Intelligence / Predictive models
Scores you can read, models that climb a ladder
A transparent 0–100 Predictive Severity score your managers can tune, advisory scores at first notice, reserve adequacy, and machine-learning claim scores trained on your organization’s own claims.
Every model is born illustrative and moves up one rung at a time — illustrative, candidate, active, retired — with a person’s reason recorded at every step. Only active models serve scores.
The question this answers
What happens when a model gets stale?
An alert goes to a person. A daily check flags an active model older than 180 days, or whose labelled rows have grown 25% since it was trained. A person always decides; nothing retrains by itself.
The scores, and what governs them
Readable first, then learned
Predictive Severity
A weighted 0–100 risk ranking of open claims. Default weights: exposure 30, litigation 18, claim age 15, fraud 12, denial 11, report lag 8, reopened 6 — weights you can change, and custom factors you can add.
At First Notice
The Severe-Claim Flag and Complexity Segmentation score a claim when it is created, as advisory chips on the claim header.
Reserve Adequacy
An adequacy band beside the claim’s reserve. In an automatic mode, a blank opening reserve can be set by the model — through the same coverage cap and authority checks as a person’s, and recorded with the model version.
ML Claim Scores
An advisory score on a claim, with its explanation. A score never writes to any claim field: autonomy is impossible, not just switched off.
The Honesty Ladder
Models trained on pooled or seed data are capped at illustrative. Promoting to active takes a reason and the typed model version.
Governance
What is recorded
Each change to a model, and each score that drove an action, leaves a record an examiner can read.
Record
Detail
Decision Record
When a score drives a real action: model and version, inputs, output, reasoning, who triggered it and who reviewed it
Replayed Answers
One decision record per claim and model version; asking again replays the stored answer
Model Report Card
Each model’s metrics, quantiles and top contributions, in your Admin Portal
Legacy Training
Your organization answers whether its legacy book may train its models
Entitlements
claims.severity_flag, claims.complexity, claims.reserve_adequacy and ml.scoring
Where we are honest
What a score is not
It is an input to a person’s decision, and the platform is built so that stays true.
It Does Not Pay Or Close
No score releases a payment, decides coverage or closes a file. The one place a model sets a figure — an automatic opening reserve — is a mode you switch on, held to the same caps and authority as a person.
Small Books Model Poorly
A model needs history. If your closed-claim volume is thin, the model stays below active rather than shipping a confident-looking one.
Bias Is Your Risk Too
A model trained on your history learns your history, including any patterns in it you would not defend. That is worth examining deliberately.
Read next
Bring twelve closed files
Six that developed as expected and six that did not. That comparison tells you more about whether modeling will help you than any accuracy statistic.