Module Boundaries: Verdict-Heavy vs Settlement-Anchored Data

TL;DR — Separating litigation data into strictly bounded modules prevents selection bias from inflating baseline claim valuations. Generative AI structures the facts, but distinct geometric models must isolate verdict tail risk from expected settlement clearing prices.

A database full of jury verdicts is a terrible place to look for a settlement value. Most litigation data providers sell access to scraped public dockets. This means they are primarily selling verdict data, trial judgments, and motion outcomes. When engineers ingest this data without strict architectural boundaries, the resulting models inevitably predict massive payouts for routine liability claims. The training data suffers from severe selection bias. The public record is a graveyard of extreme outliers, while the vast majority of claims resolve quietly behind closed doors. Feeding these two completely different realities into a single vector database creates a machine that hallucinates risk.

The cases that reach a jury are the ones where plaintiff and defense fundamentally disagree on the underlying facts or the liability probability. This is Priest-Klein selection bias in action. Routine cases settle early because the clearing price is obvious to both sides. If you train a single machine learning model on a blended dataset of settlements and verdicts, the model learns a warped distribution of risk. It pulls the expected value up artificially. It terrifies claims adjusters into over-reserving on day one. You cannot average a private compromise with a public fight and expect a mathematically sound number.

The Architecture of Separation

We enforce strict module boundaries between verdict-heavy data and settlement-anchored data. At Canotera, our generative AI layer has one highly specific job: read the raw case file. It ingests the pleadings, the medical records, and the correspondence. It processes thousands of pages per claim to structure the messy, unstructured facts. It identifies the injuries, extracts the chronological timeline, and isolates the legal theories. It does not predict the financial outcome. We isolate the reading comprehension task from the mathematical forecasting task. This strict separation of concerns allows us to process highly sensitive medical records securely, stripping out personally identifiable information before the structured data ever reaches the predictive layer.

The prediction happens in separate, purpose-built mathematical models trained exclusively on resolved cases with known outcomes. The settlement module operates strictly on private, resolved claims data. This data is inherently sparse because settlements are highly confidential. Triangulating a clearing price requires mapping the specific drivers of a new claim against the geometric space of these known, resolved settlements. We represent claims as high-dimensional vectors where each dimension corresponds to a material fact extracted by the generative AI. The model outputs a calibrated settlement range, not a single point guess. This range reflects honest uncertainty bands based on the density of comparable settled cases in that specific region and injury profile. If the local space is sparse, the uncertainty band widens organically.

The verdict data lives in its own isolated module. We use it exclusively to model the tail risk. When a claim shows signs of escalation, we query the verdict module to understand the absolute ceiling of exposure. A jury award is rarely the final amount paid. Hyman and other researchers have demonstrated that post-verdict settlements, policy limits, collectability constraints, and appeals asymmetry drastically reduce the actual cash transfer. Building a system that treats a fifty million dollar verdict as a fifty million dollar payout is a fundamental engineering failure. We model the verdict as a distinct event probability, entirely separate from the expected baseline settlement payout.

Triangulation and Honest Uncertainty

Maintaining these boundaries requires rigorous data hygiene during the initial ingestion phase. Every resolved case we use for training must be cleanly tagged by its resolution mechanism. Did this specific claim resolve via early settlement, post-discovery mediation, or trial? Our API exposes these distinctions directly to the claims workflow. When an adjuster reviews a reserve delta generated by Canotera, they see exactly which comparable resolved cases drove the calculation. The specific drivers behind each number are fully traceable to the source documents. If the escalation probability is high, the system surfaces the verdict-heavy comps to justify the warning, rather than silently inflating the baseline settlement range. This traceablity eliminates the black-box reserve volatility that plagues legacy predictive models.

Running distinct models in parallel introduces strict latency and compute constraints. Reading thousands of pages of unstructured medicals and pleadings requires heavy compute resources. We decouple the document ingestion pipeline from the predictive queries to solve this. The generative AI processes the file asynchronously, updating the structured claim record continuously as new correspondence arrives. The geometric machine-learning models run low-latency inferences against that structured record. This asynchronous architecture ensures that claims teams get immediate updates to their settlement ranges without waiting for a monolithic system to re-read the entire file. It also preserves strict tenant isolation, keeping sensitive client data completely walled off from the aggregated mathematical models.

Industry forces like social inflation and third-party litigation funding break static models. Verdicts are climbing significantly faster than historical economic inflation. If settlement and verdict data bleed together, it becomes mathematically impossible to track the velocity of this change. By isolating the verdict module, we can measure exactly how much nuclear verdicts are dragging up the settlement floor in specific jurisdictions over time. We track the drift independently. The models adapt to shifting liability probabilities and rising defense costs without contaminating the baseline settlement data. Claims teams can allocate defense spend accurately and negotiate from hard data, knowing exactly whether a plaintiff demand is anchored in reality or in the phantom threat of a runaway jury.

Building a reliable forecasting platform for insurance claims means respecting the underlying structure of the legal system. You have to model the reality of the dispute. This requires acknowledging that trials and settlements are completely different financial mechanisms driven by completely different incentives. The engineering architecture must reflect the legal architecture.

Bad architecture averages the extremes; good architecture maps the boundaries.