Working papers, technical notes, standards, data reports and model cards from Criterica Intelligence. https://critericaintelligence.com/feed.xml 2026-11-08T12:00:00Z Criterica Intelligencehttps://critericaintelligence.com https://critericaintelligence.com/research/federal-civil-duration-2025 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

All 94 federal district courts, read from the public U.S. Courts statistics for the year ending September 30, 2025: median time from filing to disposition, median time to trial, the share of civil cases pending more than three years, filings, judgeships and vacancy. The typical district disposes of a civil case in about eight months; the national median is roughly double that because a few districts carry very large multidistrict dockets. The report argues that aged inventory is the better forward signal for a funder, ranks the districts, compares the circuits and publishes the full table with sources.

https://critericaintelligence.com/research/how-to-read-an-estimate 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

A Criterica estimate is a set of labelled figures, each with a basis, a horizon, a date and a status. This paper teaches a reader to use them well: what a band is and why it replaces a point, where dollar figures and returns come from, why direction of outcome is read only from sources that record it, what a thin segment means, how a figure responds when the law or the court changes, and what evidence turns a forecast into a claim about realized outcomes for an asset class.

https://critericaintelligence.com/research/missing-document 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

A diligence file is never complete, and the party assembling it chooses what goes in. A valuation method that can rise when evidence is withheld rewards the wrong behaviour. This paper states the rule every Criterica figure obeys: removing information never raises a certified figure, a matter on hold carries no figure at all, and one exact calculation sets every number. It works an illustrative example in which a document is removed and the band widens while the certified figure falls, and sets out what the rule means for a funder or buyer reading a data room.

https://critericaintelligence.com/research/model-cards 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

Every Criterica production model carries a card: what it estimates, for which population and horizon, how it was validated, the gates it cleared, its version and the policy for retiring it. This document sets out the fields a card contains and publishes one sample card built from fleet-level figures cleared for the research series, so a reviewer knows what to ask for and what they will receive under NDA.

https://critericaintelligence.com/research/reading-a-band 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

A diligence memo carries probabilities, intervals and a stated reliability. This note defines each one, shows how calibration is measured with a reliability diagram and expected calibration error, explains the public promotion gate of ECE at or below 0.05 on a temporal holdout, and shows why a stated confidence of 78% carries no information until its calibration has been demonstrated.

https://critericaintelligence.com/research/same-file-same-number 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

Discussion of AI in legal valuation centres on invented facts. For anyone who prices, reserves or approves on a number, the larger risk is a different number each time the same question is asked. Criterica Intelligence Research designed a repeat-request test on one synthetic case summary and a leading general-purpose model, run in November 2026. The paper sets out the design of the test, explains why dispersion matters for decisions, and states the property a decision-grade estimator must have instead: same input, same output, re-estimated only when the record changes.

https://critericaintelligence.com/research/what-the-record-adds 2026-11-08T12:00:00Z 2026-11-08T12:00:00Z

Legal platforms now intake, draft and manage matters well, and their users still ask two questions the software cannot answer: what is this matter worth, and how long will it take. This paper frames the answer as value of information: decision volume, times the cost of an error, times the reduction in dispersion a calibrated band provides. It describes how the record is embedded at the level of inputs and outputs, files in and bands and reports out, and what Criterica never sees or keeps.

https://critericaintelligence.com/research/pi-mass-tort-index 2026-10-11T12:00:00Z 2026-10-11T12:00:00Z

A quarterly read of personal injury and mass tort court volume and timing. Issue 1 covers Florida after HB 837, Georgia after SB 68, every pending federal MDL and the share of federal personal injury cases closed within 12 months of filing. It reports volume and timing only, with no outcome or value figures and no forecast.

https://critericaintelligence.com/how-we-build-models 2026-10-09T12:00:00Z 2026-10-09T12:00:00Z

This note describes the lifecycle of a Criterica production model in seven stages, from real records and jurisdiction-specific training through temporal holdout evaluation and promotion gates to a versioned registry and scheduled retraining. No stage is optional and none is waived. It is the same pipeline description given in diligence.

https://critericaintelligence.com/methodology 2026-10-09T12:00:00Z 2026-10-09T12:00:00Z

This note states what a production model means at Criterica, how a model earns that label, how the court-record corpus is built, how often models are retrained and which published figures come from third parties. It separates measured quantities from third-party estimates and from illustrative figures, and it states which basis is available to institutional counterparties on request.

https://critericaintelligence.com/registry 2026-10-09T12:00:00Z 2026-10-09T12:00:00Z

This record states how Criterica decides what counts as a production model: the promotion gates, the full record of models that did not make it, and the public methodology changelog. Each model is cleared individually, every attempt stays on the record and corrections are recorded alongside improvements.

https://critericaintelligence.com/score 2026-10-09T12:00:00Z 2026-10-09T12:00:00Z

This specification defines the Criterica Score: a number from 0 to 100 attached to one legal asset, portfolio or counterparty at a stated date, against a stated reference population and horizon. It sets the mapping, the Not rated rule, the sealed-outcome rule, validation status by asset class, versioning and governance. It is issued as a request for comment.

https://critericaintelligence.com/research/duration-and-returns 2026-08-03T12:00:00Z 2026-08-03T12:00:00Z

A multiple is time-blind and an annualized return is not. This paper works through what that asymmetry does to legal-asset economics using arithmetic any reader can verify by hand: the sensitivity of IRR to duration at a fixed multiple, the convexity of the clock, four effects kept separate, portfolio tails and capital velocity. No model output appears in it.

https://critericaintelligence.com/research/seven-clocks 2026-08-03T12:00:00Z 2026-08-03T12:00:00Z

A duration estimate such as "about 36 months" is seven forecasts under one name: seven processes with different drivers, endpoints, data and responses to intervention. This paper separates them. It then re-estimates one matter from the public federal record four times to show why duration must be a living distribution with preserved history, and not a number spoken once at underwriting.

https://critericaintelligence.com/research/time-at-risk 2026-08-03T12:00:00Z 2026-08-03T12:00:00Z

Credit has ratings and market risk has value at risk, while legal assets have had adjectives. This paper defines six measures of duration risk: TaR90, Excess TaR, Duration Drift, Cash-at-Date, Delay Concentration and Recoverable Time. Each carries a formula, a unit and the decision it feeds. Five are computed from the public federal record; Recoverable Time is defined but not computed, because public data holds no intervention histories.