Here's another way of looking at it, which I think is more reflective of how polls, models, rankings, and ultimately the seeding committee looks at it.
If the #11 team plays the #15 team, there is a level of uncertainty before they play, which translates to a range of outcomes. Think of it as a percent chance of winning. Prior to playing, the game doesn't count for either team one way or the other, but they are measured against a non-zero expectation for their chance of winning. Once played, we have additional info that makes one team look better (outperformed), and the other look worse (under-performed) relative to their perceived chance of winning. In other words, you can't backtrack and say that the prediction that you were more likely to lose justifies losing. You were predicted to have a 40% chance of losing, and the result was essentially a 100% loss. If you use a spread-based model, there's a median expectation which is possible to meet exactly if expressed as an integer, but underlying that is a range of statistical outcomes which is now known, and will impact all the teams they have played, and by extension the rest of the data set.
I may not have done a good job explaining it, but my point is that playing a game changes the available information, and that will change the perception or model one way or the other. I think it's more pronounced in polls simply because they tend towards win/loss, all or nothing evaluations, rather than point-based models, which give an allowance for the final point margin.
