Blueprints & databases

Checkpoints & long runs

• A preflop training run runs as its own process. Closing the browser tab or the window does not stop it; when you reopen the app, the progress is still there.
• Quitting the app entirely does not stop a running preflop training run either. The next time the app starts, it finds the solve, shows its log and progress, and lets you stop it cleanly (a checkpoint is saved).
• Runs keep two checkpoints: the newest and one spare. When a new checkpoint has been written completely, the oldest is deleted. A run can always be resumed, and disk use stays bounded.
• To continue a run, use Continue loaded checkpoint and Additional iterations. The run keeps the checkpoint's own configuration. When continuing in the GUI, checkpoint targets are additional iterations from the loaded checkpoint: a target of 2,000 from a 10,000-iteration checkpoint saves at 12,000. For a new run, targets count from zero.

Keep an experimental control you can actually reproduce

A checkpoint continues the learning state; a blueprint is the strategy file you browse. Save the configuration of the tree and the identity of the run with any result you intend to compare. Changing rake, locks or legal actions changes the game, so it is not an ordinary continuation of the old checkpoint.

When comparing checkpoints, specify averaged or current strategy. A current mix can fluctuate while its average moves slowly. Historical EV values and final-policy evaluations are also different measurements; use a consistent evaluation for an EV-loss comparison.

Before a longer experiment, load the original run through the supported UI and inspect its settings. Continue from the loaded checkpoint’s own configuration. For a different tree, create a separate run and folder instead of repurposing the old file and losing the control.