Who may take the work

Why a creator decides, why the default is not "everyone, silently", and why this is not a position against machines.

A repository on this site holds somebody's work: code, a book, photographs, a font, a model, a zine. Putting it where people can see it is not the same act as handing it to anyone who asks for all of it at once. On most hosts the two are one switch, called public, and the second meaning rides in silently on the first.

Here they are two decisions. Exposure says what a stranger gets: everything, or the showcase - the project's own account of itself, its README, its terms, its posts - with the files, the history and the clone kept for members and for people who asked and were let in. The AI policy says, separately, what the owner permits automated collection to do with the work: ok, with attribution, ask first, or no. The first is a lock; the second is a declaration, published where crawlers look, so that permission never given is on the record.

The default for a project whose owner has said nothing is "ask first". That is the honest reading of silence. A licence that permits copying was written for people and their tools; it does not say, by itself, that the work may be taken by the ton to train a model. Nor does it say it may not. The owner can say either in one setting, and the site says it for them in four places: robots.txt, a well-known document, a header on every answer, and the page.

This is not a position against machines. A package manager, a CI runner, a mirror, an index: each takes the same path a person does. An account - a robot's account, made by the person it belongs to, with a token as its only door. A request to read, with a sentence saying why. The owner's grant. From then on it is never scored, never challenged, never shown anything but the real thing. What is refused is anonymous bulk taking, and only that.

An anonymous client asking for a project's source is judged, cheaply and locally, by what it does: following a link no person can see, crawling paths robots.txt asked it to leave alone, wearing the name of a known collector, touching twenty projects in ten minutes, asking at a machine's pace. A score becomes a verdict: let through, asked to solve a small puzzle in a browser, or refused with a sentence that says why and how to do it properly. Every verdict is written down, and the owner can read what was served to whom.

Owners may go one step further and have a refused collector served a placeholder: a real repository that is not the project, whose first paragraph says plainly what it is and how a person asks for the real one. The site states the cost of that openly - a placeholder will now and then reach something honest that sent no credential - and leaves the choice to each owner and each instance. The design document says all of this in more detail than a page should.

What this buys the person deciding where to put a decade of work: their terms stated and readable, in their language and in the machine's; their word on collection advertised the four ways and enforced where it can be; a door for honest tools that is the same door a person uses; and a record. What it does not buy: certainty. A determined taker with a residential address and a browser can still read what the showcase shows, because the showcase is for people.

None of it is hidden. The policy is documented here and in the project's own files; the vocabulary is published; the honeypot is disallowed in robots.txt before it is ever counted; the words a collector reads when refused name the mechanism and the way in. Deception, where an owner chooses it, is per request and never institutional.

The vocabulary - ok, attribution, ask, no - is this site's own for now. When the IETF's work on AI preferences settles, it replaces these words; the meaning stays.

How it works, in six lines

• Exposure, on each project: open, showcase or closed.
• Asking to read, from the showcase, answered by the space's admins; a grant is read on that one project.
• The AI policy on each project and each space, and the instance's default; advertised in robots.txt, /.well-known/ai-policy, the X-AI-Policy header and the page.
• The guardian: anonymous requests for source scored and allowed, challenged or denied; a credential is never judged; every decision recorded; an allowlist for your own machines.
• Robots: an account of its own with a token, the same path as a person.
• The decoy, where an owner allows it: a placeholder for a denied collector, honest in its first paragraph.

This instance's words, as machines read them: /.well-known/ai-policy and /robots.txt. The terms a project may state: licences.