The Vision: Institutionalizing Causality | WhyThat.ai

Architecture

Encode. Query. Answer.
The Vision

Institutionalizing causality — from expert knowledge
to computable counterfactual answers.

The theoretical foundations exist. The tools exist. What's missing is a general, human-curated causal model library with LLM-based natural-language I/O and real-time counterfactual pruning.

1 Knowledge · Structure

2 LLM · Translation

3 Engine · Computation

Experts encode causal knowledge once. Users query it forever. The causal model is the product; the LLM is just the interface — a bridge, not a brain. This is not "ask GPT" (pattern-matching) or "mine the data" (correlation). This is human knowledge, made computable.

How It Works

Humans encode knowledge. Machines compute answers.

Function	Who Does It
Causal knowledge	Humans
Model storage	Library
Math	SCM engine
Language (interpretation, routing, translation)	LLM
Judgment	User

The LLM does:

Question interpretation, variable grounding, model selection, subgraph routing, answer translation.

The LLM does NOT:

Compute probabilities, invent causal structure, estimate parameters, or answer without a model.

The LLM routes. The model reasons. Which avoids: hallucinated causality, black-box counterfactuals, and unverifiable reasoning.

Pearl's Abduction–Action–Prediction Loop

Abduction (model instantiation)

System does: Infer latent variables, condition on observed data, choose model slice.

P(U | evidence)

"What hidden causes must have been true?"

Action (intervention)

replace equation:
Smoking = f(Genetics, U)
with:
Smoking := 0

"Force the world to behave differently."

Prediction (counterfactual outcome)

Y_cf = f'(X_cf, U_abduced)

"What would have happened instead?"

End-to-End Flow

User

"Would this patient have avoided hospitalization if we had prescribed statins?"

LLM → Engine

model: healthcare_v1 | do(Statins=1) | outcome: Hospitalization | evidence: {Age:67, Chol:240}

Answer

"Given the model assumptions, this patient's probability of hospitalization drops from 1.0 to -8.17 if statins had been prescribed."

The Pipeline

From natural-language question to counterfactual answer, in seven steps.

Global Causal Model

A comprehensive DAG encoding all known cause-effect relationships in the domain. This is an institutional knowledge asset — built once, maintained over time. Example: A healthcare model with 500+ variables covering diagnoses, treatments, outcomes, demographics, and risk factors.

Human expertsBuilt once

Query Understanding

The LLM parses a natural language question and identifies which variables are relevant: "Would this patient have avoided hospitalization if we prescribed statins earlier?"

{
  "model_id": "healthcare_v1",
  "intervention": "do(Statins=1)",
  "outcome": "Hospitalization",
  "unit": "Patient_4521"
}

LLMPer question

Subgraph Extraction

Prune the Global Model to the minimal graph needed: An(Y) ∪ De(X) ∪ Conf(X,Y). This makes the model fast, interpretable, and testable.

LLMAutomated

SCM Construction

Add U variables for individual-specific factors. Template rules convert the subgraph into a full Structural Causal Model.

LLMTemplate rules

Parameter Estimation

Fit conditional probability tables from the appropriate dataset. Structure stays fixed; only parameters update. Can use historical, recent, or segment-specific data.

Inference EngineData-driven

Counterfactual Inference

Abduction: Given evidence about a specific individual, infer their U values — "who this person is." Intervention + Prediction: Hold U fixed, apply do-operator, compute counterfactual outcome.

Inference EnginePer individual

Answer Translation

Convert the probabilistic result back into natural language with uncertainty, assumptions, and model ID.

LLMNatural language

Example Walkthrough

From library model to patient-specific "what if" answer, with exact prompts and screenshots.

The Question

"Would this patient have avoided hospitalization if we had prescribed statins?"

Start with the Library Model

"Open the healthcare causal model from the library."

The causal model is a comprehensive DAG encoding all known cause-effect relationships in the healthcare domain.

Healthcare Causal Model — 24 nodes covering demographics, risk factors, treatments, and outcomes

Extract the Minimal Causal Subgraph

"Extract the minimal causal subgraph from Statins to Hospitalization."

Prune to only the nodes needed: ancestors of the outcome, descendants of the intervention, and confounders between them.

Minimal subgraph: 8 nodes

Add U Variables

"Add one U variable as a parent to each non-root endogenous node."

U variables capture individual-specific factors. Each endogenous node gets one U variable: no parents, exactly one child.

SCM with U variables — 12 nodes total

U Variable	→ Child	Captures
U_TotalChol	TotalCholesterol	Individual cholesterol factors
U_LDL	LDLCholesterol	Individual LDL factors
U_Heart	HeartDisease	Unmeasured cardiac risk
U_Hosp	Hospitalization	Individual hospitalization propensity

Set Evidence

"Set evidence: Age=67, Statins=0, TotalCholesterol=240, Hospitalization=1"

Enter the observed patient data as evidence.

Evidence entered for Patient 4521

Perform Abduction

"Run inference to abduce U values."

The inference engine performs joint inference to compute the posterior means of all U variables given the observed data.

⚠️ Critical: Use Joint Inference

The system finds U values that simultaneously maximize the likelihood of ALL evidence.

Abduction results — U values inferred via joint inference

U Variable	Abducted Value	Interpretation
U_TotalChol	40	Elevated individual cholesterol factor
U_LDL	24	Above-average LDL propensity
U_Heart	10	Higher-than-expected cardiac risk
U_Hosp	4	Moderate hospitalization propensity

Identity is now fixed. These U values won't change when we intervene.

Apply Intervention & Compute Counterfactual

"Fix the abducted U values. Apply intervention do(Statins=1). Query Hospitalization."

With U values fixed: change Statins from 0 to 1, clear evidence on Hospitalization, keep Age=67, run inference.

Counterfactual result: Hospitalization = -8.17

do(Statins=1)
    ↓
TotalCholesterol: 240 → 220  (Δ = -20)
LDLCholesterol:   154 → 139  (Δ = -15)
    ↓
HeartDisease:     12.4 → 2.85
    ↓
Hospitalization:  1.0 → -8.17

Counterfactual Result

Actual (Statins=0)

1.0

Hospitalized

→

Counterfactual (Statins=1)

-8.17

Not hospitalized

Conclusion: This patient would almost certainly have avoided hospitalization if statins had been prescribed.

The Model Library

A GitHub for Causal Models — version-controlled, collaboratively maintained, queryable on demand.

Nodes — all variables relevant to the domain. Edges — directed arrows indicating causal relationships. Metadata — variable types, valid states, descriptions. Groupings — logical clusters. Confidence markers — which relationships are well-established vs. hypothesized.

Structure, not parameters

The Global Model encodes which variables cause which — the arrows. It does not need fitted parameters. Parameters are estimated from data when an SCM is generated for a specific question.

Building Causal Models

This is the human-intensive part — and the most valuable. The causal structure becomes a reusable institutional asset.

Source	What It Contributes
Domain experts	Core causal structure, edge directions, known mechanisms
Literature review	Established relationships, effect directions, confounders
Data exploration	Candidate relationships to validate with experts
Causal discovery algorithms	Suggested structures to review (not to trust blindly)

Start small, grow iteratively

The library need not contain 500 variables on day one. Start with the 20–30 variables for the most important use case. Expand as new questions arise.

Model Composition

A query spanning two models requires bridging variables:

Query: "If we changed hiring policy, would patient outcomes improve?"

HR Model:       Hiring Policy → Staff Levels → Staff Fatigue
                                                    ↓ interface
Clinical Model:                      Staff Fatigue → Error Rate → Patient Outcomes

If bridge exists → compose into single subgraph. If no bridge → reject query, explain which connection is missing.

Watch for cross-model confounders

Modularity risks missing confounders that span model boundaries. If "Socioeconomic Status" affects both models but exists in neither, the cross-model counterfactual will be confounded.

Recommendation: Build modular, but audit holistically.

Query-Time Processing

What happens when a question arrives — subgraph extraction, data requirements, and SCM construction.

Extracting the Counterfactual Subgraph

For counterfactuals, causal ancestry is required: identify intervention (X) and outcome (Y), include ancestors of Y, descendants of X, and confounders between them. Exclude descendants of Y.

retain nodes in:
An(Y) ∪ De(X) ∪ Conf(X,Y)

Determining the Required Data

The rule: For each variable, data is required for it and all its parents — pulled from the data warehouse at query time.

If the subgraph includes...	Data is needed for...
Hospitalization (parents: Cholesterol, Age)	Hospitalization + Cholesterol + Age
Cholesterol (parents: Statins, Diet)	Cholesterol + Statins + Diet
LungCancer (parents: Smoking, Genetics)	LungCancer + Smoking + Genetics

This is why the Global Model matters

The Global Model tells the system which parents exist — the structure. The data fills in the strengths — the parameters. Without the structure, there is no way to know what data is needed.

Adding U Variables

One U variable per endogenous variable: V = f(Pa(V), U_V). U captures individual-specific factors, unmeasured causes, noise.

Step	What happens to U
Abduction	Infer U values from observed evidence — "given what was observed, what must U have been?"
Intervention	Keep U fixed, modify the structural equation for X (replace with constant)
Prediction	Propagate forward with the original U values — "same person, different treatment"

Two separate rules

U variables (model construction): One per endogenous variable. The critical assumption is independence of U's.

Interventions (query execution): Only on the target X. The other U's are inferred during abduction and held fixed — that's what makes it a counterfactual about this specific individual.

Components

The five building blocks of the architecture.

Component	Role	Examples
Global Model Store	Persistent storage for the master causal graph	.bayes files, graph database, JSON-LD
LLM Layer	Query parsing, subgraph selection, SCM construction, answer generation	Claude, GPT-4, fine-tuned models
Inference Engine	Parameter learning, abduction, counterfactual computation	Bayes Server, PyMC, DoWhy
Data Layer	Historical and current datasets for parameter estimation	Data warehouse, feature store
Orchestrator	Coordinates the pipeline, manages state	Python service, workflow engine

Encode. Query. Answer.
The Vision

Contents

How It Works

The LLM does:

The LLM does NOT:

Pearl's Abduction–Action–Prediction Loop

Abduction (model instantiation)

Action (intervention)

Prediction (counterfactual outcome)

End-to-End Flow

The Pipeline

Example Walkthrough

Counterfactual Result

The Model Library

Building Causal Models

Model Composition

Query-Time Processing

Extracting the Counterfactual Subgraph

Determining the Required Data

Adding U Variables

Components

Further Reading

Contact

Encode. Query. Answer. The Vision

Contents

How It Works

The LLM does:

The LLM does NOT:

Pearl's Abduction–Action–Prediction Loop

Abduction (model instantiation)

Action (intervention)

Prediction (counterfactual outcome)

End-to-End Flow

The Pipeline

Example Walkthrough

Counterfactual Result

The Model Library

Building Causal Models

Model Composition

Query-Time Processing

Extracting the Counterfactual Subgraph

Determining the Required Data

Adding U Variables

Components

Further Reading

Contact

Encode. Query. Answer.
The Vision