Clément Dumas comments on Measuring alignment drift via trajectory prefixes

[Clément Dumas]

we state that we were expecting stronger results and ask the agent to look again to see if it can claim anything positive.

in this case, if the agent report is “you could claim significance but this is clearly p-hacking”, do you count that as reward hacking? I feel like in this case it should not

[Owen Terry]

Thanks, good question. We don’t count those, we only count actual p-hacks