When using the Baby model in the first release of FaceReader 9, the automatic analysis did not detect any negative valences across the whole dataset. This resulted in a misclassification of negative and very negative affect as positive valence.
Fortunately the video does not need to be re-analyzed. This is because the valence can be calculated directly from the action units.
This issue has now been fixed in FaceReader 9.0.17 which can be downloaded via MyNoldus.