current impl. uses only the fulfilled/active ratio
[ ] no diff for unfulfilled predictions (should "worsen" the anomaly)
[ ] confidence: no penalty for too much predictions
eg. you predict 100 possible outcomes, how useful/trustworthy is then when 1 of your predictions is fulfilled?
[ ] can we use some SDR properties in the process?
context growing #46
TM should do inhibition
[ ] confirmed, deep down the line predictions (100 steps confirmed with the predictions) -> bigger anomaly on breakage then if on beggining; "momentum" of confidence