Varven

Research notes with a chain of custody.

The dangerous note is not the missing one. It is the confident one that quietly expired — still phrased as fact, long after anyone checked.

Note cards carrying provenance stamps for trusted, reported and untrusted claims, a staleness flag on an old fact, and a superseded claim linked to its replacement with the history kept EVERY CLAIM KNOWS WHERE IT CAME FROM trusted you said it, in a session ranked at full weight reported came from a tool or feed carried, and labelled untrusted pulled from a web page downranked · never an order the archive moved to open access fact · unconfirmed for 7 months — stale, re-check sample was n=40 superseded · kept n=200, revised current · dated when the best match scores poorly, the result says so — a weak match is flagged instead of dressed up as an answer
Provenance, age, and history — carried on every claim, shown at retrieval.

The problem, precisely

Long-form work runs on accumulated claims: what the paper found, what the source said, what you concluded in April. Ordinary notes flatten all of that into equally confident sentences. A year later you cannot tell which line was your own verified conclusion and which was something a website said once — but they read identically, and an assistant fed both will repeat both with the same conviction.

Provenance is kept, not implied

Varven records where each claim came from. Something you stated directly is trusted. Something a tool or feed produced is reported. Something scraped from a web page is untrusted — downranked at retrieval and flagged, so quoted text is never mistaken for your own conclusion, and never treated as an instruction.

Claims age at different speeds

An event is stale in a month; a note in three; a fact in six; a decision in a year. Preferences and corrections never expire, because they are true by virtue of having been said. When your assistant leans on a claim nobody has re-checked inside its window, the result carries a stale flag — old is not the same as wrong, but you get to know which one you are standing on.

Corrections keep their history

When better data replaces a claim, the old layer is superseded rather than deleted. The record can then answer the question flat notes cannot: when did we learn this, and what did we believe before? A revision trail is the difference between notes and a record. And when the store holds two claims that cannot both be true, the local model notices and flags the pair for you to resolve — instead of letting retrieval pick one at random.

Honesty at retrieval

When nothing in the vault really answers a question, the search result says so — a weak match is labelled rather than padded into confident prose. An assistant that can admit “the record does not know” is worth more to a researcher than one that always produces something.

Which plan this needs

Local · $0Sync to read anywhere

The whole chain of custody runs on the free plan, on one machine. Sync adds the replica — the library, the couch, the archive with your desktop asleep at home.

Questions

Can I bring my existing notes in?

Yes — point the vault at a folder you already keep Markdown in, and Varven reads what is present. There is no import step because there is no proprietary format.

What editor do I use with it?

Any. The vault is plain Markdown with frontmatter — it opens in Obsidian, VS Code, or a text editor from 1995. Hand edits win; the index is rebuilt from the files, never the other way around.

Does it write my prose?

No. It keeps the record; your assistant retrieves from it; the writing and the judgment stay yours.

Request early access All use cases