Writing on prediction, claims, and the science behind the platform
Posts on what we actually think.
Essays on conformal prediction, social inflation, and where generative AI helps and where it doesn't. Written by the team.
Why Traceability Beats Accuracy Alone
A model that spits out a perfect prediction with zero explanation is a liability in a high-stakes claim. Trust requires knowing exactly which medical record or pleading drove the math.
The Cost of Late Escalation Recognition
A routine claim sits quietly for months before a demand letter blows through the policy limit. Late escalation recognition is an operating model failure driven by unread documents and reactive defense strategies.
Reserve Delta, Escalation Probability, Comparables: The Feature Set
A single point prediction is useless in litigation. Claims forecasting requires calibrated ranges and traceable drivers to survive the scrutiny of a reserving committee.
Geometric Machine Learning on Resolved Cases
Large language models are word guessers, not calculators. To predict the financial outcome of a lawsuit, you must separate the extraction of text from the mathematics of risk.
Negotiating Settlements From Data, Not Gut
Plaintiff attorneys walk into mediation armed with structured verdicts and funding algorithms. Defense teams walk in with a spreadsheet and a gut feeling. It is time to eliminate this asymmetry at the negotiation table.
A Forecast for Every Open Claim: The API
A forecast isn't useful if it lives in a silo. We built the Canotera API to inject calibrated settlement ranges and reserve deltas directly into the systems where adjusters already work.
Measuring Calibration: Why We Publish Error Rates
A claim prediction is useless if you do not know how often the model is wrong. Publishing error rates forces a transition from guessing to actual risk management.
Where Defense Spend Is Wasted
The root cause of misallocated defense spend is not exorbitant hourly rates. It is the failure to understand the true exposure of a claim on day one, forcing carriers to fund procedural skirmishes while plaintiffs build damages.
Security and Data Handling for Sensitive Claim Records
Handing thousands of pages of raw medical and legal records to a third-party AI pipeline is a CISO's nightmare. Building a forecasting platform for insurance claims requires treating data as a liability and engineering for pessimism.
Neural-Symbolic Models for Legal Outcome Prediction
Generative AI is a text engine, not a crystal ball. To forecast litigation outcomes accurately, you must fundamentally separate the act of reading a claim file from the mathematics of predicting its cost.
Litigation Funding and the New Math of Claim Exposure
Third-party capital has fundamentally altered the incentives driving bodily injury claims. When the plaintiff's side plays a portfolio game to maximize returns, defense strategies built on historical payout curves become obsolete.
Designing for Traceability: Every Forecast Links to Evidence
A forecast is useless to a claims professional if they cannot defend it. Traceability requires engineering the system to link every predicted outcome directly to the source document that generated it.
Generation Is Not Prediction
Large language models are built to produce plausible text, not accurate forecasts. Confusing a statistical parrot for a mathematical pricing engine is a fast way to misprice your entire claims portfolio.
Predicting the Outcome of Disputes: What Claims Leaders Need
Setting initial reserves based on gut instinct and a quick skim of a file is a liability. In an era of social inflation and litigation funding, claims leaders need calibrated forecasts, not wait-and-see guesswork.
Ingesting Thousands of Pages Per Claim Without Losing Signal
A claim file is a chaotic data swamp of pleadings, medical records, and emails. Extracting the structural reality of a case from this mess requires treating ingestion as an engineering discipline, not a generic text-parsing task.
Conformal Prediction for Claims: Ranges, Not Point Guesses
A machine learning model that predicts a precise settlement dollar amount for a casualty claim is lying to you. Litigation is probabilistic, and your forecasting models must mathematically respect that reality.
Why Day-One Reserves Are Systematically Wrong
Setting an initial reserve is an exercise in institutional guesswork. Adjusters are forced to pick a number before the facts materialize, creating a compounding cycle of misallocated defense spend and missed negotiation leverage.
Inside the Canotera Pipeline: From Case File to Forecast
Processing multi-gigabyte case files requires strict architectural boundaries. We separate the language models that read medicals from the mathematical models that calculate settlement ranges.
Why LLMs Can't Predict Legal Outcomes
A language model generates text that looks like an answer. It does not calculate probabilities based on historical claim geometries. Confusing the two is a fast way to misprice your reserves.
AI as a Building Block in Insurance Analytics
Generative AI makes large, complex legal documents accessible to purpose-built predictive algorithms — when integrated correctly into the analytics pipeline.
Can You Trust LLMs to Predict the Outcome of an Insurance Claim?
Language models generate verbal predictions that sound sensible — but they were trained to sound logical, not to be mathematically sound.
Want to talk to an executive?
Press, partners, investors, candidates — the inbox is monitored. Tell us who you are and we'll route it to the right person within two business days.