The method

An examiner catalogues what an author supplies and performs the six procedures below against it. They write down what they did and what they saw. They do not grade the documentation, and they do not decide whether a person wrote the manuscript.

This is the shape of an agreed-upon-procedures engagement rather than an audit opinion: findings of fact, with the limits of each finding printed beside it. An examiner who writes “this was clearly written by a human” has stepped outside the method, and the record would be worth less for it.

Nothing here judges prose. No procedure asks whether writing “reads like AI”, because stylistic intuition is not evidence and a service built on it would be a service for punishing unusual writers.

Every finding carries the examiner’s own written observation. A finding without one is not a finding, and the system will not accept it.

A record carries a handful of numbers: items catalogued, days of dated coverage against the stated composition period, the largest uncovered gap of 14 days or more, undated items, procedures performed, and exceptions noted.

These are measurements of the documentation, not of the author, and no threshold is attached to any of them anywhere in this service. There is no minimum coverage. A record showing 40% coverage is published exactly as readily as one showing 98%, and neither is described as better.

You are asked to state which tools you used. A spell checker, a grammar tool, a translation aid or an AI assistant used for research notes are all ordinary things to disclose, and disclosing them costs you nothing here: there is no rating for them to lower.

Your statement is reproduced verbatim on your card, and the cross-check procedure reports whether any catalogued artifact contradicts it. A statement the procedure does not contradict has not thereby been confirmed, and the record says so.

An examination is refused in one situation: material that appears fabricated, or a submission with nothing to examine at all. The grounds are stated to the author in full, and they may ask for a second examiner.

Thin documentation is never grounds for refusal. If you kept little, the record says you kept little, and it goes into the register like any other. Refusing on sparseness would make this a grading service through the back door, and it would fall hardest on people whose way of working leaves the fewest traces.

We keep the proof, not the files. The files you upload, your manuscript and every evidence item, are deleted once your examination is over: 30 days after your record is published or the submission is refused or withdrawn, or sooner if you choose (straight away, or after 7 days) or ask us to delete them now. Files from a submission that is never sent are deleted after 90 days without activity.

What stays is what a record is made of: the SHA-256 fingerprint, name, size and your description of each file, the examiner’s findings, and the signed record. That is enough for anyone holding your manuscript to confirm it is the file that was examined, because they can compute the same fingerprint themselves. It is not enough to re-examine the work, so after deletion an examination can be repeated only if you supply the files again, and a file that does not match its recorded fingerprint is not the one that was examined.

• The fee covers cataloguing one manuscript's documentation and performing the published procedures against it.
• Every completed examination results in a published record. Thin documentation is not a failure and is not refunded, because the work was done and the record says what it found.
• If an examiner asks for more information, supplying it and being re-examined is included. There is no second fee.
• A materially revised manuscript is a new submission at the applicable fee, because a record describes one exact file.
• If we refuse an examination before performing it, the fee is refunded in full.

The register entry is a summary. What an author sends with a query is their locker card: a single page with two clearly separated parts.

From the author is theirs. A photo, a line about them, a note, a link to their website and a few more links, in a theme and type style they choose. The note can differ for each agent or publisher: every recipient gets their own link, which can carry its own message, is counted separately, and can be withdrawn on its own.

From the registry is not. The examination record and one entry for each kind of documentation (version histories, session logs, screenshots, notebooks, AI scans) are written by the examiner and drawn by the registry in fixed form. An author cannot edit them, hide an item, reorder them or restyle them. Everything lodged is listed, including anything unflattering and including detector reports, which are shown with a plain statement that no procedure was applied to them.

The author’s section is plainly labelled as theirs, set in plain text with no links inside the message, and every outside link shows the address it goes to. A reader should never have to wonder who is speaking.

Links are unguessable and can optionally carry a passphrase. Downloading the catalogued files is off unless the author turns it on for a particular recipient. Opening counts are approximate: link-preview and email-scanner fetches are ignored, and a refresh is not counted twice.

The card is built to be opened on a phone: no scripts, one request, a few kilobytes of markup, and typefaces served from our own address rather than a font service. A reader who has never heard of us can open it and judge the findings on their face.

This is a record of documents supplied by the author and of defined procedures performed against them. It is not a certification, a rating, or an opinion about who wrote this manuscript. No conclusion about authorship is drawn here, by design. The findings are set out so that you can weigh them yourself. Nothing in this record can prove that generative AI was not used. Documentation shows what it shows; it cannot exclude what leaves no trace. No AI-detection tool was used. Where an author supplied detector output it is listed as received and no procedure was applied to it, because such tools are unreliable in both directions and say nothing about how a text was made. The procedures were performed against the manuscript whose hash is recorded here. They say nothing about any other version of the work. Evidence supplied by an author may be incomplete or misrepresented. These procedures reduce that risk; they do not eliminate it.

If material behind a record is later shown to have been fabricated, the record is withdrawn. Its code keeps working: anyone looking it up sees that it was withdrawn, when, and why. A quietly deleted record would let a bad entry disappear, which is the opposite of what a register is for.