Models and metrics · 08

Choosing Fast, Balanced, Deep, and Human SL

Models shape speed, strength, and style; search controls stability. Read win rate, score lead, visits, principal variation, and ownership together.

Choose by task, not by a stronger-sounding name
Fast suits whole-game scans and interactive previews. Balanced is the everyday review default. Deep targets a few critical positions with higher-quality search. Human SL uses human-game style training to explain plausible human choices; it is not simply the strongest win-maximizing model.
The four tiers are not one leaderboard. Speed, depth, and human likeness are separate axes.
Visits measure search budget, not model size
More visits on one model generally stabilize candidate ordering, but cost more time and usage. Scan the whole game lightly, then concentrate the budget on two or three turning points. A larger model and more visits are related but not interchangeable.
Go wide before deep: use a low budget to find questions and a high budget to answer a few.
Read five fields from one consistent perspective
winrate estimates outcome probability, scoreLead the expected margin, visits the search effort, pv the principal variation, and ownership the point-by-point territory estimate. Candidate prior and policy add context. Before comparing, confirm whether values use Black's or the side-to-move perspective.
A candidate is not one score but a bundle of evidence whose fields constrain each other.

Model sources