCMT v2 Global and V3 engines run internal validation steps (the REF-001 pre-checks and four block gates). Their outcome for your run is shown in the Discovery quality panel under the result: each gate with pass/fail and its key figures, the REF-001 status and data regime, the phases and blocks executed, and per-edge confidence where the engine measured one ("not reported" otherwise). The Trust view of the Reasoning tab reads the same figures from the result (see Trust).
• The run uses the engine's own settings for each algorithm; the page sends no hyperparameters.