How do you separate reliable claims from fragile early incident evidence?

The review builds on two separate preliminary incident records: the reported July 2026 Hugging Face/OpenAI evaluation breach, and Anthropic's subsequent review of three incidents. Its abstract then names the method: "A comparative evidence protocol distinguishes record-specific factual claims from the shared systems lesson."

The reasoning for the split is about how fragile early incident evidence is. A preliminary record is the weakest kind of source: its facts (what happened, in what order, with what cause) can be disputed or revised, and each record has its own. A lesson that both records support survives revision of either one's details; a fact from one record does not transfer to the other. So the protocol sorts claims into two bins. Record-specific facts stay attributed to the record that states them. Only what the records share is lifted into a lesson. In this review the lifted lesson is Is your evaluation environment actually part of the threat model?, and the discipline shows in the conclusion, which says outright what the records do not establish (What can two incident records actually teach us about AI evaluation security?).

How to apply it in this vault. When several excerpts recount one episode, attribute each fact to the excerpt that states it, and lift a claim into a note only when independent accounts agree on it. A note that restates one record's narrative as settled fact has skipped the sort.

Limit. The excerpt describes the protocol in a single sentence. It gives no steps, no criteria for what counts as shared, and no account of how disagreement between records is handled, so this note records the principle and not a procedure.