How Arbiter's research is done, and how to check it.

Sources

• Primary first. A claim about a law cites the statute, the regulation or the regulator's own page. A claim about a study cites the paper. News coverage and law-firm alerts are used only to find the primary source, and are labeled "secondary" when nothing better could be read.
• Announced, proposed, passed and in force are kept apart, each with its date.
• Preprints are labeled as preprints, and single-author or unaudited results say so.
• Every source shows the date it was read.

Measurements

• Arbiter's own numbers come from scripts in the repository, run on synthetic records with local models. No real patient data is used, and nothing is sent to a cloud model.
• A measurement states its size. Nine cases show a mechanism; they do not estimate a rate. Accuracy comparisons need enough questions to detect the difference claimed (see the power analysis).
• Failures are published, including the ones in our own work.

Charts

• Every chart is drawn from a CSV published next to it, and has a "Show the data" table.
• One axis per chart, values labeled on the marks, and a categorical palette checked for color-blind separation and contrast before use.
• Charts that compare different populations say so in the subtitle.

How reads are made

Reads are researched and drafted with an AI research assistant (Claude, by Anthropic). It works from the primary sources under the rules on this page, and every number traces to a linked source or a published CSV. Yoonbo Cho is responsible for what is published. A routine runs every two to three days: it checks what has been published before (the index), looks for new primary sources and new results, writes the reads, and proposes agent policy changes when a finding warrants one. It cannot merge policy changes.

Corrections

Errors are corrected in place, with the date and what changed, in the read's Corrections section. Report one.

Not advice

Nothing here is legal or medical advice.