This doesn’t help much as the stddev is almost same for all of them all the time

@rjnw It helps a lot! I now see that I don’t know whether rkt is better than jags, but rkt and jags are definitely better than hk!

(I assume hk means hs)

Improvements to the plot: 1. remove the darker curves for mean+std and mean-std; leave only the shading and the thicker and darker curve for mean in the middle. 2. same number of digits after decimal point for all labels on each axis 3. sort legend to match rough best-to-worst order of the 3 curves

@carette The two things that @rjnw is writing are (1) describing the compilation pipeline from simplified hk to LLVM; (2) describing the experiment and results. So, we agree that an experimental setup section is needed. And I hope the way that section can be written is you keep asking questions and Rajan keeps revising the text to answer the questions while adhering to my two favorite writing principles: punch line first and read out loud.

@ccshan I would be most happy to continue doing that. For example, I think I understand what these latest plots are saying [good]. However, I still am not fully confident of the experimental setup. @rjnw it’s not you, it is more that Hakaru has seen way more than its fair share of comparing apples with motorcycles, so my base dubiousness level is quite high.