Why every Brain answer ships with a citation.

When the Brain answers a question about internal policy, it returns the answer with a link to the document, the revision, and the paragraph the answer came from. The user sees it inline. The reviewer sees it in the audit log. A regulator who asks sees the same chain.

This post explains how attribution works in Veetso Brain, why we made it a default rather than a feature, and what it does to drafting.

Why we made attribution a default

Answers without citations sit in a system structurally indistinguishable from a confident hallucination. We do not want our staff to be in the position of having to take an AI's word for it. We want them to be able to point a finger at the source, every time, without leaving the answer.

The default matters. When attribution is opt-in, attribution disappears. People skip the extra step under deadline pressure, and the audit log fills with answers that nobody can verify. When attribution is opt-out, attribution survives, and the friction lands on the few cases where someone has to argue against it.

We chose default-on, no opt-out for the internal mode.

The four layers that have to agree

Attribution is harder than it sounds. To get an unbroken chain from question to source, four layers of the system have to cooperate.

01. Retrieval

The retrieval system pulls candidate sources for the model to read. Each candidate is tagged with a stable identifier: document, revision, paragraph. We do not let the model see content that has not been tagged. This is the input contract.

02. Generation

The generation step has to keep the tag attached to each sentence it produces, even as the wording is rewritten. This is the hard part. Most attribution systems lose the tag during synthesis because the model paraphrases freely. We constrain the synthesis to per-claim attribution: each claim either has a source or is omitted.

03. UI

The interface has to surface the citation without burying the answer. We render the source inline as a small clickable chip next to each claim. Clicking it opens the source document at the relevant paragraph in a side panel. The friction to verify is one click.

04. Audit log

Every answer, every source it cited, every user who saw it, and every approval it received is recorded with a verifiable hash. The audit log is the same chain we would hand to a regulator without any post-processing.

How drafting mode is treated differently

When a user asks the Brain to draft a policy memo from broader model knowledge, the output is marked draft top to bottom. It does not get a citation chip because there is no source to point at. It is not citable as fact downstream: the audit log records it as a draft and refuses to let it be referenced as an answer.

This separation is essential. A regulator reading our audit log can tell at a glance which claims are sourced and which are drafted. Pretending the two are equivalent would undo the whole point of attribution.

What the regulator sees

The audit log entry for any answer is a single record with five fields:

The original question.
The set of source documents the retrieval system surfaced.
The subset of sources the model actually cited.
The answer the user received, with per-claim citations.
The reviewer (if any) who approved a downstream action based on it.

A regulator can sample any of these and follow the chain end to end. No part of it is reconstructed retrospectively. The chain exists at the moment the answer is generated.

What attribution does not solve

It does not solve hallucination on its own. A model can still cite a real source while drawing the wrong conclusion from it. What attribution does is make that mistake visible. The reviewer can read the citation, compare it to the claim, and decide whether the citation actually supports what the model said. We separate the question "does the model have a source" from the question "does the source say what the model thinks it says." Attribution answers the first, human review answers the second.

This is why attribution is necessary but not sufficient, and why we pair it with human oversight on every regulated decision. See the gates post for how the two fit together.

FAQ

Questions readers ask

What is AI source attribution?

AI source attribution is the practice of attaching, to every AI-generated claim, a verifiable link to the source the claim came from: the document, the revision, and the paragraph. The link travels with the answer through generation, into the UI, and into the audit log so a reviewer or regulator can verify any claim end-to-end.

How does Veetso Brain prevent hallucination?

We constrain synthesis to per-claim attribution: each claim either has a source it can point to or it is omitted. This does not stop the model from drawing the wrong conclusion from a real source, but it makes the mistake visible to the reviewer and to the audit log. We pair attribution with human oversight on every regulated decision.

Does drafting mode include citations?

No. Drafting mode is for output that does not have an internal source, for example, a policy memo generated from broader model knowledge. The output is marked draft top to bottom, has no citation chips, and is refused downstream if anyone tries to treat it as a citable fact. The separation is deliberate.

What does the regulator see in the audit log?

Each Brain answer is logged with five fields: the original question, the documents the retrieval system surfaced, the subset the model actually cited, the answer with per-claim citations, and any reviewer who acted on it. A regulator can sample any entry and reconstruct the full chain from question to source.

What happens if the source document is updated after the answer?

The citation points to a specific revision of the document, not the current version. If the document is updated, the original answer's citation still resolves to the revision it was based on. The audit log records when the answer was produced, so the regulator can see exactly which version of the source was in play at the time.