Reserve Delta, Escalation Probability, Comparables: The Feature Set

TL;DR — Canotera separates text extraction from prediction. Generative AI structures the case file, while geometric machine learning calculates settlement ranges, escalation probabilities, and reserve deltas based on historical comparables.

You open a new claim file. It contains a 50-page complaint, 400 pages of medical records, and a dense demand letter. The claims management system defaults to a standard initial reserve based on a simple injury code. You know that number is wrong. Social inflation and the threat of nuclear verdicts mean the actual exposure is buried in the text, but you lack the hours to read every single page today. This is the baseline reality of claims handling. Systems that attempt to solve this by dumping the entire file into a large language model and asking for a settlement value fail immediately. Language models are text engines. They predict the next sequence of words based on training weights. They do not calculate liability. They do not understand the time value of money, nor do they comprehend the nuances of venue volatility. Building a functional forecasting platform requires separating the reading of the file from the mathematics of the prediction.

We designed Canotera around a strict, non-negotiable boundary between text extraction and numerical forecasting. The generative AI layer acts purely as a reader. It ingests thousands of pages of unstructured case files—pleadings, correspondence, medicals, expert reports—and structures the facts. It maps specific injuries, venues, plaintiff counsel history, and treatment timelines into a deterministic JSON schema. Once that data is structured and validated, the generative AI's job is done. It does not guess the outcome. We pass those structured features to a separate geometric machine-learning pipeline. This pipeline is trained entirely on large numbers of resolved cases with known financial outcomes.

The Mathematics of the Output

The geometric models produce a settlement range, not a single point guess. A point prediction is mathematically dishonest when dealing with litigation. Juries are unpredictable. Plaintiff demands shift based on external factors like third-party litigation funding. The model outputs a calibrated range based on the density of historical outcomes for similar fact patterns. From this range, we calculate the reserve delta. This is a direct, automated comparison between the model's output range and the carrier's current posted reserve. If the system calculates a likely settlement range of $250,000 to $400,000, and the current reserve sits at $50,000, the delta triggers an immediate review. The goal is to set realistic reserves on day one instead of stepping them up incrementally over three years as bad news trickles in.

Escalation probability operates on a different mathematical axis. A claim might look benign on day one but carry latent risk factors. Third-party litigation funding, specific plaintiff firms known for pushing trials, or subtle shifts in medical treatment patterns often precede nuclear verdicts or drawn-out litigation. The model evaluates the structured features against the historical dataset to calculate the exact mathematical probability that the claim will breach a specific severity threshold or require extensive defense spend. This acts as a deterministic early warning system. It allows claims leaders to allocate their best defense counsel to the right files before the escalation actually materializes. You stop reacting to plaintiff maneuvers and start anticipating them.

Traceability as an Engineering Constraint

Claims professionals do not trust black boxes. If an algorithm suggests increasing a reserve by $300,000, the claims handler has to justify that decision to a reserving committee. Telling a committee that the computer said so is an unacceptable answer. Traceability is an engineering constraint, not a user interface flourish. Every output Canotera generates must trace back to the specific drivers that influenced the model. We designed the architecture to maintain an unbroken chain of custody between the final numerical output and the raw source document.

We surface these drivers explicitly. The system calculates feature importance and shows exactly which facts pushed the settlement range up or down. If a specific surgical intervention or a known aggressive plaintiff attorney drove the escalation probability, the platform highlights the exact sentences in the source documents where those facts were found. The generative AI provides the pointer back to the raw text, ensuring the handler can verify the data instantly. The handler remains the final decision-maker; the software simply does the heavy lifting of surfacing the relevant variables.

Alongside the drivers, the system retrieves comparable resolved cases. The geometric model identifies historical claims that occupy the closest mathematical space to the active file. We map the structured features into a high-dimensional space and measure the distance between the current claim and historical precedents. The system retrieves the nearest neighbors. You see the actual cases, their fact patterns, and exactly how they resolved. This changes the dynamic of a settlement discussion. You negotiate from a position of hard data, anchored by historical reality rather than gut instinct.

Latency, Security, and Real Claims Work

Integrating this into a live claims workflow requires strict attention to latency and system architecture. Ingesting a 2,000-page PDF, running optical character recognition, extracting the entities via the LLM pipeline, and executing the geometric models requires significant compute power. Processing this synchronously would result in unacceptable timeout errors. We optimize the ingestion pipeline to handle these documents asynchronously. The Canotera API receives the payload from the core claims system, processes the unstructured data in the background, and fires a webhook callback with the settlement range, escalation probability, reserve delta, and comparables. This data populates directly in the handler's dashboard. The prediction updates automatically as new documents arrive and the facts of the case evolve.

Security dictates the infrastructure. Claims files contain highly sensitive personally identifiable information and protected health information. We do not pass this data to public, multi-tenant language models via standard API endpoints. The extraction models run in isolated, single-tenant environments with strict access controls. The data remains encrypted at rest and in transit. The models do not learn from a carrier's proprietary data to benefit a competitor. The historical training data for the geometric models is strictly segregated from the active inference pipeline. We treat claims data as highly sensitive material that must be secured, processed, and restricted to authorized personnel.

A prediction you cannot explain is just a liability.