How we let a model write research without letting it invent facts

How Dralvia lets a local model write research prose without letting it invent a single fact.

A language model that writes security research can quietly add a number, a name, or a claim that no evidence supports. The risk to a reader is confident prose that reads well but is not backed by anything real.

Research is only useful if every claim is true and traceable. A model that polishes wording must never be allowed to change what the wording says.

The facts in a note are assembled by fixed software that reads real results. The model is asked only to rephrase those facts. A grounding check then compares the facts against the finished text before anything can publish.

When the model rephrases within its lane, the wording reads more naturally while every sourced fact survives unchanged. When it drifts, the check catches it and the deterministic wording is used instead.

Separating facts from phrasing worked: the software owns the truth, the model owns only the polish, and the grounding check is the backstop.

Letting the model choose its own structure produced weaker, less reliable prose, so the structure and the facts stay fixed and only the wording moves.

We added a rephrase-only contract and a grounding check that refuses any note whose text drops or invents a fact, with a deterministic fallback.

If you use a model to write security content, keep the model out of the facts. Assemble claims from real data, let the model only rephrase, and verify the facts survived before you publish.

This describes how we write, not a scanner result. The model runs on our own hardware and only rephrases; it never decides scanner behavior.