The corpus is yours.
It is never blended with anyone else's.

Ownership here is not a clause in a licence. It is where the bytes are. The originals stay where they were. The converted copies are Markdown on your disk. The index is a SQLite file on your disk. The model that reads your text runs on your machine. There is no second copy anywhere, because there is nowhere else.

Across the four proving corpora on the founder's own machine, those stores total 21.7 GB and hold 92,114 documents in 8,363,533 searchable chunks (counted 2026-09-12). Every one of those documents came from the founder's own companies' archives. None came from anybody else, because nobody else's records have ever been in a store.

There is no export feature, because none is needed. The readable copies are plain text files; copy the folder and you have exported the corpus. Open them in any editor. Nothing about them depends on this product continuing to exist.

Deletion is the same. The converted copies are files and the index is a file. Delete them and the corpus is gone from the machine. The product keeps no copy on a server, no cache in a cloud account and no fragment inside a model, so there is nothing further to request from anyone. One other thing does persist on the disk: an append-only audit log of requests made to the engine. It is on the same disk, under the same ownership, and it goes when you delete it.

The licence says the same thing in fewer words: it runs on your machine, your corpus stays yours, export and delete on request. Terms beyond that are not published yet, and this site will not pretend otherwise.

A corpus belongs to a tenant, and the engine cannot touch a store without naming the tenant. Every read and every write requires a tenant id; a call that arrives without one is refused with an error before it reaches the data. This is enforced in the storage code itself, not in a menu or a setting that could be left off.

The proving machine holds four corpora side by side — JFI 54,484 documents, HTC 23,864, SIG 13,143, AII 623 (counted 2026-09-12). A question put to one of them cannot return a line from another. They share a machine and a body of code, and nothing else.

The wall is also the reason the product can promise it never learns from your records. Retrieval is per-tenant, and there are no weights to update, so there is nothing that could carry a fact from one corpus into another. If a model were being updated from the records, the wall would be a fiction. None is, so the wall is a property of the code.

The same refusal applies outward. The model that turns your text into vectors for the meaning leg of the search runs on the device. A provider that would send corpus text off the machine to do that job is refused at startup — not warned about, refused. So the store cannot be read by another corpus, and its contents cannot be posted to anyone to be read.