Going Deeper
For Both Executive and Technical ReadersLonger pieces on how models get built and where this work sits in the wider AI landscape, written for the most skeptical person in the room.
01 What This Column Is For
The other columns answer questions about the practice. This one answers questions about the wider landscape: how models get built, what has been tried before, and where causal reasoning sits in the AI industry as it exists today.
These are longer than the other pages and less immediately practical. They are here because the questions come up in real conversations, usually from the most skeptical person in the room, and they deserve a fuller answer than a meeting allows.
02 How a Model Gets Built
How a Model Gets Built is the most practical page in this column and the one most engagements start from.
It covers how expert knowledge gets drawn out and turned into a working model, what to do when there is very little data, how the result connects to systems you already run, and what gets reused from one project to the next.
If you are trying to picture what the first eight weeks of an engagement actually look like, this is the page.
03 The Deep Dives
Eight essays, each taking on a question that comes up repeatedly.
A Cautionary History
The Watson Lesson, the most expensive AI failure in healthcare, and what it showed about the limits of systems that have no model of cause and effect. Worth reading before you approve any large AI budget.
A Practical Problem
Synthetic Data, how to build data for situations that have never occurred, how to prove it is sound, and the mistake that quietly turns a causal model back into an ordinary correlation.
The Landscape Question
Why Isn’t Causal Reasoning Already Built In?, if this matters so much, why have the large AI labs not done it? The honest answer, in plain terms.
Where Causal Structure Actually Appears, the same argument for a technical reader, naming the specific methods and research behind each claim.
Language, Cognition, and Learning
Grammar, Mood, and the Shape of Cognition, grammatical mood as a real, if imperfect, sensor for which kind of causal question a stakeholder is asking, and why that forces a sharper statement of what pairing an LLM with a causal model actually does.
Timestamps are Required, a learning organization without a decision journal is running on reconstructed memory, not evidence. Why every governance artifact in this practice is, underneath, a dated record made before the outcome was known.
Three Challenges, Not One, agreeing an ontology, classifying the rung of a live question, and reading intention from an expert’s grammar look like one problem from a distance. They fail independently, and solving one says nothing about the other two.
Lanes, Gates, and Loops, the Pipeline redrawn as a workflow diagram: three lanes for who does the work, three gates that certify or refuse, three loops that bend the line back on itself. One of the three lanes turns out to be the thinnest.
04 Two Versions of One Argument
The last two pages above are the same argument written twice, at different depths. This is deliberate.
The shorter one is for a sponsor deciding whether to fund the work. The longer one is for the person that sponsor will forward it to, typically someone technical whose first instinct is to check whether the claims hold up. It carries twenty-one references for exactly that reason.
Send the short one. Let them find the long one.
05 What Is Still Coming
A documentation set covering the EARA architecture in full is in progress and not yet published. It will sit in this column when it is ready.
Until then, the architecture is covered at working depth in the Architecture overview and the pages beneath it.
These questions come up from the most skeptical person in the room. They deserve a fuller answer than a meeting allows.