Most "source-backed" AI systems do the same thing: they split documents into chunks, retrieve the ones most similar to the question, and the model writes an answer supported by them. It looks grounded. It feels verified. It is not.
A chunk is a piece of text retrieved by semantic similarity to the question. That a passage is relevant does not mean it supports the specific statement the model ended up writing. The model can retrieve the right fragment and still assert something the fragment does not say — or paraphrase it until the meaning changes. The citation exists; the support does not.
The right unit is not the passage
Truth is not verified at the document level or the chunk level. It is verified at the claim level: an atomic, checkable statement — "the product's margin is 34%" — bound to its literal citation, the exact span of the source it came from.
In Centro de Verdad the flow is: sources → signals → claims, each with a binding to its source. Each claim keeps its raw_signal, the raw evidence it was born from. Verifying is not "finding a similar passage"; it is checking that the cited span says, word for word, what the claim asserts.
A relevant passage is not proof. Proof is the exact span that says, verbatim, what you are asserting.
RAG gives you apparent grounding; claim-binding, real provenance
The difference is not cosmetic. A RAG's grounding reassures: there are citations, it looks serious. Claim-level provenance can be audited: anyone can take the statement, go to the cited span and check whether it supports it or not. One makes you feel safe; the other lets you verify.
And there is a second benefit: if the unit is the claim, you can say things that are impossible at the chunk level — how much independent support that fact has, whether another source contradicts it, or whether simply nobody confirms it yet.
Verified · principle P1 — Provenance-First
Separating the claim from the chunk is the first step toward an AI answer being auditable, not just plausible. Everything else — the graded opinion, the traffic light, contradiction detection — is built on top of that unit.