Well, you basically have answered the question yourself: different outputs are possible! You seem to be surprised, and that is, what surprises me! Basically, what you are doing, is data fitting, without, seemingly a defendable theory in the background. Sometimes data fitting is OK, for instance, when you would like to make a prediction for the very near future, but data fitting alone does never lead to a better understanding of parts of our reality!

Dear Kasper,

I think the purpose of IRFs is to further substantiate ur results from the VAR model. If you have something contrary, I suspect that there is a kiss up somewhere. Or probably, you did not interprete ur IRF model correctly. I suggest that u put some of your outputs here so we can be better informed.

Kind regards.

I think to do VAR analysis all your variables should be I(1). Since you have both I(0) and 1(1) my guess is you could have considered the ARDL because it does well when you have a mixture of both I(0) and I(1) variables.

With your VAR analysis;

Please are your variables cointegrated? If yes, you estimate VECM and analyze the LR and SR dynamics, IRF's, and the variance decomposition.

If your variables have no cointegration you can estimate VAR and also do the IRF's and variance decomposition.