What is meant by probe accuracy ?

[alxndrTL]

Hello,
I have read your paper Linear Latent World Models in Simple Transformers:
A Case Study on Othello-GPT and I just wanted to be sure about something :

What is exactly meant by probe accuracy ? (which goes up to 99.5% at layer 6 and 7) ?
Is it the average cell error of the model ? Like it is able to predict the cell state correctly (empty, mine, yours) 99.5% of the time ?
Or is it the board accuracy (all 64 cells correctly predicted) ?

I think the answer is the former but I wanted to be sure.
Thank you very much your work !