The intelligence layer for law.
Deterministic, calibrated models built from real legal outcomes — not inference, not approximation. Jurisdiction-specific, judge-aware, and documented to actuarial standards.
Not inference. Calibrated models.
Splitifi models are production artifacts — fixed models with documented calibration curves and registry entries. They do not hallucinate. They do not drift with upstream API changes. They produce identical output at any scale.
Every prediction carries a calibrated probability. Not a directional signal, not a confidence range — a number that means what it says. If the model predicts 67%, comparable cases resolved that way 67% of the time. Calibration is tested against a hard gate before any model reaches production.
This is not a distinction without a difference. In legal practice, a miscalibrated probability is worse than no probability. Splitifi treats calibration as a first-class requirement, not a post-hoc evaluation.
Pattern-matched from training text. Probabilities are not calibrated. Output varies across runs. Cannot cite source outcomes.
Trained on labeled court outcomes. Probabilities match observed frequencies. Identical output at scale. Every prediction traceable to source data.
Calibrated predictions across every dimension of a case.
Win probability, award ranges, case disposition.
Ruling patterns, grant rates, bench tendencies.
Settlement zone, timing, issue-by-issue ranges.
Pleading quality, evidence strength, procedural posture.
Parenting time splits, support calculations, modifications.
Division percentages, characterization, valuation disputes.
Fee awards, damages, punitive multipliers by jurisdiction.
Time to resolution by case type, judge, and complexity.
We don’t predict duration once. We continuously manage it as evidence changes.
Time to resolution is the variable that turns legal outcomes into financial ones. Splitifi’s duration layer serves bands with a monitored tail, never single dates; re-conditions every estimate as procedural evidence arrives, with the full forecast history preserved; and rolls each change through to portfolio and capital views. The institutional deep dive, the Time-at-Risk metric family, and the duration research series are published by Criterica Intelligence.
Explore Duration Intelligence →From case inputs to calibrated prediction.
Select a prediction scenario. See the model metadata, case inputs, factor attribution, and calibrated output — exactly as the production system processes it.
Ranked factor contributions to this prediction.
Predicted probabilities calibrated against held-out outcomes. Calibration is required for production promotion.
Four gates. No exceptions.
Every Splitifi production model passes a four-gate promotion protocol before it handles a single live prediction. The methodology is not proprietary — it is rigorous application of established statistical standards to a domain that has historically operated on instinct and precedent alone.
Models that fail any gate are demoted to experimental status. Models that pass all four are registered with full documentation — discrimination score, calibration curve, calibration error, and training date. Every model in production can survive technical diligence because it was built to.
A minimum discrimination threshold is required. Models failing the statistical floor do not advance. No exceptions.
Every input reviewed against the decision time boundary. Any information unavailable at decision time is excluded. Train on past. Test on future.
Predicted probabilities must match observed frequencies. Models failing the calibration gate are demoted.
Production candidate validated against held-out outcomes it has never seen. Registry entry documents discrimination, calibration error, calibration plot, and training provenance.
Every prediction is explainable.
Splitifi predictions are not black-box scores. Every output carries factor attribution — the ranked contributors to that specific prediction, with direction and magnitude. Which factors increased the probability. Which reduced it. By how much.
This matters in law for three reasons: attorneys need to advise clients with specificity, not just a number. Courts may require methodology disclosure under Daubert and Rule 702. And any prediction you cannot explain is a prediction you cannot defend.
The output is court-defensible: documented methodology, known error rate, testable, peer-reviewable.
Ranked factor contributions for every prediction. Which factors drove the outcome and by how much — not aggregate importance, but case-specific attribution.
Documented methodology. Known error rate. Testable and peer-reviewable. Prediction output structured to satisfy Model Rules 1.1 and 1.4 disclosure obligations.
Each factor contribution carries sign and scale. A factor does not just matter — it matters in a specific direction, by a specific amount, for this specific case.
Production models are fixed artifacts with documented calibration. No upstream API dependency. No prompt-based inference. Identical output at any scale.
Federal. State. Provincial. International.
Production models cover all 50 US states and DC, Canadian provinces (BC, AB, ON, QC), Australian states, and England & Wales plus Scotland. The same methodology. Different jurisdictional datasets. The same rigorous validation.
The same calibration discipline extends to institutional scale. Jurisdiction-level benchmarking and funder-side underwriting inputs are published by Criterica Intelligence, the regulated outcomes intelligence platform for litigation funders, insurers, and enterprise legal. Legal capital — pre-settlement funding, law-firm capital, and portfolio finance — is served through Criterica Capital.