Liability Probability Dominates the Number

TL;DR — Generative AI cannot calculate litigation risk because it anchors on a plaintiff's stated demand. Accurate forecasting requires mapping extracted case facts into a geometric space to calculate a strict liability probability, which then defines the calibrated settlement range.

A plaintiff attorney demands ten million dollars, attaching a life care plan and an economist's projection stretching into the next century. The human instinct is to anchor immediately on that gross damages figure. Claims professionals and defense counsel spend hundreds of hours arguing over the precise cost of future medical care or the discount rate applied to lost wages. We fixate on the ceiling. But the most sophisticated defense desks understand that gross damages represent a strictly conditional state. That state only exists if liability is established. The math of litigation risk is driven almost entirely by liability probability. A high likelihood of a defense verdict collapses a massive damages model into a fractional settlement value. Conversely, absolute certainty of liability turns a moderate injury into a swift policy-limits payout. In an environment defined by rising nuclear verdicts and extreme reserve volatility, damages construct the theoretical limits of the case. Liability dictates where within those limits the actual financial exposure rests.

Generative AI fails completely at understanding this distinction. If you feed a dense pleading to a generic large language model and ask for a prediction, it acts as a credulous summarization engine. It reads the plaintiff's anchored demand and regurgitates it. A language model has no internal representation of risk, no historical memory of the local jurisdiction, and no concept of selection bias. Generative AI is exceptionally good at reading thousands of pages of medicals and correspondence to structure the facts of a case. It extracts the narrative. It does not do the math. We bridge the gap between text and prediction through a neural-symbolic structure. The generative models act as the perceptual layer, parsing unstructured text into distinct symbolic representations. They identify the exact statutory violations claimed, the specific biomechanical mechanisms of injury, and the timeline of medical interventions. Once the case is reduced to these structured symbols, the prediction engine takes over. This separation of concerns prevents the hallucination inherent in language models from infecting the strict calculation of expected value.

The geometry of expected value

Prediction requires an entirely separate architecture. At Canotera, we use mathematical, geometric machine-learning models trained on massive numbers of resolved cases with known outcomes. These models map the structured facts extracted from the documents into a high-dimensional space. By observing exactly where a new case sits relative to historical outcomes, the model calculates a strict probability of liability. It measures the geometric distance between the current fact pattern and historical defense verdicts, plaintiff verdicts, and settlement thresholds.

Calculating this probability accurately requires navigating extreme selection bias. Relying solely on public jury verdicts to estimate liability is a well-known analytical trap. As the Priest-Klein hypothesis demonstrates, the cases that actually reach a jury are not representative of the broader pool of claims. They are the highly contested disputes where the parties hold fundamentally divergent views on the likelihood of success. The obvious defense wins and the clear plaintiff victories settle long before trial. A model trained only on verdicts will severely miscalculate the true liability probability of a newly filed claim.

To build an accurate liability model, we must account for the reality that settlement data is sparse and heavily censored. Even when cases do go to trial, as the Hyman studies on litigation outcomes show, the final jury award is rarely the actual amount paid. Appeals asymmetry, collectability constraints, and policy limits warp the final distribution. The predictive model must triangulate liability signals by anchoring settlement data against verdict-heavy comparables, maintaining them in separate but mathematically related modules. We apply field-specific priors to build an honest uncertainty band around the risk.

Calibrating the settlement range

Once you establish a rigorous probability of liability, the damages forecast changes from a naive point estimate to a calibrated settlement range. This is where conformal prediction becomes critical. A conformal range provides a mathematical guarantee that the ultimate outcome will fall within the stated bounds a specific percentage of the time. If the liability facts are highly contested and the jurisdiction is volatile, the conformal range naturally widens.

This widening is not a failure of the model. It is honest error reporting. A forecasting system that outputs a precise specific dollar estimate on a case with fifty-fifty liability is lying to you. The true statistical output of such a case is a bimodal distribution: a high probability of a low nuisance settlement, paired with a lower probability of a catastrophic hit if the jury assigns full fault. Forcing a point prediction destroys the nuance required to manage the actual risk.

Claims executives need this structural clarity to set realistic reserves on day one. When a predictive model outputs a reserve delta versus the current human reserve, that delta is heavily driven by the underlying liability calculation. If the system detects an eighty-five percent chance of plaintiff liability based on comparable resolved cases, it flags an escalation probability immediately. The defense team can allocate their spend toward early resolution, rather than funding two years of futile discovery just to arrive at the same conclusion.

Negotiating from data means separating the argument over liability from the argument over damages. When you hold a calibrated forecast showing the specific drivers behind the numbers, traceable directly to the source documents, you stop negotiating against the plaintiff's anchored demand. You negotiate against the structural reality of the case. Gross damages are a hypothesis. Liability is the physics that forces the collision.