THE DISCIPLINED METHOD

A method, not a magic button

The inference delivers the value. The disciplined approach — capture, validate, gate — is what earns the right to trust it. Here is how the model gets built, signed off, and kept honest.

Approach earns the trust. Inference delivers the value.

← Back to Real-Time Causal AI

HOW WE MAKE THE INFERENCE TRUSTWORTHY

Decide. Build. Validate. Operate.

The value is the real-time inference in step 4. Steps 1–3 are the disciplined approach that earns the right to trust it.

1Decide

Suitability wizards

XMForge
CausalForge

Is causal the right tool for this problem, before building anything?

2Buildguided method

Data understanding
Causal discovery — capture engineering knowledge
Modelling — open-source methods
Testing + validation gates

refine, loop back to discovery

3Validate

Engineering Validation Gate

→ Validated Causal Model

A named gate that produces a named artifact, not an informal review.

4Operatethe product

Real-time causal inferenceranked root cause · edge or cloud

Human decides, or agent actsalways audited

Approach earns the trust. Inference delivers the value.

HOW THE MODEL IS BUILT

Three sources. One gate. One validated model.

These do not merge automatically. They converge through a gate where engineers and operators sign off — producing a model your own experts stand behind.

Engineering KnowledgeFirst principles + HAZOP

Operating PracticeOperator experience, SOPs, rules

Data EvidenceHistorian-discovered structure

HUMAN IN THE LOOP Engineering Validation Gate Two-stage engineer + operator sign-off, plus a reproducibility check

THE SIGNED-OFF GRAPH YOU OWN Validated Causal Model (VCM)

THE TRAINED MODEL THAT RUNS LIVE Structural Causal Model (SCM)

Combined is a tool’s opinion. Validated is your engineers’ commitment.

TRUST BY DESIGN

Not a black box

The discipline and guardrails are what make the output trustworthy enough to act on: validated by your engineers, not merged by a tool. Said plainly:

Nota button that auto-generates a causal graph from data Butengineering-validated models, started from templates

Nota black box Butopen, established methods, transparent and community-validated

Nota replacement for your engineers Butit captures and scales their knowledge

Notanother correlation engine dressed up as insight Butexplainable cause & effect, grounded in physics

THE METHOD AS A DOUBLE DIAMOND

Problem before solution, with a gate in the middle

Two diamonds, each diverging then converging. The suitability gate sits at the pinch, and can send a use case back to prediction or monitoring if causal isn’t the right tool.

PROBLEM SPACE

Discoverdiverge Defineconverge

Suitability gate back to predict / monitor

SOLUTION SPACE

Developdiverge Deliverconverge

PHASE 01Data Understanding

PHASE 02Question & target

PHASE 03Discovery + Modeling

PHASE 04Testing

Rigour by design: you don’t build a causal model until the problem, and the gate, say you should.

STRUCTURE BEFORE DATA

The graph is authored, not learned

Keeping the graph and the model separate is what makes the result physically meaningful — not a pile of correlations.

THE GRAPH A first-principles map of cause & effect Drawn by hand from process physics and the HAZOP study, with no historical data. Nothing in it is discovered by correlation.

THE MODEL The maths trained on that graph Fitted to the plant’s historical data, but only after sign-off. A mechanism is learned for every node in the approved structure.

Data Understanding DOCS & TAGS → HIERARCHY Organise manuals, HAZOP, setpoints and tags into an equipment hierarchy, and separate what’s in scope.

02most iteration

Causal Discovery PHYSICS + HAZOP → GRAPH Author a directed graph by hand from first principles and hazard knowledge — not learned from data.

Causal Modeling GRAPH + DATA → MODEL Only after sign-off: train the approved graph on historical data, fitting a mechanism for every node.

Testing INCIDENTS → GROUND TRUTH Run real upsets through the model and compare its ranked answer to what operators actually did.

Roots must be manipulable Only duties, flows, set points and feed can be causes. Temperatures are effects, never causes.

No cycles It must be a directed acyclic graph; a proxy variable breaks any loop.

Correlated is fine, if causal Keep both correlated variables when the physics explains the link.

manipulable process effect anomaly target

Illustrative extract · the full graph is ~51 nodes and ~155 edges.

TESTING

Ranked causes, checked against reality

Every node gets a signed score — positive drives the anomaly, negative counteracts it — and the output is checked against what operators actually did.

Every node gets a signed score

Positive drives the anomaly · negative counteracts it

reboiler duty+55%

Strongest driver, the recommended corrective handle.

reflux flow+30%

Secondary driver.

steam temp−29%

Counteracts the anomaly, not a cause.

drives anomaly counteracts

GROUND TRUTH Compared to operator action Outputs are checked against the corrective action operators actually took, with a live operator brought in to confirm the root causes.

SELF-DIAGNOSTIC The target explaining itself If the target is its own large contributor (over ~40%), a variable is missing — a built-in signal that drives the next refinement.

REVIEWED EVERY VERSION

STAGE 1 · PROCESS ENGINEERSAccuracy against the physics — are the mechanisms and directions right?

STAGE 2 · PLANT OPERATORSAgreement with plant reality — does this match what we see and do?

Never signed off in one pass — feedback reshaped the structure across versioned iterations.

TRANSFERABLE PRINCIPLES

Structure before dataAuthor from physics first; train only after sign-off.

Only manipulable variables causeTemperatures are effects — the rule that prevents most wrong graphs.

Control mode changes the pathAuto, manual and cascade are not the same graph.

Self-check for completenessThe target explaining itself signals a missing variable.

Validate against operator actionsGround truth, not just statistics. Bring an operator to the table.

Open engine underneathEstablished methods; the value is the industrial layer on top.

DECIDE BEFORE YOU BUILD

Is causal even the right tool? Decide it, consistently.

A guided wizard suite lets a domain team — without a causal specialist — decide what’s worth doing and how, so causal methods are applied the same way across every domain.

1 Application Selection Which decisions are candidates?

2 Causal vs Prediction Is understanding why worth more than predicting here?

3 Causal Readiness Is the data and knowledge in place?

4 Method Selection Which approach, and where to plug in an external library?

5 Deployment & Autonomy Edge or cloud, and the governance required?

Each wizard produces a shareable scorecard and blueprint, standardising how causal methods are applied across domains.

THE FOUNDATION FOR SAFE AUTONOMY

The method is why the output is safe to act on.

Approach earns the trust; inference delivers the value. See where real-time causal AI fits, or talk through your first use case.

Talk to an expert Back to Real-Time Causal AI