Causal inference: Reasoning tab

Reason on a discovered graph: causal paths between two variables, a counterfactual value, scenario comparison, the trust view and the engine's executive decision summary.

What the tab is for

The Reasoning tab answers questions about one causal graph: the one you have just discovered, or one loaded from the Graphs tab with Use for Reasoning. Without a graph it says No causal graph to reason on and offers Run discovery and Load a saved graph. The strip at the top names the graph, its size and the algorithm; Load saved paths appears when this browser holds a saved path search for it.

Technical mode has four views: Paths, Counterfactual, Scenarios and Trust. Executive mode (the Insights step) has three, named for non-specialists: What-If, Scenarios and Confidence. There is no Recommendations view in this release: its figures had no model behind them, and it returns when one exists.

Every view has the same shape, top to bottom: an intro (the question it answers, what to enter, what you get, and an example in the graph's own variable names), the inputs, the run button named for what it runs, the result, and a How to read this expander that says what each figure is and is not. A figure the run did not measure reads not reported; nothing on the tab is estimated in your browser.

All views reason on the graph's structure and weights. They are as good as the graph: a wrong or missing edge changes every answer below.

Paths

Does a change in one variable reach another, and along which route?

1. Pick the Source (cause) and the Target (effect).
2. Optionally set the Longest route (1 for direct edges only, up to 10 hops) and the Smallest route effect (hide weak routes), and turn on Also search target → source to look both ways.
3. Select Find paths.

Paths, PKA to Erk on the Sachs graph

The tiles show how many routes were found, the strongest route's effect (the weights along it multiplied, in the graph's stored units), the number of direct routes and the longest one. Routes lists them, strongest first; open one for its steps with each edge's weight and measured confidence. Mediators shows which variables the routes pass through; Combined effect adds the route effects up and says what that sum is not (a counterfactual estimate). Recent keeps your last pairs. Save in this browser stores the result on this device, not on the server.

The search is a plain search over the stored edges, done in your browser: it adds nothing to the graph and estimates nothing. If no route is found, the result says why (no outgoing edges from the source, no incoming edges at the target, or a route that exists only the other way round).

Counterfactual

If one variable had been different, what would the outcome have been?

1. Variable to change: the variable you set. Value as it was and Value instead start at its observed mean; change the second.
2. Outcome to predict, then Run counterfactual.

Counterfactual, PKA on Erk

The tiles show the outcome as it was, as it would be, the difference in the outcome's own units, and the confidence. Two chips say how the figure was produced: Fitted to the dataset's rows when the dataset selected on the Discovery tab holds the graph's variables (the engine fits the estimate to them), otherwise Propagated along the stored weights; and whether the estimate was stable under resampling, which the engine reports only with a measured confidence. The route the change travels and the engine's sentence follow.

Scenarios

Which of several interventions moves the outcome most?

1. Choose the Outcome to measure.
2. Add scenario for each case. Each card lists the first five variables of the graph; enter the values to set (an empty field keeps the variable at its mean, shown as the hint).
3. Predict 2 scenarios (the button counts them). Each card shows its predicted outcome, and the table under them puts the scenarios side by side with the largest driver of each prediction and the engine's confidence when it reports one.

Scenarios, two values of Raf

Trust

Trust: the run's gates

Trust (Executive: Confidence) shows the validation the engine ran with the discovery itself: the gates of the last iteration with their metrics, the REF-001 readiness figures (status, readiness, validated and rejected candidates, reliability, data quality, calibration, data regime), the per-edge confidences and the run's reasoning steps. It does not run the engine a second time.

Trust (Sachs data)

Keep in mind:

• The figures belong to the run that produced the graph on screen. A graph made with a classical algorithm (PC, GES, FCI, ...) carries no gates and no REF-001 figures, and each missing value reads not reported.
• Final gate passed refers to the last of the engine's checks (Gate 3); All gates passed is shown beside it and is true only when every gate passed.
• A passed gate means the run met its own bar; it is not proof of causation for any single edge.

The executive decision summary

Under every view, Executive decision summary is written by the engine from the run's facts: select Generate summary. Before that the card holds only the button and a line on what the engine reads.

Executive decision summary, before Generate summary

Once generated, the card shows, in this order:

• Four chips for the four things the product knows about this graph: Engine checks (passed / not passed / not checked), Confidence (measured / not reported), Consensus (run / not run) and Acyclicity (acyclic / contains a cycle / not checked). Only a check that passed or did not pass is coloured; everything unmeasured stays neutral.
• The verdict, in the same words as the Discovery tab: Engine checks passed, Engine checks not passed or Not checked by the engine. It rests on the engine's gates and REF-001 and nothing else. Why no decision? opens a list of exactly the facts that stand in the way, each with what would change it.
• One paragraph of at most three sentences: the strongest stored edge into the outcome (your target when you chose one, otherwise the variable the graph points to most), whether a confidence was measured, and whether a consensus was run.
• Key edges into the outcome, heaviest first, with the stored weight, the confidence and the direction; the first is the primary driver. A self-loop is never a driver.
• What this means and What to do next, two short lists that follow from the facts.
• One line on what the summary was built from (whether a counterfactual fitted to the dataset's rows is included) and, when the checks did not pass, that exports carry the verdict.

Executive decision summary, generated by the engine

If the engine cannot generate the summary, the card shows the engine's message and a Try again button, and nothing else: no summary is composed in your browser.

CEO Brief downloads the summary as one sheet in the style of the graph export (the wordmark, the graph's name, the four facts, the verdict and paragraph, the key edges table, the two lists, a legend): Print or save as PDF, Download PNG or Download SVG. It is built from the engine's response only.

Executive mode

The Insights step opens on What-If (the counterfactual); Confidence is the trust view. The controls and results are the ones described above, with plainer labels and one-decimal figures.

Executive mode, Insights

Common messages