I think you also have to step back bit when considering ordinal data - what is the nature of the real-world parameter that you are attempting to represent?
With an ordinal scale, it is invariably a means to try and simplistically map discrete value labels onto sub-ranges of a continuously variable (analogue) parameter, typically a subjective one. So if, for example, the parameter you wish to measure is "how do you feel your temperature to be?", you might have ordinal values ranging from "desperately cold" all the way through to "desperately hot". But the underlying parameter, subjective or not, will itself be an analogue one.
So the process of a person deciding what their answer to the question is, will itself be a translation of their real, analogue, temperature sensation, to the ordinally labelled value that seems a "best fit" for them at the time. It will get tricky to answer if their real, analogue, sensation feels to be close to a boundary between two ordinal labels , and in such boundary cases will also be much more liable to bias, depending on how they are feeling at the time - boundary condition decisions are often subject to such in-the-moment biases.
So any subsequent mapping of ordinally scaled readings onto an analogue scale, is only any use if it can adequately reverse map back to a good approximation of what the person's original analogue sensation was at the time they answered the questions.
And this of course, is about just one question, about a single category being queried of the person. Once you have multiple questions, sub-categorised or not, the potential for error between reality and measured will be high I would think.
There is an engineering technique called "fuzzy logic", which is a way of mapping natural-skills solutions into engineering ones, and the one thing for absolute sure, is that the mappings are highly unlikely to produce nice neat linear translations - if they did then fuzzy logic would be not have been necessary. (Note: I'm not sure if fuzzy logic is used so much these days).
