"A general model tells you what tends to happen. A model trained on your book tells you what tends to happen to you. The base supplies breadth no single portfolio can hold. Your data supplies the signal only your portfolio carries. The work is fitting one to the other without ever taking the raw data out of your hands."
Trained on the portfolio you actually hold.
Every bespoke model starts from your proprietary portfolio — your matters, your counterparties, your venues, your realized outcomes. That book carries signal no external dataset can reconstruct. Trained alone, it is too narrow to generalize.
Blended with the Criterica base of 475M+ real court records and 23,706 production models, the narrow signal gains the breadth to hold up out of sample. Your data plus the base beats either alone. The result is a model whose expected outcomes are tuned to the exact book being priced, funded, insured, or managed.
We never hold your raw data.
Institutional buyers do not get to hand over raw loan-level or claim-level data, and they should not have to. Models can be trained inside your own environment, or inside a secure clean room where only derived features move and the raw records never leave your control. Criterica trains against the signal without ever taking custody of the book.
Data control is the single constraint that blocks banks, insurers, and large funders from working with an outside modeling partner. Remove the constraint and the engagement is unblocked. The clean room is not a compliance checkbox bolted on at the end. It is the architecture the whole engagement is built on.
Train an outcome model on your own proprietary portfolio, blended with the Criterica base of 475M+ real court records and 23,706 production models. Your data carries the signal specific to your book. The base carries the breadth no single book can hold. Their data plus the base beats either alone.
You do not have to hand over raw loan-level or claim-level data. Models train inside your environment, or inside a secure clean room where Criterica never holds the raw records. Data control is the constraint that blocks most institutions from outside modeling. Removing it is what unblocks the engagement.
Client-only models hosted under defined SLAs, with explicit versioning and a fixed refresh cadence. Every prediction is traceable to a model version. Every version is retrained on a schedule you set, against data that stays under your control.
Audit the auditor. Independent validation and challenger-model benchmarking of your existing internal models: how often each is right, how well it separates strong cases from weak, and how stable it stays, measured against a jurisdiction-matched base. Built to support model-risk-management mandates, not replace your governance.
Outcome models a partner platform can ship under its own brand, powered by Criterica underneath. Your interface, your customer relationship, our validated data science. Delivered as a private model surface, not a logo on a slide.
Six steps from your book to a private, tuned model.
The raw data never leaves your control. The model never leaves the version trail.
Define the decision the model serves — pricing, funding, reserving, or portfolio management. Identify the target, the population, and the constraint set before any data moves.
Stand up a secure clean room, or connect inside your environment. Raw loan-level and claim-level data stays under your control. Criterica works against features, never the raw book.
Train on your portfolio blended with the Criterica base, against real court data, never synthetic, with checks that keep the model from ever seeing the answer in advance.
We document how often it is right, test it on cases it never saw, and check it against what actually happened in comparable matters. The model ships with its reliability shown, not asserted.
Private hosting under SLA. Client-only access, explicit versioning, every prediction traceable to a sealed model version.
Quarterly retrain against new outcomes and an expanding base. Reliability is re-measured each cycle so the model stays tuned to the book as it changes.