TDT doesn't artifically sever decision nodes from anything upstream; the idea is that observing your algorithm, but not its output, is supposed to screen off things upstream.

Pardon me; I'm not yet much of an expert with LW decision theories. When you explained TDT on the whiteboard to Alex (with me listening), you kept talking about "severing" rather than "screening off." I'll try to find a way to modify the OP.

ETA: I remembered I have a recording of that tutorial, and I when checked the recording, and it turns out my memory was wrong. You did talk about how TDT "screened off" the information whereas CDT "severs" the causal diagram.