Attribution under uncertainty.

Naming an actor is the cheapest part of an investigation and the most expensive thing to get wrong. A confident, premature attribution closes off lines of inquiry that a careful one keeps open. It also leaks into reports, into briefings, into press, and into policy, where its uncertainty is quietly stripped away by people who did not see how thin the evidence was.

“A model that hedges well is more useful than one that names confidently.”

Mara is trained, deliberately, to use the same calibrated language an experienced analyst uses: assess with high confidence, assess with moderate confidence, evidence is consistent with, cannot rule out. The point is not to be timid. The point is to be exactly as certain as the evidence allows, and no more.

What we ask Mara to do.

• —Surface the specific overlaps that support a name, and the ones that don't.
• —Distinguish observed behaviour from inferred intent.
• —Treat shared infrastructure as a hypothesis, never a conclusion.
• —Name an actor only when the evidence would survive a sceptical second reader.

The result is reports that are sometimes a little duller and almost always more useful. Decisions made from them have not yet been embarrassed.