I am not sure if I follow completely Bob but I think you are pointing out an important issue for "semantics immutability" versus "byte/bit- level immunity". If a client retrieves data from two different clients under a byte-level immutability contract a simple equivalence test should be able to verify the byte-level equivalence. Under the semantic immutability contract, a more complex text for equivalence would be required to fit for example the mime-type.
In practice I do not think this is an issue. If clients act under blind faith under either contract they would not text the equivalence. In fact they would usually only retrieve a particular LSID from one service. The blind faith client would process the data as if the data provider is following the contract and no more. The client could not assume byte-level immutability when there is only semantic immutability because it may indeed break the client code. Caching a byte-level representation of data from one call can not be compared with semantic data. If XML is carried in the data all operations must be consistent with XML operations. I do not see this as a problem.
Since in the biodiversity community LSID data payloads would be about a large variety of objects, clients would always need to check the data types before most processing operations. The data type information would be encoded in the metadata but could also be segregated by service provider (but even there for good form the metadata should encode the data type.) The metadata needs to encode both the physical layout of the bits and "use" (there must be a better word). For example, the data could be a Darwin core records, a dublin core records or SDD. All are XML but the legal operations over that XML are different depending on the "use". Some clients could just pass the data through without be concerned about this but other clients would need to process accordingly perhaps ignoring types it knows nothing about.
------
Unrelated to Bob's comment I would like to add a point about digital from birth vs made digital data.
What is data and what is metadata has no relation to being digital or not. There was data and metadata long before there were computers. Galileo studying the time of objects to move down an inclined plane collected data, the time, distance, angle and mass of the objects. At least the time and the distance recorded in his notebooks are data. If we re-represent his data from the notebook in digital format in 2007 so we can process it in an excel spreadsheet it is still the same data. If we just take a photo of the book we might have a different beast but as long as we leave his number as numbers it is the same data. The metadata about inclined plane experiment would include information about the apparatus used. For example he might have bells that ring at different locations/distances of the inclined plane., it might be made of a wooden frame with brass rails. All this metadata tells us about the data, it is data about the data. Similar arguments can be made about specimens. A digital representation of a specimen is still data. No one is arguing that the specimen is a species or a species concept. A specimen glued to paper or in a photo can be assigned to a species concept, meaning someone has said this is an X. As such we can treat it as an exemplar of X. If it is a type we can even say it is a very good example of X but it does not cover the entire concept of X. The image of the specimen can be data. We need not treat it as metadata just because it is digital or because there is an object or event in the world that is now primary representation. Galileo's numbers also existing in the notebook do not make the numbers in the computer any less data. We will want to add metadata to the digital numbers to tell the user that they came from Galileo's notebook.
Bryan