Even the unaligned models, I’d imagine, will provide at most a picture of what we believe or expect to see in the past. (Actual 19c / early 20c data is a tiny fraction of the training corpus.)

Yeah, that's my point (I think). What we see is the past distorted by expectations learned from its uneven collapse into the present (the training data). The aligned models then try to "correct" that past. Both encode history and its biases somehow, but bizarrely, ambiguously, nearly untraceably so.