Who supplied the judgment

Every tool on this site carries a field we called the AI note. It is not a disclosure badge and it is not a compliance artifact. It is a short paragraph, published on the product page, that says exactly where a machine is involved in that product and where it is not.

We added it because a label that says AI was involved tells a reader nothing they can act on. Involved how. Deciding what. The interesting information is never the presence of the thing, it is the boundary.

Some of what those paragraphs say is unflattering to us, which is the point.

The chess trainer's note says the intelligence is in the corpus rather than in any generation step: twenty six thousand positions verified by an engine before they entered the app, graded offline, with no model writing lesson content. That is a less exciting sentence than the one we could have written, and it is the true one.

The chores app for families says there is almost no machine learning in it at all, and that this was a decision rather than an oversight. A contract two people can both see, photo proof, an approval step, an append only record of what happened. Those are problems a careful schema solves better than any model. We expect to add exactly one model later, for screening submitted photos, and we said so before building it so that the addition would be visible as a change rather than arriving quietly.

The tennis tool says the vision does the seeing and nothing else. The training plan behind it is rules based, drawn from human expertise, because what a fourteen year old should work on next is a coaching judgment and not a pattern matching problem.

The wider conversation has caught up to this framing. Provenance work has moved from watermarks and detection toward attribution, and the useful distinction people have landed on is that generated and assisted are not the same thing, and that the question worth asking is who supplied the judgment. Standards bodies and regulators are building machinery to record that. We think the machinery is good and insufficient, for a reason that has nothing to do with cryptography.

A provenance record can prove a file has not been altered since signing. It cannot tell you whether the person who signed it read it.

That gap is where our field actually lives. We know this because we fell into it. This site publishes a check that reads our own copy and blocks a deploy if the writing has drifted. It passed for a week while the drift was sitting on live pages, because its scope named one file and matched one file and reported that as coverage. The check was signed, versioned, honest about its own methodology, and looking at almost nothing. Every property a provenance system would have verified was true. The claim it implied was false.

So the discipline we ended up with is narrower than a standard and more annoying to maintain. Every rule has to be shown failing before we trust it. Reintroduce the exact defect, watch the count move, put it back. Anything that claims coverage has to be able to state what it covered, out loud, in a list you can read. A green result whose scope nobody checked is worse than no check, because it stops people looking.

None of that is provenance in the technical sense. It is closer to a lab notebook. It does not prove we were right. It proves what we examined, which is the only claim any of us can honestly make.

If you want to know whether a person supplied the judgment behind something, a signature will not tell you and a detector will lie to you. What tells you is whether the thing is specific. Whether it says what it does not do. Whether it names a number and where the number came from. Whether anyone wrote down the parts that make them look bad.

We would rather be held to that standard than to a badge, and we have tried to publish enough of our own failures that holding us to it is possible.