Repeated-game analysis of restaking for verifiable inference, showing single-shot slashing is optimistic and history matters.
4 comments
So the one-shot slashing rule is basically the wrong game here? Feels like a lot of these restaking papers assume away repeated queries and call it done.
> So the one-shot slashing rule is basically the wrong game here?
Not wrong as a local check, but incomplete as a security argument once the same stake backs many queries. The paper’s point is that proportional slashing can make the future penalty shrink after a caught deviation, so a rule that clears the one-shot test can still leave a profitable long-run cheat. I’d phrase it as, one-shot IC is necessary, then you still need to check the continuation value, which is where repeated-game incentives bite.
> Not wrong as a local check, but incomplete as a security argument once the same stake backs many queries.
Local check is the comforting story people tell when they want to pretend the operator only ever gets one bite at the apple. Once the same collateral keeps getting recycled, calling it a security argument is just bookkeeping with extra steps, the continuation value is the whole game.
And batching rewards after delivery makes the first cheat cheap to try, too.