Trust engineering · ~6 min

Not every source carries the same weight

An ERP export, a Slack message, and an LLM answer do not deserve the same initial trust. How we assign priors per source — and why seeing the same data point twice does not make it more true.

By Juan Urrea2026-06~6 min
ERP, Slack, and LLM with different trust priors: weight bars of 0.90, 0.55, and 0.25

Every data point arrives with a prior: a starting confidence that depends on where it came from. An authoritative record —the ERP, a signed contract, the system of record for that domain— does not weigh the same as a Slack summary or a model's output. Treating them as equals, giving them the same vote, is a design error before it is a data error.

Institutional truth rules

Centro de Verdad assigns priors by source and authority. Every domain has a system of record: for sales, the ERP; for a commitment, the contract; for a ticket's status, the tracker. That origin weighs more — and everything else corroborates; it does not compete as an equal.

Note: a high prior raises the starting point; it does not close the case. An authoritative source that contradicts itself is still a problem. The prior is where you start, not where you end.

The dependence discount

Here is the most deceptive error: corroboration only counts if it is independent. Five dashboards reading from the same export are not five confirmations — they are one, shown five times. Seeing the data point repeated everywhere creates a sense of consensus that does not exist.

Five copies of the same origin are not five proofs. They are one proof, five times.

That is why the system discounts correlated corroboration: it traces which origin each occurrence comes from and does not add up the ones that share a root. Real consensus demands sources that do not copy each other.

The special case: AI-only

A data point backed only by another AI does not reach "verified". A model's output is not validated by citing another model — it takes non-AI corroboration or an anchor in an authoritative record. If all the evidence was born from a machine, the uncertainty stays high on purpose.

Verified claimprovenance (how reliable is the source?) and veracity (do independent sources corroborate it?) are distinct axes and must not be mixed.
Verified · ADR-0003 — authoritative sources

Weighing each source for what it really is —and discounting the ones that copy each other— is what keeps noise from disguising itself as consensus. Without that, repeating a lie enough times turns it into "data".

This is how we weigh evidence at Centro de Verdad. More at ver-4.comver-4.com
← All posts