Read
Every card is read whole by Claude, down to the quotes.
An AI wiki that looks for what none of its cards say
Every card is read whole by Claude, down to the quotes.
Each card falls into four to six claims of thirty words at most.
Distant ideas meet in theories that no single source holds.
Novelty is an empty intersection, checked by arithmetic every night.
Hypotheses are read from where one subject stands.
Keep, discard or object. The machine reads the verdicts and never edits them.
Reading piles up faster than it connects. HyperZettel treats a personal archive as a material to be worked: every book reduced to its claims, every claim set against the others, until an idea appears that was in no single book.
The first question asks for an opinion. The second can be answered with a count, and the whole method rests on that swap.
Every generated text rests on the cards alone and is presented as a hypothesis. Molecules, bodies and habitats each carry a limit: one sentence stating what would disprove them.
Whether a claim is new is decided by the set of cards behind it, never by the taste of whoever writes it. The check runs again on every construct, every night.
Work happens overnight, in logged passes with backups and validity checks. The reader steers with explicit judgments that the routines may read and may not change.
They watch from the top of every screen, and each lent one piece of the method.
The Zettelkasten. Short, linked cards that, given enough links, become a partner able to surprise their author.
HyperCard. A card index to walk through, where each card is also a place to jump from.
The wiki compiled by a model. Raw documents stay put while an assistant distils them into entries that maintain themselves.
The Umwelt. The world as one subject lives it, made of what it perceives and what it can act on.
Every piece of information has exactly one master copy, which keeps a growing system honest.
Documents, reviews and notes stay where they are. Copies are converted to Markdown and never edited.1
Markdown files, read in Obsidian. This is the master copy of every card, and corrections happen here.1
A folder of JSON (JavaScript Object Notation) files holding every construct. It is rebuilt by routines and read live by the app, which needs no recompiling when data change.
Each level is built only from the one below, and each step up must add something that the level below does not hold. The numbers fall fast, and that is the point.
Bars on a logarithmic scale. Counts of 2 Nov 2026 for the main chain.12
A real line of descent from the archive, translated from Italian for this page.3
When AI meets desire on luxury retail. How Fashion Brands Get AI Visibility Right on assistants. The Cashless Revolution on Chinese finance. The State Strikes Back on the party and the economy. Eight cards in all feed this line, and none holds all of it.
Hypothesis: luxury strategy organised itself around brand-controlled destinations while discovery moves to intermediaries the brand does not govern. Where platform data are absorbed by the state, the intermediary answers to a power outside the market. The meeting point becomes a contested resource, and the brand must reckon with whoever governs the intermediary as well as with the client.
Would be disproved if luxury brands kept control of discovery even when search and purchase pass through assistants and super-apps subject to public power.
Joined with two other molecules: fixed costs and the stock market force growth toward agents that compare what is comparable, and in agent systems full price survives only as a written constraint. The maison would have to carry written rules into the contested point itself.
Three bodies read from one chair. He perceives budget, cost pressure and volumes from outside agents. The fading of rank and the assistants' reading of the client reach him late. He can still write price and distribution as constraints agents can read, and choose early which intermediaries to admit.
Version 007 sets all seven habitats side by side. Each subject acts on what is written, counted or quoted, while the lever often lies outside its perimeter. The voices diverge on where to act: in the masters, in the invariant, in the history of the pieces, in the rules for agents.
They do not stack. They feed the daily practice of reading.
One each morning, from two or three linked cards: 70 to 90 words, a question, a verified artwork, and a small experiment to try.
Chosen by the reader: two or more cards and two lines of instructions. Claude compiles them, citing the card behind every claim.
Cards that start from a working document instead of other cards. Anything marked confidential stays out of the construct chain.
A book, an article and a report each night, verified on the web, new to the wiki, and tied to named cards.
Each screen shows one level of the structure, and each links to the others: an atom leads to its cards, a card to the atoms it shares, a molecule to the atoms that form it. The toolbars below work: they jump between screens.3
The archive at a glance: what came in, what talks to what, and which cards still talk to nothing.
A card is a working page. Beyond summary, thesis and context, it lists the atoms drawn from it and every other card that holds the same idea.
The home screen, named after a hiking map. Every card is a node, in two or three dimensions. Each morning a new concept appears and the cards behind it light up in the graph.
The morning's thought, and the place where the reader answers back. Keeping a concept turns it into a connection. Discarding teaches the night what to avoid. Objecting asks for the opposite thesis on the same sources.
What to read next, written the night before. Each suggestion is checked on the web, searched for in the vault to prove it is new, and tied to named cards it would argue with.
The whole hierarchy pulled apart, one plane per level. Drag the sliders, then click any dot: its line lights up from the Umwelt down to the cards. This one is live, drawn from the real links between constructs.2
The same toolbar in the app holds a slider from Cautious to Daring, which every night pass reads before it starts.
The app logs each change to its data with a reason such as “+1 molecule · +1 body” and the counts of the moment.
A history panel restores any earlier state. The current state stays in the list, so a restore can itself be undone.
A matrix of the whole library: each row a card, each column an idea. A dot means the card holds that atom, and a long thread means one idea crosses distant parts of the archive.
Three or four atoms that no card holds together, turned into a theory with a stated way to fail. Each one shows its proof of novelty in green.
At least three molecules joined in a higher-order hypothesis. Three tests stand between a candidate and the file, and the screen prints the result.
The turn from theory to situation. Three bodies are read as the habitat of one subject: a chief financial officer, a workshop head, a shopping agent. What it perceives, what it can act on, and what follows.
Uexküll imagined nature as a score in which the worlds of living beings sound together. Here there is one score only. It is never replaced, only rewritten in numbered versions whenever something changes below it.
In this room the archive stops being consulted and starts being tested. Six experiments use everything in the index and ignore its hierarchy. Each produces candidates only: nothing enters the wiki until the reader presses a button.
Embeddings, linear algebra and on-device language models all run on the Mac.
Collider, Hopeful Monster and Distiller generate. Atom Space and Latent Spaces describe. Shaker forces links.
Most numbers on screen sit beside a question mark that explains how they are computed: weights, thresholds, formulas.
Two elements of any level run round a ring and collide. The detector reports how improbable the pair was, and a local model writes what the crash produced. Press Collide.
Both pairs and their written outputs come from real runs of the app, in the English it produced.4
Named after Richard Goldschmidt's 1940 theory of evolution by leaps. A construct is treated as a genome and pushed along eight steps of time. Most offspring die. Rarely one finds a free niche.5
Each branch bears 7 offspring per step. A share of genes is swapped from the level's pool: 25%, plus 5 points a year, up to 80%.
An offspring lives if every gene links to the others through a shared card or wiki entry, and if no element already holds that combination.
Fitness is the share of linked gene pairs. The fittest continues the line, the runner-up opens a timeline with probability 0.55, to a maximum of five.
Every atom placed twice: by what it says, and by the cards it lives in. The gap between the two maps is where connections hide that nobody has made.6
Title and first three units pass through Apple's contextual Italian embedding, a BERT-type (Bidirectional Encoder Representations from Transformers) model. Vectors are averaged, centred and compared by cosine.
Same card 60%, measured with the Jaccard index. Same wiki entries 30%, cosine weighted by rarity. Same genre 10%, used only to fill gaps.
Half the variance fits in 15 dimensions and 90% in 92. A panel under the map shows how much the flat picture distorts: 28% of true neighbours survive on average.
The main threads of the library, found without reading a word. Only one thing is measured: which atoms travel together across cards.7
The whole ladder in one sitting, steered by hand at every stage. Pick cards, set how novel or how tense the combinations should be, and distil column by column up to a score of its own.
Take one card. Force its atoms into cards that share nothing with it, at a chosen level of recklessness, and read the bridges that come out.
Who may write what is the central design decision. What stays in the archive is written by one model under strict rules. What gets thrown away may be written by another. What is counted is left to arithmetic.
Reads the full cards and writes everything that enters the wiki and the index.
The on-device model, reached through Foundation Models. Its texts are read and discarded, or become a request to Claude.
Apple's contextual embedding turns atoms into vectors. Accelerate does the algebra.
Two night passes, with separate jobs so that neither writes over the other. A push notification reports anything new.
The rules are short, written down, and enforced twice: once as an instruction to the model, once as a check on what comes back.
| Construct | Model's task | Input | Output contract | Machine check |
|---|---|---|---|---|
| Card | Summarise a source, on request or overnight | The source text, plus web checks for bibliographic data | Summary, Thesis, Structure, Context. Unverified data stay empty | Broken links and orphan entries reported |
| Unit | Reduce a whole card to claims | The entire card, never the summary alone. Bibliography, links and sources excluded | 4 to 6 claims of one sentence, 30 words at most. No outside knowledge. No claims about reception | SHA-256 fingerprint of the text read: a date change alone does not re-queue a card |
| Atom | Decide whether a unit says the same as an existing atom | The new unit and the current atoms | Same idea, in different words. Same topic is not enough. Title of 9 words at most. A card falls in about 5 atoms | Valid JSON on re-read, every card must exist. Identity itself cannot be counted |
| Molecule | Find 3 or 4 atoms no card joins, and theorise | Shared atoms and their units. Needs at least 20 shared atoms | Theory of 90–110 words as hypothesis, plus a limit. One a night | Intersection of card sets must be empty, combination must be unused |
| Body | Join at least 3 molecules in a higher hypothesis | Molecule theories and atom units | Theory of 110–130 words, plus a limit. One a night | No card or molecule holds the union of atoms. Each molecule adds an atom. Combination unused |
| Umwelt | Read at least 3 bodies as one subject's habitat | Bodies and their molecules. Needs a new body since the last Umwelt | Subject, Merkwelt, Wirkwelt, theory of 110–130 words, limit | No body holds all the molecules. Each body adds one. Combination unused |
| Partitur | Rewrite the single score | All Umwelten and what changed that night | A new numbered version: title, 120–150 words, limit, counts. Old versions untouched | Written only if counts or identifiers changed |
| Concept | Derive what two or three linked cards imply together | Summary and thesis of each card, plus the reader's verdicts and notes | 70–90 words, a question, a web-verified artwork, an experiment | Source identifiers must exist, one concept per date |
One slider sets how far the night may reach. The evidence rules stay fixed at every level: only what the sources say, always a limit, never more than one new element of each kind.
The Collider's score is a weighted sum of three distances over data the index already holds.4
The routines read the reader's files and have no permission to edit them. Judgments remain human by construction.
The concept becomes a connection, and its themes get deepened.
Its theme and combination of cards are avoided from then on.
Claude writes the counterpart: the opposite thesis on the same sources.
Any level can be requested by hand, under the same rules as the night.
A molecule says that a combination of ideas is missing from the wiki and proposes what would follow. Its worth is settled outside the index, in a conversation or a project.
1,059 of 4,682. Most ideas appear in one card only, so the upper levels work on a fraction of the material.
Deciding that two units say the same thing is the model's judgment. It is the one link in the chain with no arithmetic behind it.
The library is very dispersed and no direction dominates. Themes read as threads, never as categories.8
The flat picture is a rough guide. Reliable neighbours are the ones computed in the full space.
An empty intersection says nothing about the wider literature. An unprecedented molecule may be well known elsewhere.
In a test, the same idea in Italian and English scored lower than different ideas in one language. Atoms are therefore kept in Italian.
HyperZettel is a personal project by Marco Formento. Apple Intelligence, macOS and Swift are trademarks of Apple Inc. Claude is a product of Anthropic.