I think people have misconceptions about algorithmic feeds.
You can measure what inflames people or you can predict what inflames people, measuring is always going to be more acccurate.
… and you can do that measuring by counting boosts or (on HN) taking the ratio of comments/votes. 0.5 is about the median, articles that go above 2 are well in the danger zone.
In my mind the chronological feed with boosting is pretty toxic on its own, an “explore” feed based on what is getting boosted is even worse.
YOShInOn’s main feed I look at every day is based on (1) clustering and (2) prediction of my upvotes and downvotes, I don’t particularly upvote toxic articles, in fact, I make a point to down vote them so my feed is calming and low in sensationalism. (Though there was a time I was into articles from The Guardian about how the UK is going to hell…). So that kind of algorithmic feed is not inflammatory.
If you use “boost count” or judged sentiment or the like you could make a model that amplifies or diminishes angry content. The thing is voting and boosting is not all bad, I post a lot of flower pics to Mastodon because they get boosted, up to a point discussions on HN are a good thing, If you used a sentiment model to suppress angry content in a voting/boosting process that might be very good.
Sometimes I think about making an “angry toot” model but I can’t stand the thought of looking at 5000 angry toots (well, i have thought about testing a weight loss plan based on inducing a psychogenic fever and maybe that would work…)