A pre-flight checklist for feeding data to a model

Before a dataset reaches a model, run it through seven questions. Each has a yes or no answer. A no is not always a stop, but it is always something to write down.

1. Can I name the source? Not the category, the specific origin. If the best you can say is "scraped," you are carrying risk you cannot describe.

2. Do I know when it was read? Every value has a shelf life. A read time lets you decide whether it is still good.

3. Can I see the steps between source and value? Parsing, joining, scoring, summarizing: each is a place an error hides. A record that shows only the endpoint is hiding the middle.

4. Is there a signature over the exact bytes? A signature binds a name and a version to the data. Without one, "the same data" is an assertion.

5. Can I verify that signature myself? If checking requires the vendor's servers, the check is a courtesy, not a proof. You want a public key and an offline check.

6. Is the record append-only? A dataset that can be edited in place has no past. You want corrections to arrive as new versions that supersede, with the old version still visible.

7. Do I know the license and the limits? Data you cannot legally use in production is a liability no matter how clean it is.

Why all seven

Six yeses and a clear license is a dataset you can defend in a review. Fewer, and you are making a judgment call, which is fine as long as it is written down rather than assumed.

The reason to insist on all seven is downstream. When a model's output is questioned, the question travels back to the data. If the data carries its own answers, the review is short. If it does not, you are reconstructing history from memory.

DRM3 is built so a dataset answers all seven on its own: trusted data for AI agents, signed at the source, versioned, and verifiable.