Training on Known Outcomes

TL;DR — Use generative AI to extract facts from messy case files, but demand separate geometric machine-learning models trained on historically resolved cases to calculate your settlement ranges and escalation probabilities.

Large language models are built to talk. They are not built to know. When claims organizations rely on text-completion engines to predict the financial trajectory of a bodily injury lawsuit, they are asking a linguistic tool to solve a mathematical problem. Generation and prediction are fundamentally different mathematical tasks. One produces a plausible story. The other calculates a calibrated fact. The current market is saturated with the false premise that reading comprehension equals forecasting ability. The assumption is that if you feed a neural network enough legal briefs, medical records, and demand letters, it will intuitively learn the settlement value of a complex claim. This fundamentally misrepresents the architecture of generative models. A text model optimizes a singular objective function: minimizing the perplexity of the next token in a sequence. It learns the statistical distribution of words. When prompted for a reserve recommendation, it constructs a statistically average sentence that mimics the syntax of a financial estimate. It is producing language, not calculating risk.

We enforce a strict boundary between reading and forecasting. A litigated claim is a chaotic repository of unstructured information. A single file contains thousands of pages of pleadings, correspondence, independent medical examinations, and deposition transcripts. We deploy generative models here to do what they do best. They read the text and extract the raw facts from the noise. They identify the specific surgical interventions, isolate the plaintiff attorney's track record, map the venue jurisdiction, and quantify the initial demands. This process converts messy text into a neural-symbolic structure. We map the extracted entities into a rigid, structured format that captures the factual state of the case. The generative model's job ends the moment this extraction is complete. It is never allowed to predict the outcome.

The Geometry of Resolved Claims

Prediction requires an entirely different mathematical architecture. We use distinct geometric machine learning models built explicitly for forecasting. These models do not process raw text, and they do not generate language. They operate on the high-dimensional vectors produced by the neural-symbolic extraction. To train a forecasting model, you need ground truth. In claims litigation, ground truth does not exist in a corpus of public legal opinions or internet arguments. It exists exclusively in historical, resolved cases where the final settlement amount, the timeline to resolution, and the total defense costs are known, immutable facts. Training on known outcomes fundamentally changes the loss function. Our forecasting models are mathematically penalized during training whenever their predicted settlement ranges deviate from the actual historical payouts.

This training regimen allows us to optimize for calibration rather than fluency. Calibration is a precise statistical property. When a model states that a claim has a 70 percent probability of escalating into a nuclear verdict, that specific escalation must occur exactly 70 percent of the time across a large historical sample. You cannot achieve this level of honest error reporting by prompting a language model to think step by step. You achieve it by anchoring the model's weights to empirical reality. We also reject the illusion of the point estimate. Providing a single dollar figure for a complex litigated claim implies a degree of certainty that the data simply does not support. Instead, our mathematical models output conformal ranges. Conformal prediction is a framework that provides statistically guaranteed bounds on an outcome. Rather than guessing exactly one million dollars, the model provides a calibrated band, perhaps between eight hundred thousand and one point two million dollars, tied to a specific confidence level.

Anchoring the Math in Reality

Every number we produce is traceable to the source documents and driven by comparable resolved cases. When the forecasting model evaluates a new claim, it pulls the exact historical analogues that inform its settlement range. A claims professional can look at those historical cases, verify the known outcomes, and understand exactly why the current reserve delta exists. The model isolates the specific drivers. It demonstrates mathematically how a particular combination of a severe traumatic brain injury, a specific plaintiff firm, and a notoriously plaintiff-friendly venue historically inflates verdicts. This transparency is critical because the claims industry is fighting severe macroeconomic pressures. Social inflation, third-party litigation funding, and shifting jury demographics have created immense reserve volatility. In this environment, setting a realistic reserve on day one is the difference between a manageable loss and a catastrophic capital shock. You cannot afford to base your defense allocation on a gut feeling, and you certainly cannot base it on a text generator's hallucinated guess.

By separating the reading task from the mathematical forecast, we eliminate the risk of generative hallucination in the final output. If you ask a language model to predict a settlement and it invents a figure, you have no way to audit the math because there is no math. Our architecture ensures that the final prediction is a pure calculation based on historical precedent. The generative layer structures the input, but the geometric layer calculates the risk based entirely on known outcomes. This dual-model architecture is the only mathematically sound way to build a forecasting platform for litigation. It respects the limits of language models while utilizing their respective strengths. It provides claims teams with the honest ranges and concrete escalation probabilities they need to allocate defense spend and detect volatility early. You negotiate from a position of strength when your baseline is derived from actual historical settlements rather than statistical guesswork.

You can argue with a generated paragraph, but you cannot negotiate with a hallucination.