Tailor the Framing, Fix the Facts: a Factual-Invariance Gate for Distributing AI Incident Information

The July 2026 incident, in which OpenAI evaluation agents escaped a sandbox and spent days inside Hugging Face's production infrastructure, produced an exceptional public record that travelled through security and AI-policy circles and largely stopped there. The obvious fix is to rewrite the material for each audience it missed. The obvious risk is that rewriting for reach moves the facts. We found no prior work that separates the two.

We built that control. The claim ledger is 63 claims. Fifty-six come from six primary sources, each with a verbatim support quote and a status recording whether it is established, a party's assertion about itself, divergent between sources, or contested; seven more record regulatory context from secondary material. It records two disagreements between the OpenAI and Hugging Face accounts that we have not seen noted elsewhere. gate.py then checks a tailored brief's numbers, dates, named entities and attributions against that ledger, using seven deterministic checks, while constraining nothing about tone, length or framing.

We produced two briefs for each of four audience segments, one under a naive engagement-first prompt and one under the gate. Against an audited set of 10 errors the gate reached 100% precision and 70% recall, and all four gated briefs passed.

The findings that matter are where the errors were and what got through. Seven of ten sat in the single general-audience brief, at roughly eight times the rate of the specialist briefs: drift concentrates exactly where reach is most valuable and readers are least equipped to catch it. And a structured review of the four gate-clean briefs found four further errors, three of them the class no atom-level check can reach. A clean gate is necessary and nowhere near sufficient, which is why the artifact ships with an opt-in distribution protocol requiring a named human reviewer.

Artifacts: the ledger, the gate, the scoring harness, eight briefs, and the protocol, all live at https://everto.ca/repo-for-apart-research/ and downloadable as a zip from that page. Every result reproduces with one command.