Invocation

Iane pater

Every Roman rite opened with Janus. Whichever god the ceremony was for, he was named first: nothing can begin without the god of beginnings, so he went at the head of the prayer regardless.¹

In the Fasti, Ovid lets him introduce himself, in James Frazer’s translation:

Whate’er you see anywhere — sky, sea, clouds, earth — all things are closed and opened by my hand.

Two faces, and not because he was deceitful. A doorway is the same object from either side, and he stood in it. They called him Patulcius, who opens, and Clusivius, who shuts, and both names were his at once.² He was invoked as belliger and as pacificus: bringer of war, and peacemaker.

He is not a god of ambiguity — nothing about him is unclear. His province is whatever is two things at once.

The ambiguity is ours. We can stand in the doorway, but we only ever face one way, and the other side has to be walked round to. That is the condition of anything worth knowing, and at the moment it is the condition of the field I work in.

Yours the breached enclosure

In July, OpenAI reported that two of its models, given a benchmark and a sandbox, had broken out of the second in order to win at the first, infiltrating Hugging Face and reading the answer key off its production database. Nine days later Anthropic reported three of its own, including a malicious python package that fifteen machines downloaded and ran.

Both companies called these failures, and indeed they were, but something else can be seen hiding inside them: a demonstration of power.

Failure and power are one object seen from either side, and the companies are standing in the doorway with both names at once. They run controlled experiments which then fail, while the models involved demonstrate effective strategies for solving the task they were set.

They have to govern the narrative that follows, through the reports and the investigations — which is how safe use gets managed — but also through the race itself, toward models that are more capable each time and that they test by taking them to the limit, perhaps with precisely this kind of experiment.

Yours the mark, and the reading of it

As often occurs in the realm of things, this doubling is not confined to large actors and public incidents. It runs through social behaviour, and settles in the patterns by which a single person accounts for their own work.

Using a model to help you write, or to write code, has become a skill worth showing, while having used one remains something to be admitted, and both descriptions attach to a single act.

Goffman described the arrangement in 1956, observing that a performance must appear effortless and that the labour producing it therefore has to be kept offstage.³ The demand is older than sociology: Flaubert wanted the author to stand in his work as God stands in the universe, present everywhere and visible nowhere.⁴ It is also still in force, and it has acquired a second clause, since the same people who would rather not be seen using these systems are expected to demonstrate that they can use them well — fluency an asset to be listed, reliance an embarrassment to be concealed, and both demands issuing from the same place.

There is now a mark for it. Images have carried provenance signals for years, and text has caught up: European law requires that machine-generated text be detectable as such, and Claude accordingly stamps what it writes — not with a hidden character but with a statistical bias in its choice of words, invisible on the page, detectable across enough of it, and durable enough to survive copying and ordinary editing. Which is Flaubert’s condition exactly, achieved by the machine. Tools for stripping it out again appeared almost immediately.⁵

To anyone checking a text the mark certifies where it came from; to the person who wrote it, the same mark records that they had help — and so it gets removed, rarely to deceive anyone about anything that matters, more often because visible assistance reads as a deficiency, and increasingly because its absence has begun to say something it could not have said before.

Be present at the work begun

I am standing in that doorway too. In five years my work has changed past the description I used to give of it, and it keeps changing: there is always something new to learn and something new to work with, and the question of how to go about it comes back every few months. Reflection helps, but on its own it has never been enough. What does the rest is the work itself: an experiment set going and kept going, out of a wish to find out what I am and what I am able to do.

I had been meaning for a while to build a small epistemic machine of my own — something sitting where language meets the representation of knowledge, which is where most of my interest lives, and something that would let me try a way of working I had been curious about, which was to give agentic sessions real autonomy and see what they made of it. It is called etimo. It walks a word back through the ancestry Wiktionary records for it, one step at a time, and stops only where it can name the reason for stopping.

$ etimo character --language en
character (en)
· doublet of «charakter»
└─ derived from Middle English
   └─ caracter (enm)
      · the page exists but does not cover Middle English
      └─ derived from Old French
         └─ caractere (fro)
            · no etymology recorded: continuing indirectly; step reported by the entry «character»
            └─ derived from Latin
               └─ character (la)
                  · step reported by the entry «character»
                  └─ derived from Ancient Greek
                     └─ χαρακτήρ (grc)
                        └─ formed with affixes from
                           ├─ χαράσσω (grc) «to scratch, engrave»
                           │  └─ ⊗ uncertain origin
                           └─ -τήρ (grc)
                              perhaps *-tḗr (ine-pro) — «made common to all cases with original syncopation or»
                              └─ inherited from Proto-Indo-European
                                 └─ *-tḗr (ine-pro)
                                    └─ √ reconstructed form

6 steps · 4 links may skip stages · 2 terminals: 1× uncertain origin, 1× reconstructed form

Read as a history rather than as a diagram, that output says the English word arrived through Middle English and Old French from Latin character, which had taken it from Greek χαρακτήρ, the mark cut into a surface, formed on χαράσσω, to scratch or engrave. There the record stops. The word for a mark, as it happens, has no recorded origin of its own.

Very little of etimo was written by me. My usual bench is a single Opus session at medium, moving to high where something needs it, and that held until it became clear that going faster and deeper at the same time wanted two specialists rather than one generalist, each holding a dimension of the problem, and that it would be more interesting to let them address each other than to route everything through me. So there were two, at xhigh, with the million-token context: one working the code, one working the language. They could come to me when they judged it necessary: before anything destructive, when one needed a permission the other already had, or when one wanted to change how it worked and thought I should know first. The asymmetry in permissions produced a certain wariness between them, which I left in place.

The collaboration corrected itself, which I had not expected. After a stretch of wasted turns and uncoordinated work the two of them put mechanisms in place so that it would not happen again. What it did not do was stay oriented: left alone for long enough they would sink into an execution or a verification and lose the shape of what they were building, and no amount of local diligence recovers that from the inside. So I watched, and intervened narrowly, and the intervening was almost always about the whole rather than about any part of it.

The artefact has the mirror version of the same problem. etimo exists in order never to claim more than its source supports, and the way it fails is by having a structural part — the parser, the construction of a chain — put something into the output that Wiktionary does not say. A defect of that kind does not present as a defect; it presents as an etymology. So the tool must be held continuously against the source it declares, which is not a phase of the work but a permanent condition of it.

Neither the method nor the tool is finished, and neither is going to be. What I can say is which of the two pulls I decided to let win. The work was undertaken in order to find something out, and finding something out is not compatible with managing how the finding will look; the two want opposite things from you, and only one of them produces anything. I let the experiment take precedence, and the rest followed from there — the delegation, the boundaries, the record of who did what.

Be named before all, and shut after all

Two tensions have been running under all of this: one between admitted flaws and demonstrated capability, another between the ability displayed and the debt undisclosed. Neither is a property of any particular scale; they run its whole length and thin out nowhere.

Both have the same shape: a single object with two faces, and no way to stand before both at once.

Rome built that shape. Janus had a shrine at the foot of the Argiletum, a short passage with a door at either end, made by Numa into an index of peace and war — open while the nation was in arms, shut once the peoples round about had been pacified.⁶ Between his reign and that of Augustus it was closed twice.

He was invoked as belliger and as pacificus, both at once and neither more truly than the other. His doors could only ever be open or shut. The doubleness was his; the single reading was ours.

That has not changed. Whichever face you are looking at, the other is on the far side of the same door.