How to Check a Policy Claim · Hanh D. Brown

I typed a false sentence into a citation checker and it came back verified.

Here was the sentence. Medicare pays for telehealth under 38 U.S.C. 1701. That is wrong. Section 1701 of title 38 of the United States Code is a Veterans Affairs statute. It defines the terms for veterans’ medical care. Medicare is a different program, under a different title, written by a different committee.

The checker was not broken. It answered the question it was built to answer, and that question was narrower than mine.

Most bad policy writing now comes from that gap. Artificial Intelligence (AI) tools rarely invent a rule out of thin air, which is the failure everyone was warned about and the one that gets caught.

They do something quieter. They hand back a real reference bolted onto a sentence the reference does not support. Then a person reads the mark and skips the page. It survives review. The green mark is not the check.

What makes a legal citation actually checkable?#

A citation is checkable when it works as an address rather than a description. An address takes you to one place. You can walk to it, put a hand on the page, and read the words sitting there. A description only tells you that something exists somewhere.

Legal citations are built as addresses. Title, then section. Title 38, section 1701. Those two numbers are the whole route, and you can reach the same page I reached, in the same rulebook, without asking my permission first. Nothing depends on trusting me.

Compare that to how most claims about policy actually travel. Someone says a program covers something. The claim moves from a briefing to a blog to a summary. Each stop trims it. By the third the address has fallen off entirely, and what is left is a confident sentence with no way back to its source. Nobody notices.

Good systems keep the address attached and make it narrower than the section. The policy graph I use returns a locator shaped like this: 38 U.S.C. 1701 chars 3576-3976. That is a span of characters inside one document. It names the words, not just the page they sit on.

Character spans matter more than they sound. Sections run long. Pointing at the section is pointing at a whole book, while pointing at the span is pointing at one line on one of its pages. Only one can be checked in a minute. That gap between a book and a line separates a source a reader can genuinely use from one that merely looks impressive on the page.

So the first question to ask of any policy claim is dull and decisive. Where is the address. A sentence with no address is not a weak source. No address means no source.

Why does a verified citation not mean a true sentence?#

A verification label means the tool found the reference, not that the sentence is true. Most checkers pull the checkable atoms out of a line. Citations, quoted spans, and numbers. Each atom is matched against stored text, byte for byte, and the weakest atom becomes the verdict.

My sentence held exactly one checkable atom. The string 38 U.S.C. 1701. That string resolves to a real statute sitting in the store, which is all the tool was ever asked to establish before it returned a verdict. So the atom passed. Nothing looked further. The sentence around it inherited a pass it never earned on its own.

Notice what was never attempted. No part of that process asked whether the words around the citation follow from it. The tool’s own documentation says so plainly, and says no model runs in that answer path by design. It is telling the truth about a narrow thing. A person skimming for a mark reads it as truth about a wide thing.

Think of a postman confirming that a house number exists on a real street. The house is there. He never opened the letter, never read a line of it, and was never asked to. Confirming the address is honest work. It is simply different work from confirming the message.

The same tool has a second verdict worth understanding, because it looks like failure and is not. When a sentence carries no citation, no quotation and no number, the verdict comes back unsupported. That does not mean false. It means nothing in the sentence can be tested by this method at all. A tool that says so out loud is more useful than one that guesses.

So the second check cannot be handed off to software. Somebody has to open the passage and read whether it says what the sentence claims. That is a person, with an eye on the text and a few minutes to spare. That job is still yours.

What is the difference between no rule and never read?#

An empty result carries two meanings that lead to opposite sentences. Either somebody checked and nothing binds, or nobody has looked yet. The first is a finding worth writing down. The second describes unfinished work, and reporting it as an absence would invent a result nobody earned.

My own system keeps those apart by name. It answers with one of five states and refuses to blur them.

Answered means facts were found and each one carries its provenance. No rule found means the coverage log shows this topic and place were genuinely checked, by somebody, on a date the log will give you. Not ingested means nobody looked. Ambiguous means more than one thing matched, so it asks which you meant instead of picking one.

Out of scope catches a different mistake. It fires when the question names nothing the system tracks. I asked about nursing home staffing rules in Ohio and got exactly that answer back. Nothing was missing from the shelves. My wording simply matched none of the terms the system was built to understand, which is a different failure entirely and a far easier one to fix.

That distinction saves you from a bad conclusion. An out of scope answer says the question was refused, never that the law is silent. Asking the same thing in the system’s own vocabulary can produce a full answer a minute later. Wording is not evidence.

Numbers behind all this caution are worth seeing. Read on 2026-09-01, the corpus held 3,351 documents and 132 of them had been read. At the smaller unit it held 145,208 chunks with 572 read. That works out to 0.39 percent.

Sit with that figure. More than ninety nine parts in every hundred of that material has never been opened by anyone, and no answer drawn from it can pretend otherwise.

Picture a library room where the shelves hold thousands of books and somebody has opened 132 of them. Ask whether the library holds anything on a subject you care about. The honest answer is that nobody knows, because nobody has walked down that aisle this year.

Nothing found is a fact about the reader, never a fact about the law. Any system reporting the two the same way will eventually publish a silence as a finding. Silence is not a finding.

How much weight does the source behind a claim carry?#

Not every source that mentions policy has standing to settle it. Authority runs in ranked levels, from the law itself down to a trade article about the law. Seven levels is common. Reading a claim well means knowing which level it rests on before believing it.

Five of those levels matter most in ordinary reading:

• Statute. The law itself, written by Congress, and the strongest thing in the stack.
• Final rule. The regulation an agency issued to carry that law out in practice.
• Approved instrument. A waiver or contract term binding one program rather than everyone.
• Proposed rule. Not in force, and reading one as current policy is a real error.
• Guidance and press. Manuals, agency posts and trade coverage, useful and never binding.

Proposed rules are the line that catches careful people. A proposal reads exactly like a real rule, carrying the agency name, the formal legal language and the numbered sections in the same official typeface. Nothing on the page warns you. It simply has not happened yet. Systems that handle this well give a proposal no start date at all, so it can never be served as current policy by accident.

The bottom of the list catches everybody else. A trade article about a rule is often clearer than the rule. Clarity is not authority. Compare a note from the fire chief describing the fire code against the fire code itself. One is easier to read. Only one decides anything when money is on the table. Reach for the harder page.

There is a second habit worth building here. When two sources disagree, a good system reports both with their levels attached rather than quietly picking a winner. Averaging them would hide the disagreement, and the disagreement is often the most useful thing on the page. Being told that the answer is contested is an answer.

Tier is not quality, and holding that distinction saves you from snobbery. A sharp analysis from a weak level may teach you more than a thin agency page from a strong one. Use the weak level to understand. Use the strong level to establish. Never let the two swap jobs.

Where do Medicare rules live in the rulebook?#

Once an agency finalises a rule, it goes into the Code of Federal Regulations (CFR). Think of the arrangement as a house. A title is a floor, a chapter is a hallway, a part is a room, and a section is what sits inside that room. Learn the address once and you never hunt again.

Medicare regulation sits in title 42, chapter four, subchapter B. Knowing that one address is most of the skill. A claim about what Medicare pays for should trace to a part in that range, or to the statute above it. A claim landing somewhere else is worth a second look.

Parts inside it are large. Part 405 is federal health insurance for the aged and disabled. Part 412 covers payment systems for inpatient hospital services. Part 423 covers the voluntary prescription drug benefit. Each part is not a page but a shelf, and each shelf holds a hundred documents or more.

Here is the honest state of my own corpus against that shelf, read on 2026-09-01. Part 412 holds 149 documents and none have been read. Part 405 holds 112 with none read. Part 423 holds 92 with none read.

They are all present. Not one has been opened. My system could list them all day and still tell you nothing whatever about what any of them actually require of a hospital or a plan.

Their subject labels come from the publisher, and that detail carries the whole trap. A publisher’s label tells you what a document is about. It never tells you what the document requires. Those are different questions, and only reading answers the second one.

So a list of documents looks like coverage and is not. Treating the catalogue as knowledge is how a confident summary of Medicare rules gets written by somebody who opened none of them, which is close to the failure I keep writing about in Artificial Intelligence systems generally. A catalogue is not a reading.

What should you do before you trust a policy claim?#

Three checks answer whether a policy claim holds, and they run in a fixed order. Find the citation. Read the passage it points to. Ask how much of the material anyone has actually read. Each check catches a failure the one before it cannot see.

First, find the citation. A claim with no address cannot be improved by further reading, so stop there and go find one. This check takes seconds and removes most of what circulates.

Second, read the passage. Open the section and read whether it says what the sentence says. My false telehealth line would have died here in about a minute. No tool performs this step for you. University research guidance puts it in plainer words: confirm that the source exists, then confirm that it supports the point, because those are two separate questions.

Third, ask who read it. When an answer comes from a system, ask that system how much of its material has been examined. A tool that cannot tell you is a tool whose silences you cannot trust. A tool reporting 0.39 percent has earned more confidence than one reporting nothing, because it is showing you where its own edge sits.

Nobody feels the cost of skipping this evenly. A family caregiver deciding whether a benefit covers a home visit is not doing research. They are making a decision, at night, after a long day, on the strength of a sentence somebody wrote confidently.

I keep landing in the same place, and I have changed my mind before. These tools have become very good at producing sentences that look sourced.

They have not become better at being right, and the distance between those two things is growing faster than public understanding. Reading the passage yourself has never stopped working.