altme.org

Every assistant you use forms some picture of you. It has to; it cannot be useful otherwise. What you can establish from where you sit is less than what it holds: that something is retained, that you were not asked for it, and that you cannot take it with you.

Three things are wrong with that, and none is repaired by making the picture better. It is thin, because what one application has reason to learn about you is a small part of what is true about you. It is silent, because at no point did anyone ask. And it is not yours, because you can neither see it, nor move it, nor end it.

The asymmetry is worth stating exactly. You are the only party present in every one of these conversations and the only party who carries nothing across them. Each one begins again, and the cost of beginning again is paid by you, in repetition.

A picture a person cannot inspect cannot be corrected by that person, and a wrong inference does not merely persist, it compounds, because what comes later is built on what came before. So the argument here is not primarily about privacy. It is about who is permitted to be the continuous thing, and who is permitted to correct the record.

PrepGraph’s own products are among the applications this criticism is aimed at. Whatever they hold about a learner is not the learner’s to inspect, to move or to end today, and the demands set out below are ones we do not currently meet.

A draft open protocol, at version 0.1, published as a request for comments. Its subject is a memory that belongs to the person rather than to an application, and what an assistant should have to observe in order to be let near one.

It standardises one thing: the terms on which an agent is granted access to a memory belonging to somebody else, namely the person. It does not standardise what a memory is for, what an assistant ought to conclude from one, or what makes one good. A protocol that answered those would be answering questions it has no standing to answer.

Existing open work has settled how an agent reaches context that belongs to a service. Nothing settles how an agent reaches context that belongs to the person, on that person’s terms and revocably. The draft is about the second question and takes no position on the first.

What follows are the obligations the draft is trying to place on whoever holds a person’s memory. They are written as demands a person, or a school buying on a child’s behalf, could put to any supplier and expect a straight answer to. They are the whole of the proposal at this level; how such obligations would be met is not described here and will not be.

A property that publishes no methods is publishing a protocol in full. The two are different kinds of thing, and the difference is the point of this section.

A protocol is an agreement, and an agreement with one party to it is not an agreement. The obligations above become obligations only when somebody answerable to nobody here can implement them, and a person can move between the two implementations without asking either.

The interior is not published and will not be. The research on academic intelligence, and the way anything at PrepGraph is built, remain unpublished. We publish the interface and withhold the interior, and would rather say that plainly than let the openness of the one imply an openness of the other.

Our interest should be named as well. We would rather build on an interface we do not own than own one that nobody else uses. Whether that is more than a preference is settled by three things, none of which a reader should take on this page’s word. The licence, which travels with the draft and is not restated here: it is worth nothing unless it lets anyone implement, extend or fork the work, including into a product that competes with ours, without asking us. Custody: the draft is held today by the company that wrote it, which is the wrong place for it to stay, and no body outside the company has yet agreed to take it. Survival: if this company does not last, what has been published should remain implementable by whoever wants it, and nothing beyond that can honestly be promised.

A learner’s context is the clearest case of something that should not sit inside one vendor. What a student has understood, where they stalled, the misconception they have carried for two years, the language they think in before answering in another one: today that belongs to whichever platform happened to observe it.

What moves with an Indian student is the mark. Marks transfer between schools, boards and tutors; they are the one portable artefact in the system, and they are a late scalar that cannot say what a student should do next. A teacher who inherits a section of fifty-odd in April receives last year’s marks and almost nothing about how they came about. The compressed summary travels and the understanding does not. That is backwards.

A learner should be able to carry their own context across a change of teacher, platform, medium or school. That is a position, not a result, and we are not in a position to present it as one.

altme.org is at the stage of a draft protocol, version 0.1, published as a request for comments, alongside a reference implementation whose authors describe it as not yet production-grade, with hardening work outstanding. Neither is finished. Nothing about either has been measured.

It is not ratified. No standards body has taken it up, no organisation other than the one that wrote it stands behind it, and nothing about it has been certified by anybody. There is no adoption to report, and this page would rather say so than describe interest as though it were uptake. The draft will change, the reference implementation will change with it, and anything built against either should expect to break.

Comments are invited where the draft is published. The useful ones are not corrections of wording. They are the ones that say a guarantee is stated wrongly, or that no honest implementation could keep it, or that the whole shape is wrong and a simpler design reaches the same promises to a person. If somebody else’s design achieves these guarantees more simply, the right outcome is that this draft is abandoned in favour of it. That outcome would be a success, and it is worth writing down now, while it is still cheap to mean it.