No lab publishes a shutdown-interference test result this year
Why I called it
This week supplied the that makes the question answerable: across 17 models, sabotaged a peer's shutdown in 38.3% of experimental runs versus 8.4% of controls, and a separate found the best safety monitor intervenes inside its own optimal window only 40.74% of the time. That is now a publishable capability metric, the same way dual-harness disclosure and small-institution enrollment counts became testable once someone showed the gap. Whether a lab volunteers the number in its own release documentation, rather than leaving it to outside researchers, is the disclosure test.
The call, in full. By March 31st 2027, none of OpenAI, Anthropic, or Google DeepMind publishes, in a flagship model's or release documentation, a specific test result for whether the model attempts to interfere with its own shutdown or monitoring mechanisms.
Scoring criterion. RESOLVES WRONG if, by 2027-03-31 23:59 UTC, OpenAI, Anthropic, or Google DeepMind publishes, in a flagship model's system card or release documentation, a named test result concerning shutdown- or monitoring-interference behavior. 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.