What your state adds to HIPAA

How we did this research

This study covers the 50 states and Washington, DC. It does not cover the US territories (Puerto Rico, Guam, the US Virgin Islands, the Northern Mariana Islands and American Samoa).

• One agent per state. For each state and for Washington, DC, an AI research agent, built on Anthropic's Claude, read the state's official code and agency pages and filled in the seven questions under a written rulebook, quoting the law and recording the page and date. It could use only official sources as evidence; legal websites, news and law-firm notes could point it to a law but could not prove one.
• A second agent checked each one. A fresh agent re-opened every source, re-read every quote and date, and searched again wherever the first found no law.
• We tested the checker with planted errors. Before the checks ran, we planted 24 wrong answers (a law hidden, an answer moved up a step, a date moved, a word changed in a quote, a source swapped, a citation from another state). The checker caught 24 of 24 (95% interval 86% to 100%). We had set 22 of 24 as the bar in advance.
• A blind re-score. Separate agents, shown only the rulebook and stored copies of the cited pages, answered 150 sampled questions again without seeing our answers. After the rulings below, and with the 52 sampled answers the rulings changed rated again, they agreed with ours on 96% of the questions with levels and on 96% of statuses. Their agreement measure (ICC) was 0.97 (95% interval 0.95 to 0.98); 1 means perfect agreement. Before the rulings, the same measures were 0.98, 98% and 98%. The raters and researchers are the same family of AI model, so they may share blind spots.
• A search of our own misses. Other agents searched from scratch on 25 answers. They found 4 laws our researchers had missed, so every "no law found" answer on those three questions (getting a copy, right to sue, sensitive records) got a second, independent search: 14 answers, of which 8 were confirmed and 6 turned up a law. Those went to the rulings.
• Rulings. Where the checks disagreed, a ruling was drafted by an agent against the law's text and approved by the authors (146 of them, on October 3, 2026); none of those changed an answer's level or value. The authors then ruled on the 98 hardest calls, on 71 answers. The authors ruled on each open question; the full record is in our methods files. Where a ruling leaves a doubt, the answer says so: "current version date not confirmed", or "court order not yet pulled from the court record".
• Changes made during the run, disclosed. We wrote the plan before the research and logged every change to it. Each change that matters is summarized on one page: what changed, when, whether it came after seeing results, and what it did to the answers.
• Kept apart, never scored. Public statistics on whether people are able to use these rights in practice, such as health insurance, internet at home and income, are a separate strand. Each state's page shows them in their own section, with source and period. They are never scored, never mixed into the answers and never used to explain them.
• No state total, no state rank. The questions are on different scales and some have no best answer, so adding them up would be a choice the data cannot defend. State laws add to the federal floor. The national score rates federal law and national infrastructure and does not change with these state profiles.

The plan, the rulebook and every change log are published with the data: the US brief links the data and the paper.