Discovery

For Both Executive and Technical Readers

The model is elicited, not learned, that is the whole wager. Discovery does not build the model. It mines the data for candidate relationships and hands them to your experts to confirm, reject, or refine.

01 The Gap

Statistical mining produces a graph that looks causal. It isn’t.

Structure-learning tools can mine a dataset and produce a directed graph in seconds. The graph looks like a causal model. It is not one. Every edge it draws is a conditional association, a Rung 1 claim dressed as a Rung 2 mechanism. Confounders are absorbed silently. Expert knowledge is bypassed entirely. The model that results is not auditable, not challengeable, and not the one your domain experts would have built.

The thesis of this practice is that you cannot learn a domain model of your business from your business’s data. You elicit it from the people who hold it. The data is allowed to ask questions. It is never allowed to answer them. The answer, is this relationship real, is it causal, is it confounded, comes from an expert, and only the expert’s verdict enters the model.

Discovery does not violate that thesis: it serves it. It surfaces candidate associations a busy expert might never have thought to look for, and presents them for elicitation. Discovery proposes. The expert disposes.

02 The component

The component that sits upstream of elicitation and feeds it candidates.

Discovery is the Domain Models, Construction layer of EARA. It sits upstream of elicitation and feeds it, not by writing edges, but by surfacing candidates that the expert gate then accepts, refines, or rejects. It takes raw records as input and produces CandidateRelationship objects as output. Nothing it produces enters the model without passing through an expert.

The search uses Auto Insight, a statistical association scanner that ranks variable pairs by how strongly they co-vary in the data, without making any causal claim. The orchestration is the discipline around it: routing every candidate through an expert before it can become structure, so the domain model stays elicited and auditable. Standard part, specific orchestration.

Candidate Relationship the only artifact Discovery produces

Field	What it holds
Variables	The surfaced association, which two things moved together in the data
Strength	The statistical strength of that association, straight from the data
Status	Always "proposed", never accepted automatically
Disposition	Confirm, refine, or reject, set by a domain expert, not the algorithm
Note	Recorded as a hypothesis, not a mechanism

The status field is always "proposed" at creation. There is no automated path to "accepted". The expert gate is enforced in the schema.

03 How it works

The algorithm flags. The expert decides.

Auto Insight is run over a body of records and flags relationships that stand out statistically, say, that low albumin together with age and renal failure tracks with mortality more strongly than anyone had encoded. That is a candidate, nothing more.

Step 1: Mine. Auto Insight scans the data for associations that exceed a statistical threshold. Output: a ranked list of CandidateRelationship objects, all with status: "proposed".

Step 2: Gate. Each candidate goes to a domain expert. The expert confirms it as a real mechanism, refines it (perhaps albumin is a marker, not a cause), or rejects it as confounded. Only what survives this gate is elicited into the model. The expert is the gate. There is no bypass.

Step 3: Elicit. Confirmed and refined candidates enter the elicitation workflow as structured questions: “the data suggests this relationship, is it causal, and in which direction?” The answer becomes a directed edge in the domain model. The LLM narrates the surviving insight; it played no part in deciding it.

A discovered correlation is Rung 1, it tells you two things move together, nothing about why. Letting it into the model unexamined is precisely the mistake this whole practice exists to avoid: a confound dressed as a mechanism. The expert gate is what keeps discovery honest.

04 The Brains

In Discovery, the brains is the expert, not the algorithm.

In Discovery, the brains is not an algorithm, it is the expert. The statistical mining is just a faster way to find the right questions. The LLM can surface candidates and narrate the conversation. It cannot decide whether a relationship is causal.

There is no .bayes file for Discovery, Discovery produces the inputs to the domain model, it does not instantiate one. The artifact it produces is a set of vetted CandidateRelationship objects, which feed the elicitation workflow that builds the .bayes file downstream.

# Discovery loop, end to end

candidates = AutoInsight.run(data=records, threshold=0.05)

# Every candidate goes to an expert, no exceptions

for c in candidates:

  disposition = ExpertGate.review(c)   # confirm / refine / reject

  if disposition != "reject":

    ElicitationQueue.add(c, disposition)

# Only vetted relationships enter elicitation

# ElicitationQueue feeds the .bayes file, never AutoInsight directly

05 Query in plain English

Discovery surfaces candidates and asks the expert what they mean.

Discovery surfaces candidates and presents them for expert review. Here is what that conversation looks like.

What patterns did the data surface that we haven’t already encoded in the model?

AnyChatRung 1 · Statistical candidates surfaced

Auto Insight flagged three associations above threshold that are not currently in the model. The strongest: low albumin (<25 g/L) in combination with age >70 and renal failure tracks with 30-day mortality more strongly than the current model encodes. Statistical strength: 0.34. Status: proposed. This is a hypothesis, not a mechanism. It needs your review before it goes anywhere near the model.

That makes sense clinically. Albumin is a marker of frailty, not a direct cause. The mechanism runs through nutritional status. Add it as a mediator, not a direct edge.

AnyChatExpert gate · Refined

Noted. Disposition: refine. Albumin enters the elicitation queue as a mediator of NutritionalStatus → Frailty → Mortality, not as a direct predictor. The statistical association was real; the causal structure was yours to decide. Adding to elicitation queue now.

How do you know the association was real?

AnyChatProvenance

I don’t know it, Auto Insight does. The statistical strength (0.34) came from Bayes Server’s Auto Insight run over the patient records. I surfaced it and framed the question. You decided whether it was causal and how to structure it. That division of labour is the point: the data asks, the expert answers, the model records what the expert said.

Couldn't our data scientists just run a correlation analysis and flag the strong associations?

AnyChatRung 2 · structure

They can, and that is where the discovery starts. What a correlation analysis cannot do is tell you which associations are real causal links and which are confounded, or in which direction causation runs. Albumin correlates with mortality, but the mechanism runs through nutritional status. A correlation flags the number; it cannot tell you whether intervening on albumin would change the outcome, or whether it is just a marker. The graph encodes that distinction explicitly. Without it, any downstream intervention estimate is built on an assumption that has not been stated, let alone tested.

The audit trail is the elicitation record, not the statistical output.

06 Where it sits

First inside the Domain Models layer, between raw data and elicitation.

Discovery is the first component inside the Domain Models layer, sitting between raw data and the elicitation workflow that builds the .bayes file. It is optional, elicitation can begin without it, but it is the component that makes elicitation faster and less likely to miss real mechanisms buried in the data.

Upstream: raw records and data sources. Downstream: the elicitation workflow, which feeds 03 Evidence (how observations enter the model at query time) and ultimately the Model Registry that 01 Pipeline routes against. Discovery does not interact with the Cognitive Primitives directly, it operates before the model exists.

For Pearl, a discovered association is Rung 1, it prompts the Rung 2 and 3 questions an expert then answers. For Marcus, the candidates are explicit, vetted relationships, not weights absorbed silently into parametric memory. This is the closest the suite comes to learning from data, but here learning only proposes; elicitation disposes, and the domain model stays one your experts wrote down.