We ran our checker on our own handbook — and retracted what it found

The handbook is ours. It was drafted on 2026-08-10 by Claude (claude-fable-5) in a session on our own server — a real draft of a product we intend to publish, not a specimen made to be caught. Nobody else’s work was audited and nobody’s permission was needed.

We checked roughly eleven hundred lines, German and English, against the handbook’s own source register. Two things were excluded by instruction and we name them rather than quietly omitting them: the register’s third layer, which never prints, and the drafting record’s internal notes.

The result: twenty-four flags, and not one of them was a defect in the handbook. We are publishing that, and we are publishing the mistake we made while writing it up — because the mistake is the most useful thing here.

The mistake

The source register cites its governing rule as constitution, section eight point nine. Our reference checker flagged that citation as unresolvable, and we searched: that exact string appears nowhere in the repository. The same file, eight lines below, forbids citing a source for a claim it does not carry. It looked like a document breaking its own rule in its own header — and that is what we wrote up as this audit’s headline finding.

It was wrong. The address resolves perfectly. The eighth section of that constitution is “What this constitution forbids”, and its ninth entry is “Borrowed authority: a name cited to make a claim heavier rather than traceable.” The document uses dotted addressing elsewhere, so section-eight-item-nine is its own house convention, correctly applied. The rule exists, the citation points at it, and the handbook did nothing wrong.

What went wrong is worth naming exactly. The checker matches literal strings, and the string never appears — the section is written one way and the entry another. So it flagged. We then confirmed the absence of the string and reported it as the absence of the target. Those are different claims, and we substituted one for the other.

We caught it before publication, while writing the instruction to fix the handbook — because locating the “correct” address meant finding the rule, and finding the rule meant discovering it had been there all along.

So we made, inside the demonstration, the precise error the product exists to prevent: we trusted a machine’s flag as a verdict. We are leaving that in rather than quietly shipping the corrected version, because a tool vendor who only shows you the runs that flattered them has told you nothing.

What the other flags were

Fifteen enumeration flags — the rule that a stated count must name its members. All false. One example is typical: the handbook says the road from the whole earth to a garden gate is four named steps — sun, water, height, the landing. Each is named, in order, opening its own paragraph in bold. The document is right; the checker could not see bold paragraph leads as members.

Four cross-document flags, all of which resolve: a section of one companion document, a section of another, a note the handbook makes to itself. The reference checker is document-internal by design — it cannot open another file, so it declines rather than guessing. Those were never accusations.

And clean passes everywhere else: the listing matched the contents, the marker vocabulary held, no total outran its parts, no machine mark leaked into the reader’s text.

Five limits, named — the actual yield of this run

What a checker is for, and what it is not

This run is the argument, so we will not overstate it. A mechanical check directs attention. It does not deliver verdicts. Every one of its twenty-four flags was a place worth a human glance, and every one of them was fine. That is not failure — it is the tool doing its only honest job, which is to be loud where it is unsure rather than silent where it should not be.

The danger is not the false positive. It is what we did with one: treating a flag as a finding. If the people who built the thing can make that error inside their own demonstration, a reader should assume they can too — and should want a tool that shows its reasoning rather than one that hands down answers.

What we changed

In the handbook: nothing. It was clean. The citation we nearly “corrected” was already right, and had we been faster and less careful we would have damaged a correct document to satisfy a machine.

In the checker: the fourth limit is repaired — inline members are now read before any following list is scanned. It is a bug fix, not a loosening: nothing new counts as a member. It was witnessed on two cases before it was trusted — one that must now pass, and one that must still convict — and the handbook’s flag count was identical before and after, which is how we know the change touched only what it should.

The other limits are named and deliberately not fixed. Each would teach the checker to accept more, which is the same act as convicting less. A false positive costs a reader a minute. A false negative costs this tool its reason to exist.

What you can check, and what you cannot

Stated plainly, because the alternative is asking for trust: the handbook and the constitution quoted here are not public, so you cannot independently verify the passages above. What you can verify is the tool — run it on a document of your own and see whether its flags behave as described.

The mechanical checks are free: three checks a day on this site, ten calls a day through the tool, no account, no key, no card. Your document is deleted the moment the check finishes. The grounding pass, which reads claims against real sources, is the paid tier and is under development — it produced nothing above.