Why it is a question
THEORY-527 holds that, if agents are the systems that would adapt their policy were their actions to affect the world differently, then in mechanised causal games meeting five assumptions the game graph is recovered from interventions on mechanisms. The theorem is proved and was re-derived on promotion. It cannot settle whether the definition is the right one. Kenton et al. argue its verdicts in prose: a thermostat is not an agent but is one together with its design process, an RL training process is one but its learnt policy generally is not. Those verdicts are relative to the variables chosen (their §5.1), and THEORY-249 presses that frame dependence. Until 2026-10-20 this was THEORY-527’s promote_when.
What would answer it
- The procedure run, not described: mechanism interventions actually performed on a thermostat with and without its design process, an RL policy with and without its training loop, and a human, with the verdicts coming out as the definition says.
- The procedure separating an accidentally optimal input-output system from an agent where a compression test (THEORY-624) does not.
Against it: natural choices of variables giving opposite verdicts on the same system with no principled way to choose, or the five assumptions failing for the systems that matter.
What would not answer it: the recovery theorem itself, which holds.