This is an interesting service. A few suggestions here:
* Use a symbol to mark the last move before the questions (this can be worked out most of the times but in real game this information is available)

* Consider remove problems with very obvious answers (such as stone under atari, or forcing moves in the middle of a common joseki)

The system apparently learns from very modern pro games. It doesn't like old fusekis. And it does not allow more creative moves. I am a 5-dan but I scored very badly initially until I started playing the most common moves. Occasionally, a perfectly reasonable move is considered 'WRONG'.

The system will clearly become more useful if the database backing it is much larger, and if a lot of people use it improving the accuracy of the difficulty ratings of each problem. For the former I can help - if you are interested in using my Go4Go database to back your system, please contact me.

- I like the idea of showing the positions again that I got wrong, but sometimes it seems too recent, having something like 10 others between them would feel better before you see something twice(obv others might agree)

- The rank seems to be pretty inflated, the site ranks me 5 stones stronger than I actually am on kgs, not top priority, but a little more correlation would be nice.

- I understand that that the sample games seem to be from pro games mostly, but especially weaker players like me see a lot of whacky openings in reality. Right now it seems like the problems are really good for teaching you to not make mistakes, but teaching you how to punish your opponents mistakes seems like a desirable skill to develop as well.

The site learns how to rank players, based on the ranking information provided by other players. So it should improve over time. And of course, since its measuring fuseki knowledge, how good/bad one is at fuseki can make a difference too. For the time being most 5kyus and weaker will probably get slightly inflated ratings.

What constitutes enough data for your model?

I've 56 games for this position in my database. White's approach here is rather a lot less common than a and b here. I'm curious what your cutoff is here for viable moves for white? I really like the site idea, I'm more curious about the algorithm here than anything else.

Indeed the site is more about learning the established and prevalent fusekis. I recently relaxed the requirements for what moves are considered acceptable. A "correct" answer was a move that at very least 1/4th as often played as the most popular choice on the board. Tho even moves that are only played 1/16 as often as the most popular choice was accepted without a complaint, however these wouldn't affect your rating. A bit like, the site is uncertain if its really right or wrong so it ignores the answer. Anything less than 1/16th as often was considered wrong even if there was an example of it being played by a pro. The new relaxed requirements are 1/5th and 1/25th, but they haven't had time to propagated throughout the system yet.

Right now I'm only using games since 2011, but I plan to add a few more years worth of games to give a bit more diversity to possible answers. Tho I dont want to go too far because some older fusekis are considered inferior and not just out of fashion.

Also, you can link directly to the board position on the site from the tsumego sections "link to answers".

011 seems an unusually recent date? Perhaps offer "styles" or groups of particular pros for people? It might be interesting. E.g. someone studying Shukasu might find it interesting to go through fuseki questions based on his games. Ditto perhaps someone looking at Lee Changho or Lee Sedol. I don't know, just musing.

Indeed features somewhat similar to what you suggest here are already planned. I just want to ensure the existing functionality works smoothly first.

Once the site starts zeroing in on a rating, it seems pretty unwilling to revise its opinion. Thus in the three times I've tried so far, twice I've gotten off to a bad start and wound up with a rating around 8k, and once I got off to a good start and got a (more accurate (imo)) rating of 3-4k.

Common fuseki positions are shown from recent pro games, and you are asked to guess the next move. You rank is then estimated based your answers. In turn the positions shown to you are based on your rank. So after the first few positions you should start getting position that you can answer correctly 80% of the time.

There was a HUGE bug, which caused the first 100 problems to be totally random for each users, only after those first 100 problems did it start providing users with problems appropriate for their rank. It should now be fixed. Sorry about that.