Architecture
For Both Executive and Technical ReadersTwelve capabilities beneath the pipeline, grouped by what they are for, not by their numbering.
01 What This Layer Is
Operations describes the steps a question travels through. This column describes the capabilities those steps draw on: how the model is assembled, how it takes in information, what you can ask it to do, and how you know whether to trust the result.
Twelve pages. You do not need all twelve to evaluate the approach. They are grouped below by what they are for rather than by their numbering, and the last section says which ones matter for which kind of buyer.
02 The Foundation
Two pages set the ground rules for the whole system. Both rest on the same underlying object: the models in this practice are Bayesian networks: maps of what drives what, with the strength of every link written down as a probability. The engine that computes them is Bayes Server. Every capability on this page runs on that foundation, and Foundations covers where the ideas come from.
Pipeline explains the single organizing rule: the language model touches only the two ends, understanding your question at the start, and writing the answer in plain English at the finish. Everything in between runs on fixed rules and arithmetic. Ask twice, get the same answer twice.
Composability deals with what happens when a large model is assembled from smaller ones. The joins are where things break, and this page is about keeping them sound.
03 Building the Model
Two pages cover where the model comes from and how information gets into it.
Discovery is the one that surprises people. The model is built from what your experts know, not learned from data alone, that choice is the whole wager of this practice. Discovery searches your data for candidate relationships and brings them to experts to confirm, reject, or refine. The experts decide.
Evidence handles a problem every real system faces. Information arrives hedged: probably, usually, the inspector thought so. Forcing that into a yes or no throws away exactly the part that mattered. The system takes confidence as confidence.
04 The Reasoning Services
Five pages cover what you can actually ask the model to do. The Cognitive Primitives is the umbrella: a set of reusable reasoning services that run on the shared model. The model is the ground. These are what runs on it.
Inquiry, which single missing piece of information would change your decision. A model that only reasons over what it already has is passive; this is the question that saves money, because it tells you what is worth going out and measuring.
Scenarios, name the worlds you care about (recession, rate cut, supplier failure), store each one, and compare what the model says across all of them at once.
Optimization, search across everything you could do and return the best option under your constraints. Inference tells you what is true; this tells you what to do about it.
Pattern Analysis, an answer on its own is a verdict without a reason. This supplies the reason: which factors separate this conclusion from the alternatives, and which evidence actually moved it.
05 Trusting the Answer
Three pages exist because a number without a reliability story is not usable in a regulated decision. In practice these come up in every serious buying conversation.
Robustness, the answer is a number. What a regulated decision turns on is how fragile that number is: if an expert’s estimate was off, or a reading was missing, does the conclusion hold?
Validation, there is a real difference between “the math is correct” and “this matches the world.” Both are reported where the data allows. Where only the first is possible, that is stated plainly rather than blurred.
Monitoring, a prediction engine answers the question it was asked. A model built from explicit mechanisms can answer a second one: should you still trust this right now? Conditions change, and a model that cannot notice is a liability.
06 Which Ones Matter to You
Validation, Robustness, Monitoring. These are what an auditor or regulator will ask about.
Discovery and Evidence. Together they are the answer to “we do not have enough data for this.”
Pipeline first, then Composability. The one rule and how it survives scale.
Inquiry and Optimization. What to learn next, and what to do next.
The model is the ground. The reasoning services are what runs on it. Knowing which is which is most of understanding this architecture.