Suppose that originally a sentence has a matrix clause in the simple present tense and a subordinate clause in the past perfect tense. Then, if I stand at a point in time after the event described by the matrix clause and retrospectively report this sentence, the matrix clause can be shifted into the simple past, while the subordinate clause must remain in the past perfect.
Now suppose another sentence originally has a matrix clause in the simple present tense and a subordinate clause in the simple past tense. If I again stand at a point in time after the event described by the matrix clause and retrospectively report this sentence, the matrix clause can also be shifted into the simple past, but the subordinate clause will now be shifted into the past perfect.
Therefore, when additional temporal expressions and other contextual temporal information are relatively limited, we face the following problem: because two subordinate clauses that originally had different tense forms are mapped onto the same tense form after the matrix clause is shifted into the simple past, the past perfect in the retrospectively reported sentence may correspond to two different possible original tense states. Consequently, when we attempt to reconstruct the original temporal information in reverse, we encounter ambiguity. In other words, this process does not have unique recoverability at the syntactic level.
I am not sure whether the following two pairs of sentences can illustrate the idea intuitively:
Original state
A. I know that he had left before I arrived.
B. I know that he left yesterday.
Retrospective report from a point after the time of the matrix clause
A'. I knew that he had left before I arrived.
B'. I knew that he had left yesterday.
In other words, there seems to be no way to avoid this problem purely through grammar. To prevent this loss of information, we would need to supply additional temporal information, such as explicit time expressions or contextual cues.
(This is a question, not merely an observation. If you agree with my point, could you provide relevant literature to show whether there is scholarly support for this view? Conversely, if you disagree, please cite sources that challenge it.)
P.S.
If expressed in more abstract and rigorous terms: the rule of sequence of tenses is not an injective mapping because there exist different original tense configurations that are mapped onto the same resulting tense configuration. Since such a mapping does not satisfy the necessary condition for having an inverse function, it lacks an inverse mapping. Therefore, the original temporal information cannot be uniquely recovered.
Incidentally, this problem may become even more apparent when analyzed within Reichenbach’s theory of tense, since the temporal relations between Speech Time (S), Reference Time (R), and Event Time (E) provide a more explicit representation of the information that is lost during tense shifting.