Did you use the archived blockchain for training?
This is why I really think that steemit is going to shine. It may be an immature social media platform, but it's a fantastic AI incubator. Not sure if you already saw it, but this old post of mine is relevant.
https://steemit.com/blog/@remlaps/a-far-fetched-prediction
I hadn't seen your post, thanks for pointing me to it. Good thoughts!
I did use the archived blockchain; I trained my model on about a month's worth of posts before the bot went live. I think a week of data would have been fine, but all the data was there - why not use it?
The cool thing is that all the data is still out there: I can simulate any voting algorithm I want. I have a lot of work to do before this is operational, but I'd love to use this to publish a set of benchmarks that people can use to evaluate their own curation (manual or automated). The Steemwhales.com curator trending lists already play this role crudely, but because you can't tell which of those is autonomous or whether their algorithms are adaptive, they don't really give you the information of a standardized benchmark.
Anyway, I'm rambling. I like your thoughts.
It's a perfect training ground for a genetic algorithm.