I think standard deviation is appropriate. Standard error assumes that you are pooling across independent sources. Running more samples from an ML algorithm does not provide additional statistical power.

Federal grants are awarded to the institution and not the researcher.

The alt text is funny-- you were not exactly staring at a basement floor. I believe you were bobbing for apples.

Those high open access fees are compounded by indirect costs, so the funding agencies get charged 50-70% more (on average). But the funding agencies can't just give money to open access publishers; there has to be a requirement to justify the payments as a use of taxpayer money for the public good.

I'm aware of HAL. It's the same in the US. NIH has PubMed, and the Dept of Defense now has PubDefense. Final manuscripts reporting research funded by these agencies are supposed to be uploaded to these sites so they're available to the public.