The Mechanism
Prediction begins the moment a matter's defining characteristics are known: case type, injury or claim category, jurisdiction, venue, assigned judge where applicable, and procedural posture. Those characteristics are matched against a corpus of resolved matters sharing the same characteristics, and the model returns a probability distribution over resolution classes, win, settle, dismiss, and a separate distribution over duration to resolution. The output is never a single number presented as certainty. It is a set of bands, an outcome probability with a stated confidence interval, a duration estimate with a low, base, and high case, and a support size describing how many comparable matters the estimate rests on.
The mechanism depends on comparability, not similarity in the colloquial sense. A model does not ask whether two cases feel alike; it asks whether they share the specific, quantifiable characteristics that the training population shows are predictive of outcome and duration, and it discloses which of those characteristics are present, absent, or uncertain for the matter being scored.
Prediction is deliberately the first stage rather than an input folded into underwriting, because everything downstream inherits its errors. If the probability distribution is wrong, no amount of careful pricing, structuring, or portfolio discipline downstream corrects for it; the error simply propagates through every later stage wearing the disguise of a well-run process. Isolating prediction as its own function is what makes it possible to find and fix that error at its source rather than hunting for it across five stages at once.
Inputs and Outputs
Inputs are the matter's defined characteristics as captured at intake, plus the historical corpus of resolved matters those characteristics are compared against. Outputs are threefold: an outcome probability distribution across resolution classes, a duration distribution to resolution, and a confidence layer describing the support size and comparability of the training population to this specific matter. A prediction that ships without that third output, the confidence layer, has shipped an opinion rather than a data product, because a probability with no stated basis cannot be checked, challenged, or relied upon by an institution accountable for the capital it deploys against it.
The output format matters as much as the number itself. A single point estimate invites a false sense of precision that a range does not, which is why every prediction ships as a band with a stated low, base, and high case rather than as one number an underwriter can anchor on and stop questioning.
What Can Go Wrong
The most common failure is treating a thin comparable population as if it supported a precise estimate, producing false confidence rather than an honest wide band. A second failure is training population drift: if the corpus a model draws on no longer reflects current venue behavior, current judge assignments, or current procedural rules, the model is answering a question about the past rather than the present. A third, more subtle failure is allowing the prediction to be revised after the fact to match a result the underwriting desk prefers, which is why the standard this platform follows treats prediction as a function that must never see, or respond to, what a capital decision downstream wants the number to be.
A fourth failure worth naming directly is conflating prediction with merits analysis. A probability band describing the distribution of likely outcomes for a class of comparable matters is not a legal opinion about any single matter's merits, and a model that is asked, implicitly or explicitly, to stand in for the attorney's own professional judgment about the case in front of them has been asked to do something it was never built to do.
How Criterica Implements This Stage
Criterica Intelligence owns prediction as a discrete function, producing outcome, duration, and settlement-window models from a real court-record corpus, refreshed as new resolutions accrue rather than trained once and left static. Every prediction carries a model version, a timestamp, and the exact inputs it was given, captured before any underwriting or capital decision touches the number, so that a later review can distinguish a matter where the model was wrong from one where the model was right and a downstream decision priced it badly anyway. Criterica Group's standard requires that audit trail as a baseline governance control, not an optional enhancement.
This stage links directly to the next: the prediction produced here is the sole input underwriting is permitted to price against, with no side channel for a desk to request a rerun until a preferred number appears. That boundary is enforced by process, by keeping the two functions' incentives separate, not merely stated as a policy.
Discuss the Predict Stage
Institutional buyers and capital partners evaluating the Regulated Outcome Lifecycle.