Concept

Writing

Paste a text and read it two ways. Nothing is uploaded, nothing is saved, and no model runs — which is a design constraint rather than a privacy setting.

Writing sits at /writing and is the one part of this app you can use without an account. You paste a text into a room, and the room reads it. Each room has its own box, because the two want different kinds of text and a paste that followed you between them would arrive in the wrong room half the time.

The two rooms

Room	Reads	Finds
Stories	Fiction	Who is in your story, and where — the Braid, the cast, the pace.
Logic	An argument	Where the reasoning turns, and whether your terms stayed put.

Nothing leaves the page

The text you paste stays in your browser, and so does everything derived from it. There is no upload, no request to a model, and no row written anywhere. Close the tab and it is gone.

This is not a promise about restraint. The rooms are built so that they cannot send anything, because every instrument in them runs on a fixed list of words and a set of rules. That constraint is the next section, and it is the reason the guarantee is cheap to keep.

And if that ever changes. That may not be true for ever: some readings need more than a fixed list can give. If an instrument here ever does call one, it will be something you switch on for that reading — never a default, and never without the room saying so. Said here before there is anything to say it about, because the day a guarantee stops being true is the day it is most expensive to have overstated — and a visitor who believed this one deserves to have been told how a change would be made, not to discover it.

Why no model runs

An instrument only ships here if it can work from a closed class — a finite list of words, written down in advance, that anyone can read. therefore, because, hence and given that are all of the argument markers there are. said, told, according to and declined to comment are the attribution ones. Neither list has to be guessed at.

A name does. That asymmetry is the whole design, and it has a date attached: on 17 August a structural pass over a story returned Thursday and Tuesday as characters, because the only way to find a character is to guess which capitalised words are people. according to never needs guessing, so the room that reads for attribution stands on firmer ground than the room that reads for cast — and the rooms say so rather than pretending to be equally sure.

It buys less than it looks like. A closed class cannot invent a phrase, but it can match a real phrase in the wrong construction — and measured across 435,000 words, it did exactly that about two thirds of the time. The correction is a second deterministic rule rather than a model: the instrument checks what sits at the far end of the phrase before it counts the hit. Precision here is bought with more rules, and the rules are in the repository.

Where the method runs out

In one place, measurably, a closed class cannot carry the decision at all. The word pound means money or mass, and marking it ambiguous does not say which reading applies here:

Text	Hits	Every one of them
The Jungle	4	mass — beef, nails, a weighing scale
Pride and Prejudice	24	money — Regency fortunes

A list that excludes the word is right four times out of four on one book and wrong twenty-four times out of twenty-four on the next, and no edit to the list repairs it, because the word is doing both jobs correctly in both places. So the decision moves to a context rule — still deterministic, still model-free, kept separate and reported — which looks for the frames the two readings actually appear in: a pound of nails, a weighing verb just before it, or a bare per-unit rate.

That boundary was found by measurement rather than argued from first principles, which is the only way a boundary like it is worth anything.

It is also the honest reason the guarantee above carries a caveat. A reading that needs more than a fixed list can give is not hypothetical — this is one, and there will be others. The shelf already distinguishes what the structural pass can do from what it cannot, and an instrument that needs more than it says so on its own card rather than quietly reaching for it.

Getting a reading onto the canvas

A reading is yours to keep, and the route out is deliberately dull: the room produces markdown you can read before it goes anywhere. The philosopher’s room emits the claims it found as nodes joined by typed relations — implies and grounds, which are converses of each other — and you paste that into your graph the same way you would paste anything else. The paste format page covers the mechanics.

Claims, not terms. The rooms emit whole quoted claims rather than the words inside them. A node for justice would need a definition, and the room does not know what your text means by justice — but it can quote a sentence exactly, and a quotation is something you can check. Every claim that reaches the canvas is a verbatim slice of what you pasted.

What the rooms will not do

They do not summarise. Nothing produced here is a sentence you did not write.
They do not score your writing. A drift readout says a term moved, or held steadier than chance; it does not say whether that was good.
They do not improve. Running the same text twice gives the same answer, which is the point — an instrument you cannot predict is one you cannot check.