Colophon
Written over, never erased.
A method — and a Claude skill — for recording, measuring and disclosing the human and AI contribution while a text is being written.
Colophon records how a text was made, as it is being made, and publishes a record anyone can inspect — instead of a label that has to be believed.
The problem
There are two ways to tell how AI was used in a text, and both fail
One asks a machine to guess reading the text. The other asks the author to declare it. Both fail on the case that actually matters: a text written by a person and a model together.
Detector accuracy on AI text edited by hand.
Share of papers that disclose AI use, in a study covering more than 164,000.
The solution
Capture the process while it happens
The register — the audit trail of a text — is the main component of colophon: append-only, no event overwritten, every revision added rather than substituted.
The register is opened when you start writing, and collects and evaluates all the changes. At the end, it is closed and “frozen” using a signing key.
Colophon then calculates a set of KPIs, and a synthetic representation is made to give at a glance the contribution of human and AI. The method works on how the document is built rather than simply looking at the final result.
The measurement
Words and ideas are two different measurements
Who wrote the words and whose content is it have different answers, and the distance between them is the most useful thing the method produces. Horizontally the human share of the words, vertically the human share of the ideas. The measured text is a point, and the cell it falls into puts it in one of four categories.
The quadrant below is an example, taken from a real case.
47%
of the words are the model's.
31%
of the content is.
In this case the model wrote far more words than it brought ideas, because it came in mostly at revision, research and titling — and the first draft is 86% the author's. A single number would have said 47% and hidden that.
What you get
One document, that carries its own evidence
At the end, the content you created comes with three things attached, created by colophon:
- a note under the text
- Short, in plain language, generated from the measurement rather than typed. It says how much of the words and how much of the thinking came from the model.
- the record itself, inside the file
- Every step of the writing, in the order it happened, sealed. It is not a summary of the work: it is the work, kept.
- a way for anyone to check it
- The checker travels inside the document. A reader drops the file on it and gets an answer with no need to access external services or information. The same checker is served here too.
The site can disappear. The document still checks.
What the method does not prove
What the method does not prove
The register proves that an event was not altered after it was recorded and that it existed at a given date.
It does not prove the register itself is complete. No voluntary system can: you can record faithfully or you can omit, and cryptography does not tell the two apart.
The value of a register like this one is not in the proof: it is in the responsibility the author takes on by publishing it, and in the fact that it can be inspected.
The rest of the site
The rest of the site
Off this site
The site explains the method; the method itself is kept on GitHub. Anything you would download, file or fork is there, and nothing here is a copy of it — with one exception, which is the checker this domain serves.