Cite a figure
Every price on every board carries a permanent, dated citation. This page is what that citation is, how to get one, and how long it will keep resolving, which is a promise rather than an intention and is written here because it has to exist before the first citation is minted rather than after.
Getting one
Every figure the record holds carries a cite mark beside it, on the boards, on a sku's own page and on every period in its ledger. The mark leads to that figure's citation:
https://vitna.ai/cite/{id}{id} is the observation the figure is, which is also what /api/v1/observations/{id}/proof takes and what every tape row carries as observationId. That page holds the figure, the document it was read from, the moment it was read, the signed day it sits in, and the citation itself in three shapes.
A figure with no mark is one this copy of the site cannot cite. Citations resolve out of the record, and an instance serving the committed snapshot has no leaves to name; vitna.ai serves the record, so this affects local checkouts and mirrors rather than the site.
The three shapes
They differ by how much identifier each can afford, and the difference is reasoned rather than decorative.
| Shape | For | Carries |
|---|---|---|
| Short reference | A slide, a chart footer | The figure, the date, the url |
| Footnote | A vendor review, a memo, a report | The document's short hash and the signed day, and what the signature does not cover |
| Contract schedule | A schedule or an annex | The document's full sha256 and the source url |
A footnote takes this shape. It is an illustration of the form rather than a live citation: the figure, the hash and the id below are stand-ins, and the real ones come off a board.
> Anthropic Claude Sonnet 5, input price: $3.00 per 1M tokens. Read by Vitna > from www.anthropic.com on 23 Aug 2026 at 09:00 UTC, document sha256 4f2a..., > in the signed day 2026-08-23. https://vitna.ai/cite/{id}. No provider signs > its own pricing page: the signature covers Vitna's record of the fetch and > not the provider's page.
A derived figure carries its arithmetic
Some figures are not printed on any page. A provider that sells a sandbox by the 20-minute session has an hourly rate implied and unpublished, and Vitna computes it. The record refuses such a figure without the sum that produced it, and the citation carries that sum in every shape including the short one. One the record holds today, in the clause a citation prints it in:
> Derived rather than read: $1.92 per 20-minute session for a 64 GB container, > times 3 sessions per hour.
A computed rate quoted on a slide with no arithmetic beside it is the one way this record could put a number into somebody's argument that no document carries. That is why it is in the short form too, where the line is otherwise kept as brief as it can be.
The promise
The url is the citation. Everything else in a pasted line is there so it reads on its own; what has to survive is that vitna.ai/cite/{id} keeps answering.
- A citation is minted only for a reading the record holds. No url is handed out that does not resolve on the day it is copied.
- It keeps resolving for as long as vitna.ai serves anything. That is the whole of the commitment on availability, and it is deliberately not dressed up as more.
- An id is never reused and never repointed.
witness.observationis append-only: the application roles hold noUPDATEorDELETEon it, ids come from a sequence that is never rewound, and nothing is deleted. If a citation ever stops answering, it will not have come to answer about a different figure, which is the failure that would actually damage somebody quoting this record. - Four things about the answer cannot change: what was priced, on which meter, at what figure, and the document and moment it was read from. The page around them can be redrawn, the wording can improve, the forms can gain a fourth. Those four are the citation.
- A misreading is corrected by a later reading, never by editing this one. If Vitna read a page wrong, the citation still resolves to what was read, because that is what it says it is; the correction is a later observation, and the sku's own page carries both with the documents behind each.
- If the path shape ever changes, `/cite/{id}` keeps answering. A url in somebody's contract is not a thing to tidy up.
What a citation does not claim
That the provider published the price. Nothing here can. No provider signs its own pricing page, and the signature covers Vitna's record of the fetch rather than the page it fetched. What ties a figure to a provider is the archived document named by its sha256, and verifying a figure is fetching that url yourself and comparing hashes.
That the price is still that. A citation is of one reading on one date. The sku's own page lists every period it has held since, and draws each price that has moved on a true-time axis.
That the provider's page still says it. A page that has changed since is the ordinary case, and the archive is what makes the disagreement arguable rather than a matter of opinion.
That the day was signed when you copied it. Roots are cut once a day, after the day closes, so a figure read this morning is provable tomorrow. The citation says which of the two it is rather than implying a root that does not exist yet.
Citing something that has to outlive us
If a figure is going into a document that has to stand on its own, take the proof with it:
curl "https://vitna.ai/api/v1/observations/{id}/proof"That answer carries the leaf's own fields, its hash, the path of siblings up the tree, the root the walk should reach, and the signed statement over it. It is self-contained: a reader holding that file and the published key can check the arithmetic with nothing from this site.
The verifier is given away. platform/witness/attest.mjs ships inside the Apache-2.0 vitna-runtime package, so an independent implementation of the check is a dependency rather than a rewrite. The site that serves these urls is not open source, so what survives Vitna is the checkable artifact and not a mirror somebody else can stand up.
Two things are worth archiving beside a citation you intend to rely on: that proof response, and the day's row from the attestations, ideally fetched on a different day from a different network. The mechanism is that the statement was in the open at the time, so an edit made later has to contradict something already published, and a copy somebody else holds is what makes that bite.
What the record does not archive for you
The archived copy of the provider's document is behind a token, because it is a third party's page and republishing it is a question this record does not need to answer. The hash, the url and the fetch time are public, and those are what a check needs. If your argument depends on the page itself, keep your own copy of it: fetch the url the citation names and archive what comes back.
