Trust overview

You are going to take these numbers into a meeting. Three things make that safe, and each one has a page behind it.

What you'll learn

• The three commitments and what each rests on
• Where the formal documents live
• What we will not claim

One: the numbers are conservative

A judge model reads every answer. An independent verifier audits the ones the judge is not certain about. A disagreement means the citation is not counted, and a check decided by a single model is labelled as such rather than borrowing the confidence of an agreement that never happened.

The effect is a number that errs towards understating you. See how we verify citations.

Two: the evidence is readable

Every counted citation resolves to the answer it came from, with the sentence, the surface and the date. You can check our judgement yourself in Proof, and you should, at least once.

The same applies to absence. An unmeasured surface is reported as excluded rather than as zero, and an unscoreable change is reported as not measurable with a named reason. See when a number is missing.

Three: one client's data never reaches another

Workspace separation is enforced where data is read, not by filtering in the interface. See data isolation.

The formal documents

For a DPA, sub-processor list or security questionnaire, email support@outercite.com.

What we will not claim

• That every citation is checked by two models. The verifier audits the uncertain ones. That is what the pipeline does, and it is what we say.
• An industry-wide citation rate. Nothing in the product produces one, so we do not publish one.
• That a projected lift is a forecast. Projections carry no measured error band, which the product says on the panel itself.
• That an absence is a zero. This is the one that costs us on a chart and is worth it.

Try this in Outercite

Ask the assistant "Show me the exact sentences AI wrote about us" and check one row against the engine yourself. Trust in a metric should survive one manual check.

Related