The reader's condition sheet

What Habitat says about two public bodies of documents as of 14 September 2026, and what it does not say. Every number here was measured on the stored files and is the same number the live reader shows. Nothing here is a claim about what anything means.

What Habitat is reading

An act is one clause. Somebody does something to something, in a particular way. Habitat breaks each document into acts and measures each act on seventeen features of language: who is acting, whether the doing sits in one person or is spread across several, whether the writing hedges or commands, how it marks time, where it draws a boundary.

Two bodies of documents. The letters between John Adams and Thomas Jefferson, and a set of writing about climate and migration.

	the letters	the climate writing
acts	1,103	356
acts where Habitat knows who is speaking	411	13

Habitat knows who is speaking only when somebody tells it. Each letter carries its own header, so the correspondent is on the page. In the climate writing a person declared the speaker for three documents and an institution for two; the rest were left undeclared and their pronouns stay unresolved.

What it can say

Habitat now reads the person, not just the verb. Features like whether a subject is animate were being read off the clause's verb. Where the speaker is known, those features are now read off the person doing the acting, and two features that had sat frozen on every act came alive.

The two bodies of documents no longer sort apart. Every part of the climate writing used to score above every part of the letters on one summary number, which looked like a real difference. The letters have since moved into the same range, so that difference is gone. Recorded as a result, not as an improvement.

Almost any two clauses have something in common. Two clauses tend to share the same kind of happening, the same shape of who does what, and the same value on whatever this writing never moves. That common part is a large floor under every pair. Habitat sorts by what is left once the floor is removed, and it shows a reader the pairs where the leftover is largest. The floor is the same size inside each voice, across voices, and in both bodies of documents, which says it belongs to how Habitat reads rather than to the writing.

Two readers, one set of letters. Read through Adams's own habits, Adams's passages rank at or above Jefferson's; read through Jefferson's, the reverse. Yet the two orderings of the whole correspondence agree almost entirely. Habitat does not sort the letters by who wrote them.

What it cannot say yet

Whether any of this is meaning. The seventeen features are unrelated to surface style and to the text embedding, and being unrelated to those is not the same as measuring meaning. Settling the question needs passages judged by people, frozen in advance, used only to try to prove Habitat wrong.

Three of the features are currently one feature. How many participants a clause has, how varied their roles are, and how spread the agency is all move together almost perfectly, because each act enters carrying one participant. Whether an act should carry the verb's other participants is an open question.

Two of the features read nothing in these documents. One sits at its default on nearly every act; the other moves only where the speaker is known. This writing does not contain what those two look for. They are unobserved, not broken, and Habitat holds them ready.

One pair moves together because of how it is built. Where the speaker is known, two of the features read the same fact about that speaker, so they agree by construction. Separating them means finding a second source in the document for one of them.

A prediction that failed, kept on the page. Before running, the author predicted that 55 to 75 per cent of top pairs would show a negative leftover. The real figure was 86 per cent, against 96 per cent for randomly chosen pairs, which refutes the prediction and does so in the opposite direction from the one expected. Prediction and refutation both stay here.

How to check any of it

The files are stored alongside the code that made them. The test suite passes, 711 tests on the shipping version. Every intentional change to the inputs is logged with its date and its reason. The numbers reproduce, we publish what works and what does not, and the logs are available on request.