About readitout.net

This project was created by Henry Hemming and started on July 12th, 2025.

Problem Details

The ruleset is Japanese, with a 6.5 point komi.
Problems are generated by KataGo during self-play by random sampling from policy.
To ensure quality, each problem is checked by the KataGo model kata9x9-b18c384nbt-20231025. The model uses 1,000 playouts to verify which side wins, and confirms that playing the game out with 30 playouts per move also results in the same winner.
The collection contains over 41,000 unique problems, somewhat evenly spread out over the ranks from 20kyu to 9dan. Each problem can be displayed in any of eight orientations. More problems will be generated and added daily.
Problems are assigned a rating by having 32 different KataGo models attempt to solve them. These include b18c384nbt-humanv0 and kata9x9-b18c384nbt-20231025.
This process provides a rating for both the models and the problems. We then use europeangodatabase to look up how 1 rank difference impacts winrate, which is caused by reduced variance in points by stronger players. This provides us with the curvature of rating to rank, which is then anchored by the b18c384nbt-humanv0 model's 1-kyu rating to the 1-kyu rank, and a similar anchor is used for the 10 kyu rank.

Your rank changes quickly at first, but rank change based on each tsumego result slows with each problem you fail and speeds up when you answer correctly. With a cap at how stable a rank can get, which means there is a limit how fast your rank can decrease but not how fast it can increase.
Once your rank stabilizes, it will take roughly 10 wins to go up a rank or 10 losses to go down one.
Even high rank problems can be guessed correctly, but not frequently enough (4/5) to maintain your rank.
To help you learn, problems you fail are shown again later, whenever the system thinks you have 80% chance of remembering the solution.
Answering problems that are from links and not part of your own progression does not affect your rank.

It can be installed from your browser as a standalone app.
The review mode has numerous positions pre-evaluated by kata9x9-b18c384nbt-20231025 with a minimum of 30 playouts, usually hundreds of playouts.
For positions that have not been pre-evaluated, it uses a server to provide kata9x9-b18c384nbt-20231025 with 1 playout to analyze positions.
Upon failure, if you played too fast and you get better results when you think for longer, a slow red screen will appear.
Upon failure, if you failed a repeat problem, for which you didnt spend much time reviewing, and you get better results when you spend more time reviewing the solution, a slow red screen will appear.
The statistics page tracks factors influencing your problem-solving probability. Ideally, time-related metrics should be close to zero, while a positive trend in the "per problem solved" metric indicates learning.