Real data only.
Every model in the Criterica fleet is trained on real adjudication data: filed records from federal courts, state courts and international jurisdictions, with outcomes as the courts recorded them. No proxies and no imputed labels. A model only holds up under institutional capital if it learned from what courts actually did.
Jurisdiction-specific models.
Not generalist classifiers. A model trained on SDNY cannot generalize to the 9th Circuit without systematic error. Jurisdictions differ structurally — in local rules, jury pool demographics, judge temperament, procedural norms, and historical base rates. These are not noise. They are signal. The Criterica fleet is built one model per jurisdiction-case-type combination, not one model for everything.
Probabilities, not labels.
The output is a number. 0.73 means a 73 percent chance, and it is built to mean that, right about 73 times out of 100. Not "high confidence." Not a sentiment label. Not a traffic light. Litigation funders, insurers, and enterprise legal teams allocate capital against probabilities, not categories. Labels destroy the information that makes allocation possible. Every Criterica model outputs a real probability you can compare directly across cases, jurisdictions, and time.
We hold back models that cannot prove themselves.
A model stays on the shelf until it can prove itself against real outcomes it was never shown. A model that looks perfect is treated as a warning sign, not a win, because it usually means it had already seen the answer. A meaningful share of the models in the registry are held back as stubs, pending more data or a cleaner signal. That is not a failure. That is the standard working. An overfit model in production destroys institutional trust faster than an honest gap in coverage.
Audit-first go-to-market.
The first engagement is always a diagnostic audit using the client's own data. Not a demo. Not a pitch deck. A diagnostic run that shows buyers their own blind spots in their own portfolio. The fastest path to institutional trust is not explaining how the models work — it is showing a fund manager that their current underwriting process is missing 23% of the risk in their existing book. First sale is always: "Show me my own data."
Capital validates intelligence.
Capital partners deploy real money against the same models sold to Intelligence buyers. This is the only credible validation that matters. Checking a model against what actually happened is necessary but insufficient for institutional trust. The question that separates real intelligence infrastructure from analytics theater is: does anyone have money on it? The capital side has money on it. That is not a marketing claim — it is an underwriting commitment that creates accountability no demo can replicate.