Online analysis · 02

How to analyze an SGF with KataGo online

From importing a game record to reading graphs, candidate moves, and variations: complete a focused online SGF analysis without installing a local engine.

1. Import the SGF and verify game settings

   Open Read19 Review and choose an SGF file or paste its contents. Before analysis, verify the players, komi, rules, and move count because these settings affect KataGo's evaluation.

   An SGF stores more than moves. It may include rules, komi, handicap setup, player names, event data, and timestamps. After import, compare the final few moves with the original game, then inspect fields such as HA, KM, and RU. Missing setup stones in a handicap game or an incorrect default komi puts every later graph on the wrong premise. Before creating a public link, remove names and event details that should remain private.

2. Match search effort to the question

   Use a lighter search to scan the whole game, then reserve more analysis for positions you genuinely want to study. More visits usually stabilize rankings, but they do not turn an AI result into the only human move.

   Allocate search to questions, not evenly to every move. The first pass only needs enough stability to reveal major swings. Once key positions are identified, increase effort there. If several candidates exchange places while remaining close, they may genuinely be comparable; more search is for measuring the gap, not manufacturing a single answer. Record the analysis preset you used so later comparisons retain their context.

3. Read the trend before local variations

   Use the win-rate curve for changes in expected result and the score-lead curve for the approximate margin. Select a turning point before comparing the played move, candidate moves, and variations instead of chasing every number from move one.

   Begin with the move number on the horizontal axis rather than the highest or lowest point. Pause before a swing and confirm whose turn it is, the current score lead, and the weakest groups on the board. Handicap games and large leads particularly punish win-rate-only reading: win rate may remain near an extreme while score lead still shows whether the game is closing safely. In a close game, one or two points can move win rate sharply without the shape collapsing.

4. Test branches with TreeView

   Open a candidate branch in TreeView, play through it, then return to the fork and create a branch for your own reply. The goal is to preserve comparisons and recognize urgent points, move order, and weak groups—not memorize one main line.

   A variation is a branching tree, not a video to memorize. Follow the main line only until its purpose becomes clear, return to the first branch, and try the reply you would most naturally play. Observe whether KataGo keeps the same direction or leaves immediately. Testing at least one natural reply outside the principal variation tells you whether the recommendation depends on a precise tactic or represents a more durable whole-board idea.