Sure ideas and facts would be spread out. The weights are cumulatively involved in generating the tokens. The error margins from them would also be accumulati

Sure ideas and facts would be spread out. The weights are cumulatively involved in generating the tokens. The error margins from them would also be accumulating for some paths while also decaying for some others. If the training and inference process could track, it should be possible to figure the error margins for the output tokens. However, such an approach would be too compute intensive to be practical.