When making a recommendation a level of confidence is give to the accuracy of that recommendation.

When recommendations are deferred to the user and the level of confidence changes through time then a decision aid may wish to send a new/updated recommendation with a higher (or lower) confidence.

This (or another) section could do with some narrative as to the expected model for achieving this: i.e. does the decision aid send a new recommendation and expect the application to detect that a previous recommendation has now been superseded.

The Percentage type does not express how it is to be represented when mapped into a PSM. For example it is represented by a double in DDS. The specification should give advice on whether a percentage is expected to be mapped from 0-100 rather than 0-1 particularly as the type is often used to represent attributes which are named probability.

The description of the modelProbability attribute in the Confidence class states that this relates to probability given model inputs. There are circumstances where this it would be appropriate for this to be conditional on the broader context as well (i.e. not just on those parameters that are model inputs).
This is about whether the case in point is representative of the training (or design space). Whilst an estimation could be made on how well represented the inputs were in the training set (e.g. are they near a corner case of which there are few prior examples), or it could be known that the current example is from a very different context (e.g. geographical location) from the training set. This would lead to model probability being set low on a somewhat different basis.

Recommendations should enable the supply of a rationale or rule as a descriptor to aid an operator in particular.

Many of these attributes (e.g. Algorithm Type) are not naturally Quantities
Descriptor or Quantifier would be better classifiers for the type