The Corpus
Criterica's models are built from a single sourcing discipline: every record enters the corpus from a public court or regulatory source, not from a vendor summary or a third-party aggregation layer that has already interpreted the underlying filing. The corpus spans verified, deduplicated court and regulatory records across state and federal civil dockets, judicial decisions, and regulatory filings in every jurisdiction the platform covers. Breadth alone is not the discipline. Every record in that corpus has been through the same identity-resolution and deduplication process before it is treated as a distinct, usable observation, which is what separates a verified record base from a raw ingestion pile.
Raw ingestion, before deduplication, overstates what is actually known, because the same underlying case, ruling, or filing frequently appears in more than one source system, under more than one identifier, with formatting that differs enough to look like a separate record until it is resolved back to the single matter it actually represents. The verified corpus is what remains after that resolution work, and it is the record base behind every answer a client receives.
How Records Are Sourced
Sourcing starts at the primary record: a docket entry, a filed opinion, a regulatory disposition, retrieved from the court, agency, or regulatory system that generated it, or from a data provider whose own feed is itself traceable back to that primary system. A record's source is captured and retained alongside the record itself, not discarded once ingestion completes, because the ability to trace a fact back to where it came from is what makes the corpus usable for institutional diligence rather than only for internal modeling.
Coverage spans the litigation and regulatory activity that drives outcome, duration, and settlement questions across the verticals this platform serves: personal injury, mass tort, commercial, intellectual property, employment, insurance-adjacent claims activity, healthcare and lien matters, real estate and construction, government contracts, and regulatory and environmental proceedings. Each vertical draws on the overlapping court systems relevant to it rather than a bespoke, siloed source per vertical, which is part of why cross-vertical patterns, a judge who sits across case types, a venue whose behavior is consistent regardless of claim category, are visible in the corpus at all.
Deduplication and Identity Resolution
Deduplication resolves the same underlying matter, party, attorney, judge, or entity across the different identifiers, name formats, and filing conventions that different court and agency systems use for it. A single case can appear under a state docket number, a related federal removal number, and a secondary reporting service's own internal identifier, and identity resolution is the work of confirming those all refer to one matter before the corpus counts it once. The same resolution discipline applies to parties and counsel: an attorney or a repeat corporate litigant appearing under slightly different name formatting across filings is resolved to a single entity, which is what makes judge-level, venue-level, and party-level pattern analysis reliable rather than fragmented across artificial duplicates.
This work is continuous rather than a one-time pass. As new records are ingested, they are checked against the existing resolved corpus before being added as new observations, and the deduplication logic itself is periodically re-validated against the full corpus to catch cases where an earlier resolution pass missed a match or, less often, incorrectly merged two distinct matters that only appeared similar on the surface.
What Feeds the Model Fleet
Not every record in the corpus feeds a production model, and not every model draws on the same slice of the corpus. The platform runs a fleet of production predictive models, each trained on a data population, a target definition, and a filter selected for the specific outcome, duration, or settlement question it answers, and each registered against a signature, its target, filter, sample size, and validation performance, that prevents two models differing only in name from being treated as separate assets. Models sit alongside governance artifacts and reference tables that support the fleet, verification logic, reference taxonomies, calibration utilities, which are kept separate from the predictive models rather than folded into them, because a governance artifact is not a model making a prediction.
The corpus also supports intelligence products that are not models in the predictive sense at all: judge-level profiles built from verified judges' decisional histories, and case-level research drawing on the cases analyzed across the platform's coverage. Both draw on the same underlying verified, deduplicated corpus described above rather than a separate, less-verified data source.
Labeling and Real-Data Discipline
Every production model in the fleet is trained on real filed records only, a discipline re-verified fleet-wide in a universal retrain and applied without exception across every vertical and every jurisdiction the platform covers. This matters because a model's outputs are only as trustworthy as the population it learned from: a model can make a verifiable claim about what real courts, judges and regulators do only if it learned from what they actually did.
Labels themselves are handled with the same discipline the corpus demands elsewhere: an outcome label is only as good as the record it was extracted from, and where a label's basis is contested, ambiguous, or dependent on a data source with known limitations, the platform's practice is to withhold or qualify the claim rather than publish a precise-looking number resting on a shaky foundation. That discipline is what allows the platform to state its conclusions plainly rather than defensively.
Questions on Data Provenance
Diligence teams, capital partners, and institutional buyers evaluating the corpus behind Criterica's models.