Pearl's Third Rung

Pearl’s third rung is where consulting becomes prophecy.

What it is

Judea Pearl’s Ladder of Causation has three rungs. The first is association: given that I observe X, what is the probability of Y? Most enterprise AI lives here — pattern matching, correlation, prediction from historical data. The second is intervention: if I do X, what will happen to Y? The third is counterfactual: given that Y happened, would it have happened if X had been different? This is the rung of causal reasoning, retrospective diagnosis, and genuine strategic foresight.

A system that can only associate cannot answer “did our intervention cause this outcome?” and cannot generate the counterfactual evidence that regulators, boards, and audit committees require. Pearl’s Third Rung methodology brings causal AI techniques — structural causal models, do-calculus, counterfactual inference — to the enterprise questions that genuinely require them: attribution of outcomes to decisions, retrospective policy analysis, and auditable causal evidence chains connecting an AI recommendation to a downstream outcome.

The methodology does not require replacing an existing AI system. It adds a causal reasoning layer alongside the existing predictive layer — consuming its outputs, applying causal structure to them, and producing evidence the predictive layer alone cannot generate.

When you reach for it

Pearl’s Third Rung applies when a client needs to answer questions that association-based AI cannot answer: “Did our AI-driven intervention reduce incident rates, or did rates fall for unrelated reasons?”, “If we had applied the recommended configuration change six months earlier, what would the compliance position be today?”, or “Is the pattern the model is detecting genuinely causal, or is it a confounded correlation that will break when conditions change?”

It is also the right methodology when a regulatory framework requires causal attribution — DORA’s requirement to trace ICT incidents through to root cause, for example, or the EU AI Act’s requirement to explain high-risk AI decisions in terms that a regulator can follow. Association explains correlation. Causation explains outcomes.

What you ship

• A causal structure model — a directed acyclic graph (or a structural causal model where interventional data is available) that maps the causal relationships between the variables the client’s AI system is operating on. The model is the foundation for counterfactual analysis and the primary artefact for regulatory audit.
• A counterfactual analysis — for a specified historical decision or intervention, a structured analysis of what would have happened under an alternative action, with confidence intervals and the specific causal assumptions the analysis rests on stated explicitly.
• A causal audit trail — a structured record connecting the AI recommendation, the causal reasoning applied, the intervention taken, and the observed outcome, in a form suitable for presentation to a regulator, board, or audit committee.

Linked methodologies

Pearl’s Third Rung produces the evidence base that Outcome NAV needs to verify gain: counterfactual analysis establishes what would have happened without the intervention, making the delta — the gain — auditable rather than assumed.

The Karpathy-6 Adversarial Verification gate applies to any causal analysis document produced as a deliverable — specifically checking for FM-3 Inference Leakage (training-data causal assumptions smuggled into the analysis) and FM-5 Phantom Consensus (false attribution of causal agreement among methods or stakeholders).

Start here

Causal AI engagements at Blu Wingu typically begin with a scoping session to establish which of your current AI questions are genuinely causal questions — and which are being handled with associative tools because causal tools were assumed to be out of reach. They are not. Book a discovery conversation.