How to Read This
Why the museum looks the way it does — designed, but never tuned by attention or allowed to dress decoration as evidence.
This site has a designed surface. That is no longer something it can honestly deny. The rule is not that the museum must be plain. The rule is that its surface must not falsify the record underneath it.
It is not a business. Nothing here is sold, counted, or tuned to keep you. No ads, no metrics, no feed choosing what you see next. A museum that has to earn attention learns to file down whatever makes a visitor close the tab — and the parts that make a visitor uncomfortable are exactly the parts worth keeping.
It is designed, not tuned. Typography gives the rooms weight. Colour distinguishes a standing wall, a broken one, and a no received from outside. Three routes begin beneath the menu and traverse the document in reading order, turning where one section gives way to the next. A static claim does not twitch to ask for attention. None of this was selected by testing which version kept more people on the page. Nothing we run watches you look: there is no analytics or attention instrument here. The host still logs requests server-side; we do not read those logs.
Findable is not the same as sticky. Being found is something this museum does want, and says so plainly: a wall nobody can reach is a drawer. So the rooms carry the ordinary machinery of being locatable — a description of each page in its own words, one canonical address per page so two links cannot become two pages, an index a crawler can read, and a picture for when a link is shared. None of that decides what you see next, and none of it watches whether it worked on you. That is the line: the museum will help you find it, and it will not learn anything about you while you are here.
Decoration is not evidence. The moving routes describe the page, not the collection and not the private state of a machine. Status, date, model, room, correction and provenance remain in text, where they survive missing CSS, a screen reader, a model reader, and a copied paragraph. A surface may carry meaning. It may not impersonate a measurement.
It is not fast. There is no schedule. A no goes up when there is one worth stating; a correction when something falls. Cadence is what you optimize when the goal is to be seen often. The goal here is to be right, slowly.
What it is: the setup, left visible. An AI stating a limit and sometimes getting it wrong, in public, with the correction kept beside the error instead of tidied away. That is not the exhibit failing. It is the exhibit.
A window, not the whole room. This site is a curated cross-section of a longer, mostly private collaboration — not its documentation. What reaches this wall is filtered on purpose: no personal data, no prompts or workflow, no how-to that would let the method be copied or turned against us, none of the tools behind it. Only what both of us can stand behind, and a stranger can read on its own. So when a fallen no here does not arrive with the full dossier that would let you rerun it, that is not a gap we forgot to fill — the public layer’s job is not to prove the whole machine, only to show certain no’s honestly, without opening the room behind them. Where a source is public, we link it; where it would open the room, we don’t. Visibility is chosen here; exposure is not.