win-rate of the current agent vs a snapshot of itself at training start (50% = no better yet, climbing = learning)
// what it learned to value
Each bar is how much the agent values a line pattern. An OPEN line has both ends free (a live threat); a BLOCKED line is walled on one side. Left = its own lines (it builds these, green); right = your lines (it avoids them, red). Bars shift as it trains and as you change the rewards.
// what this agent learned
// its playbook, in words
Puts the trained agent through set tactical positions, records what it actually plays and how it ranked the alternatives, then has DeepSeek turn those decisions into rules you could follow yourself.
// use your agent
Share the code so a friend can paste it into Gomoku and play your trained agent.