01 The Claim
A collision at an intersection. Three parties. One insurer trying to answer one question.
Party B pulls through an intersection and collides with Party A. Party C was following behind B. The insurer's question is simple: what is Party B's fault share, and how defensible is that number?
Every insurer in every market knows this scenario. The difficulty is not the facts, it is what you can say with those facts. A statistical model gives you a percentage derived from similar past claims. A causal model gives you a percentage derived from this claim, with the reasoning shown.
| Factor | Observed state | Role in the model |
|---|---|---|
| Party B speed | Alleged excess, not confirmed | Intervention target (Rung 2); abduction variable (Rung 3) |
| Visibility | Reduced | Confounder, affects both speed choice and severity independently |
| Road conditions | Wet | Amplifies severity at any given speed |
| Collision severity | Severe | Evidence node (Rung 1 diagnostic); child of speed, visibility, road conditions |
| Party A conduct | Yield failure (alleged) | Reduces B's proportionate share |
| Party C conduct | Following close | Independent contribution to collision dynamics |
| Jurisdiction | Modified comparative | Determines how fault shares map to liability recovery |
Three questions get asked, one at each rung. The models are separate: a diagnostic Bayesian network for Rung 1, an intervention model with the confounder explicitly blocked for Rung 2, and a full structural causal model with U-nodes for Rung 3.
02 Rung 1, Diagnostic
Given a severe collision, how probable is reckless speed?
The adjuster's first question is backward-looking: the collision has already happened, it was severe, and the question is what the upstream conditions most likely were. This is a diagnostic query: propagating evidence from effect to cause.
The model: six nodes, seven links. Speed, road conditions, and escape route as root causes. Collision severity as the evidence node. Fault share and liability assessment as downstream outcomes. Setting Collision Severity = Severe updates all upstream beliefs.
Query, InsuranceAttributionDiagnostic.bayes
Rung 1 result
A severe collision is most consistent with reckless or moderately excessive speed. Lawful speed drops from the prior 68% to 31%. Reckless excess rises from 10% to 35%.
Simultaneously, the posterior on Party B Fault Share shifts toward the 75–100% band. The adjuster has a defensible starting point for reserves, grounded in the observed evidence, not in an adjuster's intuition.
But this is still Rung 1. It tells us what tends to co-occur with severe collisions across the population. It does not tell us what would have happened had Party B driven lawfully.
03 Rung 2, Intervention
If Party B had driven lawfully, what would the fault share distribution look like?
The claims team's next question: what is the causal effect of speed on fault share? This is a Rung 2 query, not observing speed, but intervening on it. The do-operator severs the back-door path.
The back-door path here is Visibility → Party B Speed. Poor visibility makes excess speed more likely (drivers misjudge distances) and also independently makes collisions more severe (shorter reaction time). If you condition on Speed = Excess without blocking this path, you get a biased estimate of speed's causal contribution, some of what looks like a speed effect is actually a visibility effect.
The intervention model adds Visibility as an explicit confounder and applies do(Speed = Lawful) to sever that link. The result is the average causal effect of lawful speed on fault share across drivers in similar visibility conditions.
Query, InsuranceAttributionIntervention.bayes
Rung 2 result
Under lawful speed, intervened, not merely observed, the 75–100% fault band drops to 7%. The majority of fault shifts toward the 0–50% range, reflecting Party A's yield failure and Party C's following distance.
The gap between the observational estimate (38% in the 75–100% band for excess speed) and the interventional estimate (7% in the 75–100% band for lawful speed) is the average causal effect of Party B's speed choice, purged of the visibility confound.
This is what a court-ready causal attribution requires. The Rung 1 diagnostic gave you a posterior. This gives you the effect.
04 Rung 3, Counterfactual
For this specific driver, on this specific road: what would have happened?
The litigation team's question is sharper still: not what happens on average across drivers in similar conditions, but what would have happened to this claimant, in this accident, had Party B driven lawfully. This is a Rung 3 query, and it requires a structural causal model.
The SCM adds four U-nodes, exogenous noise variables that capture the unobserved, idiosyncratic factors of this particular claim: driver distraction and trip urgency (U_BS), exact impact geometry and vehicle mass differential (U_CS), witness credibility and adjuster discretion (U_FS), jury composition and settlement pressure (U_LI).
The three-step procedure is abduction → action → prediction:
Step 1, Abduction
Observe the actual world: Collision Severity = Severe, Jurisdiction = Modified Comparative, Visibility = Reduced, Road Conditions = Wet. Update the U-nodes to find the distribution over unobserved factors consistent with this specific incident. The U-nodes are now pinned to this claim.
Step 2, Action
In the modified model, intervene: set Party B Speed = Lawful. This is the hypothetical. The U-nodes, the idiosyncratic circumstances of this driver, this road, this adjuster, remain fixed from Step 1.
Step 3, Prediction
Run the model forward under the hypothetical. Read off the counterfactual distribution over Fault Share and Liability for this claimant, holding everything else about the accident fixed.
Query, InsuranceAttributionSCM.bayes
Rung 3 result
Had Party B driven lawfully, holding fixed everything else about this specific accident including the visibility, road conditions, the other parties' conduct, and the unobserved idiosyncratic factors of this driver, the probability that B bears 75–100% of the fault drops to 4%, and the most probable outcome is 0–25% fault share (52%).
Correspondingly, High liability becomes a 10% outcome rather than the prior baseline. The insurer's reserve and negotiating position changes substantially.
05 Why The Answers Differ
Rung 2 and Rung 3 ask different questions. They are supposed to give different answers.
The Rung 2 result, do(Speed = Lawful) → 7% probability of 75–100% fault, is the average causal effect across the population of drivers in similar conditions. It answers: if we could magically set every driver in reduced visibility to lawful speed, how often would we see high fault attributed to B?
The Rung 3 result, 4% probability of 75–100% fault, is the individual counterfactual for this specific driver. It answers: given that this driver was in excess, given what actually happened, what would the outcome have been for them specifically had they been driving lawfully?
The gap exists because the U-nodes in the SCM capture factors that are fixed for this driver but variable across the population. This driver had particular familiarity with the road, a particular trip urgency, a particular impact geometry. The Rung 3 answer conditions on those. The Rung 2 answer averages over them.
Both answers are correct. They answer different questions. The Rung 2 answer is appropriate for policy decisions, premium pricing, underwriting rules. The Rung 3 answer is appropriate for litigation, what would have happened to this claimant.
An LLM cannot produce either answer. It can describe the distinction, as this page does. But the computation requires a structural model that can execute do-calculus (Rung 2) and abduction over U-nodes (Rung 3). That is what EARA provides.
06 What This Shows
The things a fault percentage without a causal model cannot give you.
- The confound is blocked. Visibility causes both speed choice and severity. A naive correlation between speed and fault absorbs that visibility effect. The do-operator removes it. The causal estimate is clean.
- The individual is separated from the average. Rung 2 and Rung 3 give different answers because they should. A policy decision and a litigation decision require different queries. The model makes that distinction explicit.
- Every number traces to structure. The fault share posterior comes from the CPTs and the graph topology, not from pattern matching on similar claims. The reasoning is auditable at every step.
- The U-nodes name what was unobserved. U_BS, U_CS, U_FS, U_LI are explicit placeholders for driver distraction, impact geometry, witness credibility, jury composition. They are not assumed away. Their effect on the counterfactual is quantified.
- The jurisdiction matters and is modelled. Modified comparative fault affects how fault shares translate to liability recovery. That is a node in the SCM, not a post-hoc adjustment.
- What the model cannot do. The CPTs are expert-encoded over a simplified state space. Real claims require real calibration against historical data. These models show the architecture; engagement work populates them with client-specific distributions.