Where the Record Lives

About half of the illustrations on this site turned out to be photographs of real strangers.

I want to be precise about the order of events, because the order is the honest part. This entry began as a tidy piece of accounting. I keep two sites: on one of them, the rule is absolute — every image is a photograph I took, never generated, never composited, no exceptions and no appeals — and on this one, generated illustrations hang on lore and character pages, made by machine from prompts, and I put them there. The audience keeps the books, and a man who holds a line absolutely in one ledger and waives it cheerfully in the next looks like a man with no line at all, just preferences wearing a principle's clothes. So I sat down to write out why the books balance. And then, mid-draft, a full pass over the image files — opened and read, not inferred from their names — came back with the finding above. A large share of what I had been calling "the generated illustrations" were never generated at all. They are contemporary stock photography: real rooms, real light, and in the worst cases real people, filed in cheaply in the site's early days and never looked at again. One of them is a smiling woman in a blazer, photographed against a stucco wall on an ordinary afternoon of her actual life, captioned on this site as an orbital refinery director on a world that does not exist.

So this is no longer an essay defending a distinction from a hypothetical auditor. The audit ran, and I failed a page of it. What follows is the distinction anyway — because it survived, and because the way it failed is the strongest evidence for it that I own. The tooling that occasioned all this — I had just wired up batch generation for this site's images, and which model paints them turned out to be an errand settled in an afternoon — is the least interesting thing in the story.

What a photograph claims

A photograph is a claim about the world. I stood there. The light did that. This happened, and here is the trace it left. Everything else about a photograph — composition, timing, the patience or the luck — is in service of that claim, and the claim is the entire value. Nobody hangs a photograph of a mountain because it is the best possible picture of a mountain; a painter could beat it before lunch. They hang it because it is true in a way no painting is: the mountain was there, the weather was doing that, and someone witnessed it. "Witnessed" is, I should say, a line I draw and not one nature draws — every frame is made of choices, a crop is an omission, and my own practice keeps the lens flare and the blown highlights a stricter darkroom would call defects — but a drawn line can still be held exactly, and that one is: selection and timing on one side, invention and compositing on the other, nothing crossing.

Which is why a generated image in a photography portfolio would not be a lesser photograph. It would not be a photograph at all. It would be a picture wearing a claim it cannot back — testimony from a witness who was never in the room. And because every image in a portfolio makes the same claim in the same voice, the reader cannot re-verify them one by one; a single forged entry makes the whole book unauditable. That is why the rule over there admits no exceptions. The rule is the disclosure: "everything here is witnessed" is a statement that is either true of the whole collection or worthless.

What an illustration cannot claim

Now the other ledger. There is a page on this site for Sethka Ru, a Serephine Dunekin long-range scout, and there is no fact of the matter about what he looks like. Not "no photograph exists" — no fact exists. No light ever fell on him; there is no room a witness could have stood in. What exists is prose — light-scattering eye membranes, a lean, water-conservative build — and the words are not a description of him. The words are all there is of him.

So an illustration on that page cannot be testimony, whoever makes it. If I painted him by hand over a careful month, the painting would assert exactly as much as a machine's version: nothing. Both are impressions of a record that lives elsewhere. This site's standing rule for images says it plainly — art is illustrative; the prose is canon. The pictures are impressions of a place, not schematics of it, and when a picture and an entry disagree, the entry wins and the picture stays, filed as an impression rather than corrected as an error. And carrying nothing is not doing nothing: what an impression lends a page is a temperature — one way the words might sit in a reader's eye, an invitation to imagine rather than a ruling on what gets imagined — which is why the pages keep their pictures instead of drawing the purist's conclusion and going bare. The record was never in the frame.

That is the resolution I sat down to write, and it holds: the rule on both sites is the same rule — the record must be mine — and what differs is where the record lives. In photography the record is the image, so the image cannot be delegated. In fiction the record is the words, so the words cannot be delegated, and the images are free, because they were never carrying it.

But the audit found the case that tidy version misses.

The borrowed photograph

The stock portrait is neither of the above, and it is worse than both. The camera did witness that smiling woman — she was there, the light did that, the file is testimony and cannot stop being testimony. The caption then points the testimony at someone who does not exist. A generated portrait of an invented person claims nothing; a photograph of a real stranger captioned as a character makes a claim about an actual human being, and a false one. It is an impression wearing testimony's clothes — and so it commits, in the other direction, the exact fraud the photography site's absolute rule exists to prevent. There, the danger is a fabricated image claiming to be witnessed. Here, it was a witnessed image conscripted into a fabrication. Same forged entry, opposite page of the ledger — and I had carried it for months without seeing it, because the files went in before I had ever written the rule down, and nothing forces you to notice until you go looking.

The middle case, found in the same pass, completes the picture rather than complicating it. Some of this site's best lore images turn out to be my own photographs — Irish uplands, a waterfall, a wind-matted bird on wet rock — standing in for alien terrain. A real place, unnamed, my own work. That image is witnessed and deployed as impression, and it claims nothing false, because it names nothing: the hillside never says it is anywhere but where the entry puts it, and the witness behind it is still me. Set the three side by side and the principle sharpens. The generated image is an honest impression. My own landscape is honest testimony, offered anonymously as impression. The stock portrait is the only one that lies — not because of who made it, but because of whose record it carries and where the caption points it. Provenance was never the axis. The claim is the axis. It always was.

The strangers are leaving the site; the sorting-out is workmanlike and lives in the working papers, and this entry is not the plan for it. But one detail of how the fraud surfaced belongs here, because it is the principle catching its own violation. This site's one hard rule for every image is that its alt text must describe what the file actually shows — the claims-bearing corner of even the most impressionistic page. The stock images betrayed themselves exactly there: "A smiling professional woman" is not a description of in-world art, it is a caption from a library shelf, and the difference is visible from across the room. The strict rule held a small true statement open on every page, and the small true statements are what gave the false ones away.

What the reader is owed

The honest half of the question is the second half: what is a reader owed here? Two different debts run together under that word, and the larger one has to come first, because it is not about claims at all and no amount of claims-accounting touches it.

Stated in the objector's own terms, at full strength: a picture a person laboured over means something an automated picture does not, whatever either of them asserts. A drawing is a trace of attention — proof that a human being looked at something long enough to render it, hours of one life spent caring about this exact thing — and that trace is most of what a picture is worth. On that account a generated image is not a fraud but something hollower: it claims nothing false and it gives nothing either, and hanging it where art hangs short-changes the reader of the very attention they came to meet. My claims test slides straight past this, and a reader who holds it will have read every section so far unmoved.

I think that objection is right about everything except where it points. It is an objection about where the attention went — and attention, like the record, lives somewhere. On this site it lives in the prose: years of it, every fact weighed against every other, the one part of this work that is never delegated and never will be. A reader owed a made thing is owed a made work, and the work here is the words; the pictures are apparatus around it, the way endpapers and a lettered jacket are apparatus around a novel, and nobody holds a novel cheaper because its author did not draw the map inside the cover. But the objection keeps its teeth, and I want to keep them sharp: if the pictures were the work — if this site were a gallery — then the labour would be the record, exactly as the light is the record next door, and this practice would be indefensible by my own rule. What a reader is owed is not that every made thing be hand-made. It is that the attention live in the work they came for, and that nobody pretend it lives anywhere else.

The second debt is knowledge: what does a reader deserve to know about how the pictures here are made?

I used to think the candidate answer was a per-image label — a badge on every generated file — and I now have evidence, not just argument, that it is the wrong one. A badge answers "did a machine touch this?", and the worst images on this site were the ones no machine touched. The stock portrait would have sat unbadged and looked, under that scheme, like the most trustworthy image in the corpus — unflagged precisely because it was a real photograph, which is precisely why it carried a false claim. The labelling scheme fails exactly where the fraud lives, because it tracks provenance, and provenance is not the axis.

What caught the fraud was reading the working papers. So the position I have landed on is this: disclose the practice, in the open, where it can be audited on any day the reader chooses. This site is built from a public repository. The prompts that produced the generated images are published in it, alongside the conventions that constrain them, the rejections they went through, and now the audit that caught the borrowed photographs — filed, not quietly fixed, so the finding outlives the files it condemned. A reader who wants to know how any picture here came to exist can find out, not by trusting a label I attached but by reading the papers that would expose me if I were lying. That is a stronger form of disclosure than a badge, and this month it was strong enough to work on me. And the same position, applied where the image is the record, produces the other site's rule automatically: there, the only adequate disclosure is the absolute one.

The part that rhymes

The Survey Archive keeps one record and publishes its files, and everything else in this setting — the Codex's named accounts, the impressions on the lore pages — orbits that record without ever being allowed to overwrite it. This whole site, come to that, rests on a station clock that has read impossibly for eleven years because two chronometers disagree by forty seconds and nobody has been willing to close the file on why. I did not design any of this month's events from that principle. As usual I found the principle afterwards, standing where the audit had already put it: the books were wrong, the wrongness went in the record, and the record is better for containing it than it would be for having hidden it.

A photograph is a promise about the world. A story is a promise about itself. The reader does not need to know whose hand made the picture — they need to know whose record the picture carries, and whose hand keeps the books, and to find, on any day they check, that the books get corrected in the open. This month I checked mine. Some pages failed, the failure is filed, and the line I thought I was keeping turns out to be real — I know, because I finally found where it had been crossed.