United States · state laws · methods

The changes we made to the plan, and why

We wrote the plan for this study of the 50 states and Washington, DC before any research, and logged every change to it before applying it. Where a change came after we had seen results, its entry says so. An answer's level is how far a law goes on a question, and its value is a named choice such as opt-in or opt-out. These are the changes that matter for the answers. Across all the rulings, the level or value of 19 answers changed. The entries below add up to 20, because one answer was changed twice.

The small codes (D-20 and so on) match the entries in our full working record. How we did the research, in plain words, is on the states page.

The check on the research agents' work records was rebuilt during the run D-20

When: October 3, 2026
What changed: Each research agent's record of what it did is checked by a script. The first version failed 28 research runs, mostly for harmless ways of writing commands. We rebuilt the check to read commands properly, had a separate agent try to break it, and fixed what it found. Under the rebuilt check, all but 2 runs passed; those (Minnesota and Wyoming) were run again from scratch. Later fixes to the same check were tested on records built to pass and records built to fail. An agent that edits its own finished output now fails every check.
After seeing results?: Yes. We changed the check after reading the research records, which is why we disclose it.
Effect on the answers: No answer changed. The change decides which research runs count, not what any answer says.

The blind re-score raters read stored copies of the cited pages, with no web access D-22

When: October 3, 2026
What changed: In the first blind re-score, the raters opened pages beyond the sources they were given. We ran the whole re-score again with each rater locked to stored copies of the cited pages and no access to the web. The first run's ratings were set aside.
After seeing results?: Yes, after seeing how the first run's raters worked.
Effect on the answers: No answer changed. It changes how the reliability figures were measured.

A second, independent search for missed laws D-24

When: October 3, 2026
What changed: Other agents searched from scratch on 25 answers and found 4 laws our researchers had missed. A rule we set before the research said that so many misses trigger a second search, so every "no law found" answer on the three questions where a law was missed got one: 14 answers, of which 8 were confirmed and 6 turned up a law.
After seeing results?: No. A rule set before the research triggered it.
Effect on the answers: No answer changed by itself. The laws it found went to the rulings below.

Rulings drafted by an agent and approved by the authors D-25

When: October 3, 2026
What changed: Where the checks disagreed, an agent drafted a ruling checked against the law's text. The authors approved 146 rulings. The 98 judgment calls, on 71 answers, were held back for the authors' own rulings.
After seeing results?: Yes. Rulings settle what the checks found.
Effect on the answers: No level or value changed. Changed the status, but not the level, of one answer (New Hampshire).

The authors' rulings on dates, sources and the hardest calls D-26, D-27

When: October 3, 2026
What changed: The authors decided six general questions, among them the five below, and then each remaining call. Two of these rulings depart from our written rules, for the flagged answers only. A law limited to one kind of data, or to insurers, counts for health data outside HIPAA. Payment data counts as health insurance information. A provision whose current version has no official date keeps its answer and shows "current version date not confirmed". A default date set by statute counts unless the law sets its own. Two court orders found only in law-firm or press copies count, and show "court order not yet pulled from the court record".
After seeing results?: Yes. These rulings settle what the checks found.
Effect on the answers: Changed the level or value of 18 answers. Changed the status, but not the level, of one answer (Alabama).

A last round of checks before publishing D-28

When: October 3 to 4, 2026
What changed: We searched for official dates for provisions that still had none, quotes our checking tool could not match on the stored page were re-read, and the 52 sampled answers the rulings had changed were rated again. Reliability is reported both before and after the rulings.
After seeing results?: Yes. We drew up these checks from the rulings and ran them before publishing.
Effect on the answers: Changed the level or value of one answer (Connecticut).

The authors' rulings on what that last round found D-29

When: October 4, 2026
What changed: A Massachusetts regulation sets an opt-out rule for record sharing through the state's exchange, and five Massachusetts dates are counted under the state's 90-day rule for new laws.
After seeing results?: Yes. These rulings settle what the last round found.
Effect on the answers: Changed the level or value of one answer (Massachusetts).

Display corrections D-30

When: October 4, 2026
What changed: A researcher note on Washington, DC's genetic records was rewritten to match the ruled answer, and 13 provisions that lacked a status were given the status of their answer.
After seeing results?: Yes, found while building these pages.
Effect on the answers: No level or value changed.

The complete working record is held in our research files and will be published with the next paper version.

Built by SuperTruth, which checks whether a record can be trusted before an AI acts on it. About SuperTruth · How we used AI

Health Record Rights Index by SuperTruth. Research tool, not legal advice. Text CC BY 4.0.