What building means
Most modeling work the field calls "model building" is parameter fitting on a structure that was assumed. Causal modeling reverses that: the structure is the deliverable, the parameters are the tail end. The work is in the elicitation, the disagreement reconciliation, and the explicit declaration of what the model does not cover.
These three pages cover the building process from three angles — from the artifact that often starts the engagement (a risk register), through the method itself, to the part that turns out to be hardest in practice (elicitation).
Pages in this cluster
From Register to Graph → You have a risk register. It cannot answer the question on the table.
A risk register is an inventory built around the wrong question. It records what might go wrong and how bad it would be. It does not model how risks connect, which failures cause which others, or what happens to the portfolio if you intervene on one variable. Every board question that matters — should we renew this portfolio, should we replace this asset, should we extend this credit — requires the causal structure the register does not contain. This page is the practical walkthrough from flat table to working graph.
How to Build One → What would have happened had we acted differently?
This is a step-by-step walkthrough, not a conceptual overview. It covers the same three-step counterfactual procedure as the interactive Rung 3 page — use this version if you want the written derivation rather than the live demo. Your data tells you what happened. It cannot tell you what would have happened. That gap — between observation and intervention — is where causal models operate.
Why Elicitation Is Hard → Causal AI rests on expert knowledge. Getting it out is harder than everything else combined.
A causal model encodes mechanism knowledge data alone cannot recover. That mechanism knowledge lives in expert heads, and getting it out is the work the modeling literature spends the least time on and gets wrong the most. Probability elicitation has been studied formally for fifty years; the findings are unflattering. This page names the difficulties honestly and describes the practitioner techniques that handle them.
Reconciling Expert DAGs → Three experts asked to draw the same DAG produce three different DAGs. The disagreements are the work.
The cardinal mistake in group elicitation is letting the experts draw together — the first DAG anchors the others. The fix is procedural: each expert produces a DAG in isolation, then the disagreements are classified, reconciled through structured discussion, or recorded in a disagreement register for later evidence. Walked through on a high-school curriculum example in the Pearl tradition.
Reconciling Expert Parameters → When data describes a node distribution one way and experts describe it another way, the procedure is not to average them.
In the linear Gaussian convention, edge weights start fixed at one and the elicitation is about parent node distributions — means and variances. When data and expert estimates disagree about a distribution, the gap is usually structural rather than parametric: a selection variable, a definition mismatch, or a population that doesn’t transport. Bayesian updating is the right resolution mechanism once the structural fix has been applied, not before. Walked through on the same curriculum example as the DAG reconciliation page.
Where this fits
Building is the second half of an engagement. The Foundations cluster covers what a causal model is; this cluster covers how one comes to exist.