Ah, you are using BiB. I'll need your addon export to be able to see what you are seeing.
Just looking at the weights you are using... I can tell you what the problem is.
The first thing to understand is that stat weights cannot rank special effects like trinket procs, legendary effects, set bonuses, etc. To rank those, we use actual simulation results. We do this by doing a representative set of simulations and then using some statistics to approximate a ranking. In order for this process to work, the stat weights need to be a predictor of DPS.
Once you normalize the stat weights, they are no longer a predictor of DPS and they become incompatible with the special effect rankings stored with the gearing strategy. What has happened is that the stat weights you entered actually result in negative DPS, causing lower item level trinkets to rank higher, because you have inverted the slope!
If you want to use those weights, you should convert them to numbers that are closer to actual DPS. The easiest way to do this is to multiply the normalized weights by the weight for intellect that the gearing strategy calculated, which is 11.25. Then, you need to manually tinker with the "Offset". You can do this by clicking the "Benchmark" button and see how far off the DPS is on average. Keep changing it until that number is as low as you can get. I did this for these weights and ended up with this:
Using those, the rankings for the trinkets aren't all wacky - the different item level versions of items all rank in order. The actual ranking of special effects will not closely match simulation results, though. As you can see, the suggestions from the machine learning result will get you about 10x closer to the simulated DPS than these custom stat weights. But, you did say that you wanted to sacrifice a bit of DPS in this situation to get more haste.
The problem you are going to run into here is that stat weights have no way to tell the optimizer: get just a little more haste, unless you also set a cap on haste. These weights are going to dump a ton of mastery from your gear and replace it all with haste.
Oh, and also - that "KynBalSimComp" rotation is some old debug rotation. No idea why that is in the public options. You shouldn't use that for gearing strategies - it is bad. Use "Default" and then the machine learning will work waaaay better. I'll get that deleted out of the public options.