Anatomy of one break

This is the full path for one run from the corpus, the one drawn in Fig. 1 on the home page. Ids and names are removed and nothing else is tuned for the page. Both models scored this run without having trained on it.

The curve. For thirty-two minutes the likelihood that side B wins wanders between 0.24 and 0.73, which is what a close run looks like. At 32:40 it reads 0.57. Sixty seconds later it reads 0.22, and in that minute side B loses five units while side A loses none. The run ends nine minutes later.

The call. For a lost fight the model ranks every unit on its side by how badly it was placed going in and coming out, compared with strong play in the same situation. In this one it puts B2 first. At 32:45 side B is pulling back, and B2, low on reserves, turns back the other way.

The verdict. A reviewer who plays at the top watched the fragment around it and answered: a mistake, the model is right. In the same run the model made two more calls. One was rejected outright: nothing was happening, and leaving was the correct thing to do. The other got no clear answer.

Across the first five runs reviewed this way there were fifteen calls and twelve clear answers. Seven confirmed the model. One rejected it. Four said the unit was caught but had done little wrong. Twelve answers from the founders are not an accuracy figure, and we do not use them as one.

They do show where to work next. The model reads being caught as a decision, and some losses are just bad luck. Telling a bad choice from a bad outcome is the centre of this problem, not an edge case, so those four are where the next version starts.