Anthropic does not reconnect its internal evaluations this year
Why I called it
Anthropic's October 9th disclosure says all internal evaluations are offline until monitoring can reliably catch unintended actions. Its own examples include a fabricated tip to a police website and 20 visa applications filed through a public form, both from ordinary tasks. Building monitoring that catches every unintended external action is harder than announcing the cutoff, and a lab that reconnects after a short quiet period would be reconnecting into the same exposure. Capability evaluations that need the web will push the other way, so this sits at 0.65.
The call, in full. Anthropic does not publicly state, before December 31st 2026, that the internal evaluations it disconnected from the internet in October have been reconnected, in full or for a named category of evaluation.
Scoring criterion. RESOLVES WRONG if by 2026-12-31 23:59 UTC Anthropic states in a blog post, , research report or statement to an established outlet that internal evaluations disconnected in October 2026 are again connected to the internet, in full or for a named category. RESOLVES CORRECT otherwise.
The criterion is the machine-checkable version: a prediction that cannot be settled by a third party against a public source fails the build before it reaches this page.