I see, so the hypothesis is there's some way of determining "how much choice" it has for each output token. The less choice it has, the more we will call what i

I see, so the hypothesis is there's some way of determining "how much choice" it has for each output token. The less choice it has, the more we will call what it's doing "verbatim copying". Once the threshold is crossed and you want to "trace" it, though, I'm still not clear on how you can map back through the latent space and figure out where in the corpus it's copying from. Other than the technique suggested upthread, where you just literally search the corpus as a separate operation.