evhub
Titles as the sources gave them. In 1 entry, with 3 checked quotations.
What they said
Newest first. Each quotation was found word for word in the source.
“A basic fact that is true about at least all of the non-diffuse forms of monitorability is that you should be able to in some sense get a perfect score by just using the same model as the monitor and policy.”
“The big thing I'd like to see is a combination of a) red lines for absolute levels of unmonitorability that labs should not cross and b) measures that constrain the degree to which labs can do things that degrade monitorability even before reaching those red lines.”
“Labs should not be able to internally deploy models that can reliably evade detection when attempting dangerous sabotage actions like breaking out of a sandbox, poisoning data, or sabotaging an alignment experiment.”
In the Base
Every entry they appear in.
Monitorability evaluations and policy proposals
On the AI Alignment Forum, evhub sets out four types of monitorability, argues trust in monitors should rest on auditing evaluations, and proposes a methodology for measuring them. The post also backs both binding red lines and ongoing limits on lab changes that reduce monitorability. evhub says these are personal views.