1 passage about one person’s own medical history was taken out of this page. Nothing on this site comes from any individual’s medical record.
From the archive: the structured content behind one room of the 3 September portal, shown field by field so nothing is lost.
How do you know that's true?
Generated
2026-08-29
Owner
Fred Schwacke | Vermont Synergy Initiative
Notice
Drafted by Claude 29 August 2026. Every example is from that day and is on disk. Nothing here is graded. Fred has not reviewed it.
Id
how-do-you-know
Room
methodology
State
alpha
Title
How do you know that's true?
Standfirst
Seven questions to ask about any claim. Each one is obvious after somebody asks it and invisible before. That gap is the whole problem.
Why a sequence
A checklist gets skimmed by someone who already believes the answer. A question asked one at a time does not. Every question below was skipped in a single working day by a machine with the whole archive in front of it, and each skip cost something.
Depth note
Everything opens plain. Click a question to see why it matters. Click again for what actually happened. Nothing is inferred from who you are.
Questions
N
1
Ask
Does the quote exist where it says it does?
Plain
Ask it. Then stop trusting it.
Why
This is the check everybody runs, and it proves the least. It tells you the words were written and are still there. It tells you nothing about who wrote them, what they meant, or whether they still hold.
Happened
Thirty claims were drawn at random from this project's own archive. Twenty-nine passed — the quote was real and the character range still resolved months later. Twenty-five of those thirty were still wrong to use.
Measured
29 of 30 verbatim. 25 of 30 unusable.
N
2
Ask
Who wrote it?
Plain
The question that catches the most, and the one nothing mechanical had caught.
Why
A machine writes about seventeen words for every word it is given. Eighteen months of that leaves an archive roughly nine parts machine to one part person. A search over it returns the machine's own prose, and a machine reading that back cannot tell the difference.
Happened
Two sentences written by an AI were handed to their supposed author as his own work. He recognised them by voice, in seconds, from a screenshot on a phone. Every mechanical check had passed them.
Measured
25 of 30 sampled passages were written by the assistant, not the person whose archive it was.
N
3
Ask
Is it a whole thought?
Plain
Read it alone, out loud. If it cannot end, it was never a claim.
Why
Most extracted passages are not sentences. They are spans. A phrase cut from the middle reads like a finding because it is grammatical. It is not one.
Happened
Three things were filed as claims. A question the assistant had asked. A line of writing critique about a personal letter. A piece of conversational flattery. None of them was a claim about anything.
Measured
Of the first fifteen examined, two were usable as claims.
N
4
Ask
What question produced it?
Plain
The one nobody thinks to ask, and the one that changes what the words mean.
Why
A sentence written to answer 'explain this to my wife' reads exactly like a sentence written to define a term. Same words, different object. Only the prompt tells you which.
Happened
Measured
The prompt sat four messages above the quote and no extraction had ever looked at it.
N
5
Ask
What did the conversation conclude?
Plain
A passage from the middle of a discussion is not what the discussion decided.
Why
Answers tend to end with an offer, not a conclusion. The conclusion often comes several turns later. It usually comes after someone pushed back.
Happened
The same conversation reached something specific. Butyrate and a form of selenium turn out to act on the same target, by two separate routes. The open question was whether the second adds anything the first does not. That came three turns later and was never extracted. Only the giftwrap was.
Measured
Four ways of finding a conclusion by position were tested and all four failed. First paragraphs are headings 95% of the time. Last paragraphs end in a question 23% of the time. The conclusion turned out to be findable not in the answer but in what the reader said next.
N
6
Ask
Has it been corrected since?
Plain
The most expensive question, because it fails silently.
Why
A record that cannot say which of its claims are dead will hand you a dead one. It will look exactly like a live one.
Happened
Creatinine is a waste product in blood, and a lab reports it against a normal range. A search returned a claim that one hospital had changed that range, and that nineteen such ranges had moved. It was the top result. The seventh result was its retraction, written four days later in the same conversation. The ranges had not moved. The real figure was five. The original query had confused differences between laboratories with changes over time. Both were in the archive. Both verified. One search returned both.
Measured
Two different AI systems read that output an hour apart and both repeated the dead claim. A scan then found 2,038 places where this archive corrects itself, in 479 notes, none of which had ever been marked.
N
7
Ask
What would it forbid, if it were true?
Plain
The question that sorts a finding from a feeling.
Why
A real claim rules something out. If it is true, something specific must be false and you could go and check. A claim that forbids nothing is not a mechanism. It is a mood.
Happened
'Eleven of 103 graduates matched at the medical centre' forbids something — you can count and find otherwise. 'This work improves by being corrected' forbids nothing. Both may be worth saying. Only one can be graded, and only one can be wrong.
Measured
A claim whose central verb forbids nothing cannot carry a grade above C, however large the study behind it.
What to do
Heading
What to do with a bad answer
Body
Nothing here throws anything away. A claim that fails a question is marked with what it failed and kept, because a discarded claim can never be revised and a marked one can. The weight travels with the reasoning that set it, or a reader downstream sees a number and not why. The two errors are not symmetrical. Treating a weak claim as firm is the obvious one. Discarding a weak claim because it isn't firm is more common and more expensive, because you never see what you lost.
Frontier
Heading
The open edge
Body
We tried to automate question seven and failed. A scan for causal words that are never defined found that every causal word in this archive is defined about two percent of the time — 'trigger' no more than 'produce'. The detector could not tell them apart, so it would not have found the thing that made 'trigger' worth finding. That one was found by reading, with domain knowledge, in conversation. On this evidence that part does not mechanise, and we are recording the failure rather than quietly not mentioning it again. What we think the real signal is: not that a word goes undefined, but that it is used in two incompatible senses in the same literature. That is a narrower target and we have not tested it.
Help wanted
If you have run these questions on a literature you know well, we want to hear where they broke. Especially question seven, and especially if you found a gap that turned out not to be one.
Footer
Drafted 29 August 2026. Every example is from that day and is on disk. Nothing here is graded, and a grade is the owner's to write.