Calculator Calibration

Each case comes from a published appellate opinion hosted by the court itself. The calculator receives a verbatim excerpt of the opinion's facts and each side's contentions, with every sentence stating an award, a finding or a holding removed. The outcome is stored separately and a script refuses to run any case whose award appears in the input. Every case goes through the same live tool a user would use, not a copy of it.

Two comparisons are shown. Main claim: the calculator's per-claim expected values, excluding any attorney's-fee claim, against the trial award on the main claim. All-in: the calculator's top-line range and best guess against the whole trial judgment including fees. A hit means the actual figure fell inside the predicted range. The tool is asked what a trial court will do, so appellate changes are noted but scored separately.

Read the per-case hits with care. The calculator returns probability-weighted expected values: a claim it gives a 60 percent chance of recovering $100,000 is carried at about $60,000. A single case that was then won outright will land above that figure by construction, and one that was lost will land below it. The fairer test of a probability-weighted tool is whether its predictions add up across many cases, so the scorecard also shows total predicted against total awarded on the main claims, excluding attorney's fees, which are lumpy enough to swamp the damages themselves.

Each case names the model that produced its analysis. Cases run before Sept 23, 2026 used Claude Opus 5; the live tool moved to Claude Sonnet 5 on that date because the slower model exceeded the hosting platform's request limits on about half of real-length inputs.

The inputs, outputs and scoring script are in the public repository so the result can be checked: case_valuation_project/backtest. This tests the model, not the intake form: a real user's description is messier than a court's statement of facts.