Continuous publishing

Most publishing systems make one decision unusually easy: create something new.

This experiment tests a different default. When a new signal, piece of evidence or argument appears, the first question is not what should I publish? It is what does this change in the body of work that already exists?

The decision

Before creating another asset, interrogate the corpus and choose one of five actions:

1. Extend — the existing argument still stands, but the new signal adds evidence, a mechanism or a meaningful qualification.
2. Correct — the existing position is inaccurate, overstated or no longer true.
3. Merge — the proposed new piece is substantially a restatement of something already published.
4. Split — one piece has developed into two genuinely distinct arguments.
5. Create — the signal supports a genuinely new argument that does not belong inside existing work.

The point is not constant revision. It is to stop the economics of publishing from making duplication the easiest option.

What AI changes

Every part of this was possible before generative AI. The constraint was coordination cost.

A person maintaining a large corpus has to find the relevant work, reconstruct its context, compare the new idea against previous positions, make the editorial decision, implement the change, preserve links and metadata, and decide whether the revision should be surfaced again.

AI makes enough of that comparison and implementation cheap that the published layer can remain active.

The human decision becomes more important, not less: does this actually change the argument?

The current implementation

Editorial Intelligence is testing the model on itself.

The website preserves original publication dates while allowing material revisions to carry an updated date. Revised pieces can resurface without pretending they were newly published. GitHub provides a reviewable history of what changed.

The working pattern uses different AI environments for different jobs. ChatGPT is often the conversational layer where a new observation is explored and challenged. Claude working against the repository encounters the actual corpus, tests whether a proposed addition duplicates existing thinking, and implements approved changes. GitHub preserves the record.

That is not a requirement of the method. It is the current implementation.

The training exercise

This is also a teachable Editorial Intelligence skill.

Give a practitioner a new signal and an existing body of work. Do not ask them to write a post. Ask them to inspect the corpus, choose extend / correct / merge / split / create, and defend the decision.

That trains editorial judgement across knowledge rather than prompting technique.

What happened on 20 August 2026

One morning provided the first concentrated test. Existing pieces were extended with new arguments; a newly made distinction was corrected when further thinking exposed its weakness; the site gained explicit revision handling; revised work was made visible without being presented as new; and an essay that had developed two arguments was split while its original URL and inbound links were preserved.

The accompanying Field Note documents the sequence in detail: I could not have published like this before AI.

The broader economic argument sits in AI didn’t create Editorial Intelligence. It made it economically viable.

What still needs proving

A one-person corpus is the easiest possible governance environment. The harder test is organisational: multiple authors, approval rights, regulated claims, competing interpretations and responsibility for canonical positions.

So this remains an experiment rather than a universal prescription.

The hypothesis is narrower:

AI can make maintaining relationships between published ideas cheap enough that organisations no longer have to treat publication as the end of thinking.