Models now sit inside the rooms where records are kept — the documents, the datasets, the files someone has to stand behind. Until now, nothing held what a model does apart from what a person does. Every tool on offer makes the model's voice part of the record it touches, and every record becomes, a little, the model's.
Your highlights, your notes, your reading build your own record, and nothing rewrites them. When a community asked that nothing ever learn from their documents, the architecture could enforce it rather than promise it.
See how that held →A model can read beside you and speak about what it sees. It cannot write into your record. That is a wall in the architecture, not a policy.
The place where you can watch that wall tested isn't built yet. This note stays until it is.
Records here only grow; nothing is revised in place. Every number published reproduces from its own bytes, on your machine. When we have been wrong, the withdrawal is printed where the claim was.
Read the corrections, dated →Three kinds of material, one unadjusted instrument. Every number below reproduces from its own bytes.
Twenty-seven documents from one cape: the town and its agencies, the people who live here, working life, the marsh and fisheries, and the study of what may come. Habitat reads how far each voice bends from the room they share — and every one of the five is more tangled on its own than the room is.
the Cape Ann reading →On a public survey, dropping the rows that produce a finding moves the reading 32 times further than an innocent change of the same size — and that ratio holds whichever way you measure the data's own churn. It says the step moved the finding. It does not say the finding was wrong.
the worked case →Seventeen thousand public mortgage filings. Reorder every row: the file's hash flips, the reading holds byte for byte. Change a single value: the file barely moves, the reading does. A checksum reports both of those backwards.
recompute it below →Seventeen thousand mortgage filings enter. Habitat reads their shape and publishes one number. Recompute it here, from the bytes this page serves — nothing is sent.
Reorder the rows and the reading holds, byte for byte, while the raw-bytes hash moves. That difference is the whole claim — and a checksum cannot make it.
164 letters between John Adams and Thomas Jefferson, 1812–1826. Choose a passage and the field hands back the passages it couples to, each carrying the words that admitted it. Your margin notes persist. The letters never change.
A user-composition can be read, highlighted, composed into, reviewed as a living document, and commented on, wholly intact.
An agent reading the user-document can speak about what it sees, but it cannot write into the record.
The world has, at present, two so-called boundary tools. A watermark: the generator marks its own output, and the mark travels with the artifact without provenance. A classifier: trained to answer is this generated? — a question that gets harder every month. Both interrogate an artifact that can be endlessly remade.
Habitat hands you a receipt.
Apparatus