How to trust the data your AI agent uses

An AI agent inherits the quality of the data under it. A model can reason well and still be wrong because the fact it stood on was stale, misattributed, or quietly edited after it was read. Most data arrives with none of that history attached. You get a value and a promise.

Four things make a piece of data trustworthy enough to act on.

The source

Not "the web" or "a database," but the specific origin: which filing, which feed, which sensor, at which address. If you cannot name the source, you cannot check it.

The time it was read

A price, a ruling, a score: each is true only as of a moment. A fact without a read time is a fact without a shelf life.

Every step since

Raw data is rarely what reaches the model. It gets parsed, joined, scored, summarized. Each step can introduce an error that the final value hides. A trustworthy record carries the chain, not only the endpoint.

A receipt you can verify

The first three are claims. A receipt is what lets someone else confirm them without taking your word for it: a signature over the exact bytes, checkable against a published key. Put together, these are an ingredient label for data. You would not feed a person food with no label; an agent deserves the same.

A signature is worth being precise about. It proves who recorded the data and that the bytes have not changed since. It does not prove the claim is correct. You can sign a mistake. What signing buys is accountability: a name on every observation, and a value you can re-check instead of re-trust. That is added trust, laid on top of the source, never a replacement for reading the source.

If you are building an agent and want the data under it to carry that label, DRM3 is trusted data for AI: a signed record of facts and datasets, each with its source cited and a receipt you can check in the browser.