Dishaya / Verification Methodology

What "Verified" Means, Exactly

A trust product should tell you precisely what its badge does and does not promise. Here is ours, in plain language, including the limits.

Updated 8 July 2026 · methodology published before the numbers, on purpose

The check, step by step

When Dishaya researches a question, it does not write a fluent report and cite later. It works in the opposite order:

• It reads real web sources and pulls out specific factual claims, each with the exact sentence from the source it came from.
• It requires that quoted sentence to appear verbatim in the fetched source. If the quote is not actually in the source, the claim is dropped before it is ever written. This step is mechanical, not a matter of opinion.
• For each surviving claim, a verification step judges whether the quote supports the claim, and assigns one of four labels.
• Only then is the report written, and every report closes with a Confidence Ledger: how many claims were checked, how many held, and what it cost.

The four labels

• Verified. The quoted source passage directly supports the claim.
• Partial. The quote supports part of the claim but not all of it, for example the claim adds a number, a scope, or a certainty the quote does not contain.
• Unverified. The quote does not actually support the claim.
• Contradicted. The quote says the opposite of the claim.

Claims that fail the check are disclosed in the report, not quietly removed.

What this does not promise

Verification is a rigor tool, not an oracle. It tells you whether a claim is supported by the source it cites, not whether that source is itself correct. A confidently wrong source can still produce a "Verified" label, which is why Dishaya also tracks how reliable different sources tend to be over time. Treat any output as a well-checked draft to build on, not as professional advice. For legal, medical, or financial decisions, have a qualified human review it.

How we hold ourselves to it

We measure the verifier against a set of human-labelled examples, including deliberately hard cases designed to trip it (a quote taken out of context, a number that drifted, a claim attributed to the wrong party). As the private beta produces enough real runs to report honestly, we will publish a recurring Verification Accuracy Report with the numbers, whatever they are. We are publishing this methodology first, before the numbers, so the standard cannot be moved later to fit the result.