The mechanics, plainly.

From photographs to points

When photographs enter the archive — on the desktop, on your own machine — each one is read by a vision-language model (SigLIP2), a model trained to place images and text in the same space of meanings. Each photo becomes a point in that space: photos that mean similar things sit near each other, and a phrase lands among the photos that answer to it. No tags, no albums, no filenames — the geometry is computed from the pictures themselves, locally.

From points to a map

That space has hundreds of dimensions; the screen has two. The atlas projects the points down to a plane once and keeps the projection. The result is the field — every photograph’s resting place, similar ones near each other. Because the projection is computed once and reused, the map is stable: the same archive always lays out the same way, the same phrase always finds the same region, and close ideas sit near each other. Photographs you bring in later project into the same map — the layout never refits under you. The plane is tiled without gaps — every photo holds exactly one tile, and there are no empty patches.

Naming gathers

A subject is anything you name. Name one, and the photographs that answer to it gather into a single connected region, opened at the spot where the subject’s best example already sits in the field. Anchoring at a real member matters: the average position of scattered photographs lands in the meaningless middle of the map, while a real member’s position lands in the subject’s true neighborhood.

The edge is exact — every member inside, every non-member outside — and just beyond it sit the subject’s genuine neighbors in meaning. Inside, members need not look alike: a person’s region holds the same face across decades of different scenes. Make a subject permanent when you care about it: it becomes a landmark you can summon anywhere by name — @bird, @mama — and place again in any narrative.

People are found by face, not by name

Faces live in their own space, separate from meanings. The word John would gather every man in the archive; your John is matched by the geometry of his face across your photographs. Naming a person means assigning faces to a name. The two spaces are never mixed — a face match and a meaning match are different kinds of evidence, and the atlas never pretends one is the other.

Contrasts are dimensions

A contrast is, strictly, a named dimension: the tension between two concepts — two poles. Warm / cool, near / far, remembered / forgotten. Any pair of phrases can stretch one, so there is a continuum of possible dimensions, not a fixed menu. Name one or two and they map directly onto the screen’s axes — the horizontal is warm-to-cool. Every contrast also carries a band: one stretchable range along the dimension. Leave it wide and the whole tension shows; narrow it and the view holds to just that stretch. The poles bend where every photograph sits — the open field and the named regions alike, though never at the cost of a region’s integrity: a member may drift within its subject’s region, but it never leaves it. Contrasts do double duty: in Find and Create they bend where photographs sit; in Remember and Experience they set the order photographs play.

The spell language

What you type into the Narrate panel is a spell — a statement in a small language the whole app speaks. Bare words name meaning: the sea at sunset. The operators and, or, not combine them: mama and papa, parrot not toucan. Three markers do the rest. @ summons a subject you have already named — @mama, @bird. ! states an invariant, a fact that must hold — !brightness 20..80 keeps the view inside a tonal range. # works an affordance of the instrument itself — #mode remember switches the view the way the mode picker does. And vs stretches a contrast — warm vs cool 20..45 names the dimension and narrows its band in one line.

Spells stack in the panel, and the stack is the narration the field answers to. Each line reshapes the field further; silence a line and its force is withdrawn. The same statements can be spoken by you, by a saved narrative replaying, or by the resident guide — one language, three speakers.

Four modes, two geometries

The modes are verbs: find — the map; remember — the timeline; experience — the story wall; create — the canvas. Underneath there are two geometries. Find and Create are space-like: the same field, arranged — one stable, one expressive. Remember and Experience are time-like: the same photographs laid into strips and stories and played in sync, ordered by capture date or by any contrast you set. Two readings of one body of photographs, not two collections.

No edges

The field is a torus: pan past one side and you arrive at the other. The world wraps — there is no edge to fall off and no privileged center. Only places, held stable.

On the web, privacy is a property, not a policy

The server only ever holds what you chose to publish, at the resolution you chose to publish it — never your originals. Full-resolution photographs never leave the desktop. From there, sight is a matter of relationship. Anyone may walk the field, but meets each photograph at a few dozen pixels — enough to tell it’s a photograph, even that it’s a face; never enough to look closely, and with no path to anything sharper. The invited see an archive at its published resolution. The caps are built into what exists on the server and what each visitor is served — not into a setting that could be flipped. What you publish is the shape of your attention, not your negatives.

Vocabulary

archive

All the photographs, treated as one body; a record of attention.

field

Every photograph’s resting place on the plane, similar ones near each other.

subject

Anything you name. Make one permanent and it becomes a landmark you can summon by @its-name in any narrative.

person

A subject identified by a face rather than by words; named by assigning faces.

spell

What you type to name and combine subjects: bare words, @subjects, and / or / not, ! invariants, # affordances, vs contrasts.

contrast

A named dimension — a vs b — with a band you can narrow along it.

relevance

The one dial between the full atlas and the exact slice: at zero, what you named lights up inside everything else; turned all the way up, only what you named remains.

narrative

A saved composition of spells, contrasts, and view. Return to it, replay it, record it frame by frame, and export the recording as a movie.

mode

One of the four verbs: find — the map; remember — the timeline; experience — the story wall; create — the canvas.

sigil

A name bound to a structure — the app’s oldest word. It survives in the names: SigilAtlas, and SigilML, the spell language.

atlas

The archive seen as a stable map; the map seen as a world you stand inside.

The vocabulary is a lexicon

These are the plain words — the default, chosen for a reader the instrument hasn’t met yet. Underneath them sits an attention language: a few concepts with deep reach — contrast, relevance, invariant, narrative, sigil — of which the plain vocabulary is one rendering. Coming: declare who you are — physicist, linguist, semiotician — and the instrument’s chrome renames itself into your discipline’s terms for the same structure, each term glossed back to plain. The spell language never forks: whatever the labels say, @, !, #, and vs mean what they mean.

The apps, in detail →