Compatibility with the future is a guaranteed "cannot complete a valid Candidate Recommendation" clause.
PF definitely sympathizes with your desire to have the content create a responsible document object model regardless of the AT uptake of the information at the moment. We want that to. But the way to achieve this is to have a concrete proposition as to the information that has to be available in machinable form; not to say "anything the user may infer from the rendered content must be encoded in a machine-recognizable notation." That isn't going to work. The way to success is, for example, take a label. The requirement here is indicated by the dialog:
Q1: Can the user do something, here?
A1: Yes.
Q2: what can they do?
A2: <action>
Q3: what in the user experience tells them that?
A3: This <label>.
If the association of A3 as label for the action is machine recognizable, and the user should be able to recognize A2 from A3, then we are done.
That is the general pattern to be replicated across the essential information to answer questions of
"Where am I?"
"What is _there_?"
"What can I do?"
.. at a fine enough grain so that
the answers are always in the context
that the use can recall or associate
with their current place-in-browse (e.g. focussed object or reading cursor).
Proposed Change:
Build on contribution expected from IBM as to how to reword 4.1. We need to connect several things: information required for an 'informed' user browse; machinable notations so UA can map to API the AT understands; format specs that afford the ability to have a shared understanding among author, author-automation, user-automation, and user.
Spell out information requirements in Principle 3 -- what the user needs to be able to understand (including about what they can do).