A large language model turns the claim file into structured features.
Mathematical machine learning models then use those features alongside millions of resolved cases to predict settlement ranges, escalation probability, and calibrated confidence bands.
Large language models are trained to write plausible text, not to forecast outcomes with a measurable error rate.
A claim forecast needs the opposite: a model whose accuracy is measured on historical cases withheld from training and published with each prediction.
So we split the work. Language models read the file and turn it into structured features, clauses, parties, dates, medical events. A separate set of mathematical models, trained on millions of resolved cases and calibrated against cases withheld from training, produces the settlement range, the escalation probability, and the comparable cases that anchor each forecast.
Ingest
Embed
Predict
Calibrate
Retrieve
Publish
The pipeline above is the path each claim travels; the layers below are the model components that power those stages — they work together, not as a one-to-one map of the steps.
Combine classical algorithmic technique with learned components, optimized for accuracy and tightly bound intervals on legal outcomes. This is the umbrella layer that produces every published forecast — the embedders, geometric ML, signal processing, and conformal prediction below all feed into it, rather than running as separate parallel steps.
Language models live inside the predictive framework as feature extractors. They emit normalized representations of clauses, entities, and events so the geometric layer downstream can reason over rich, comparable input rather than raw text.
Cases live as positions in a learned geometric space. Similar cases sit close to one another, and the model uses both the position and the local geometry to forecast outcomes and retrieve comparables.
Time-series methods identify trends and anomalies — escalation signals, jurisdictional drift, social-inflation patterns. This layer surfaces why a file is moving the way it is.
Every forecast comes with a calibrated band and a published coverage rate — the share of cases whose true outcome lands inside the band, not a measure of the engine making mistakes. An "80% band" means that across many claims, the real outcome falls inside the band about 80% of the time. So a "$425K–$550K" damage range is a calibrated 80% interval, not a midpoint with a guessed margin.
Neural-symbolic legal outcome prediction — the combination of neural representations of legal text with symbolic reasoning to produce calibrated outcome predictions.
The legal calculator infrastructure that turns model output into actionable analytics — predictions, ranges, scenarios.
A graph neural network for tracking parties, claims, events, and dates through complex documents, so downstream models reason about relationships rather than text fragments.
Bring one anonymized file and we will run it on a thirty-minute call. You leave with the forecast, the conformal interval, the drivers, the comparables, and answers to the technical questions on the call.
