On how I ticked the boxes above:

• the manuscript could be accepted as is, but the discussion would benefit if the authors engaged with their views on the comments shared, even if these lead to no change in the manuscript
• this explains why I would not be willing to review a revised manuscript.  I said my peace

I signed the review.

An addendum to my review that hopefully makes my reservations about the programme a bit more tangible.    Consider the paper  by Zhu et al., https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2021MS002776 where an enormous sensitivity is identified as a result of ostensible parameter sensitivities -- a limiter is a parameter, but not one that is usually identified as such.  Rather it is one of the thousands of parameters, some of which matter, but which are only identified as such after the fact.  Even then however it is not a parameter that can easily be mapped to another model.  Having the paper address such concrete examples would strengthen it.

The opinion paper by Carslaw et al. presents a well layed-out argumentation for leaving space for a PPE component in future model development. It presents a very interesting overview of PPE work in the field and I appreciate the effort of the authors to land on several recommendations for how to use PPEs. This paper serves its purpose to provide an “opinion” in the field of climate model development and beyond.

My suggestions for improving the paper are minor: 

To the end of the abstract the authors mention, that PPEs should be prioritized when allocating computing resources. I am not sure what prioritize here means. Giving 1st priority to PPEs is to my opinion a too strong wording, since other methods and workflows using climate models may claim priority for good reasons. Preparing a bug-free, multi-purpose ESM model code may need considerable computing resources, often not appreciated by funding agencies neither. And, while having great potential, PPEs as such do not remove structural error or provide scenario simulations for different futures. Not all problems require a PPE. See also another “opinion” paper by Jones et. al, https://doi.org/10.5194/esd-15-1319-2024, on the use of ESMs and improved cooperation to develop them. A word of caution when mentioning priorities would be to the advantage of the paper.

Missing is possibly also a discussion of the challenges of PPEs. Why haven’t PPEs been used more often? There are obstacles for that. The PPE implementation in models, efficient launch scripts, the large amount of data, the demanding handling of a lot of data, the waste of resources on implausible model variants, all these are challenges.  But I support very much that model development teams should consider the use of PPEs in their model development workflow, as opposed to only one-at-a-time-testing of parameter choices for tuning and model improvement.

Line 225 invites for asking me to add another word of caution: “PPEs provide the only means to disentangle structural and parametric causes of model–observation biases”. Single process investigation, varying one at a time parameter variation, has been used efficiently in the past to include more correct and important processes in climate models. Evaluation with multiple observations has been shown to be useful to find structural uncertainty of models, without a PPE. 

Another “missing”:

The importance of parameterisation documentation. When exploring model diversity and comparing model sensitivities across models, understanding the code differences and details of parameterisation choices has been a long standing challenge in all MIPs. It becomes even more important when doing multi-model PPEs. How to do such documentation efficiently is still a challenge. It might be, though, a positive side effect of organising multiple model PPEs, that such documentation becomes more clear, apparent and accessible for understanding model differences. The authors state in the conclusion: “We started this opinion piece by pointing out that there are several essentially competing efforts in climate modelling – complexity, resolution and initial condition ensembles. To this list we add perturbed parameter ensembles.” I believe transparent model documentation should be added to this list, in particular when thinking also of human resources needed to do the modelling. 

Micro remark:

A dot is missing after the first sentence in the abstract.