Under the assumption that a text has an architecture, that architecture is accessible only though abduction. The proof then becomes the pudding, because the role that the pieces play just so in the text validates the structure in the reconstruction.
In the emplotment of historical texts, the facts play multiple roles in the discourse with the audience. They provide points of stability through agreement ("The Normans conquered England in 1066") or they teach novel facts and interpretations (whose novelty is of course dependent on the level of expertise of the readers). The whole point of offering yet another emplotment is to correct the existing ones, and that requires either new facts or new interpretations (and preferably both).
Writers therefore have a model of the questions and the lack of information that the readers have and they write toward that model: to answer the questions and to supply the information that will make the readers agree with their architecture.
Perhaps it is too much to say that each text creates a language for a special purpose (or a domain-specific language, to use that technical term), but probably each discipline and sub-discipline does. And that functions both as short-hand (writing takes time even with word processors) and as a model for thinking about the right types of problems.
We would like it to be the case that the architecture and its realisation using the facts and the interpretative fragments is mechanically verifiable. That makes it easier for the writer to adjust the whole without breaking the parts, to experiment with attitudes and stances, and to adjudicate if choices in one place amount to consistent stances in other parts.
What is decidedly necessary is a way to rule out alternates and make clear why they are being rejected. This is part of dealing with the questions and the expectations of the audience. Didn't Odo of Bayeux commission the tapestry to suck up to his half-brother that he had insulted by siding with his enemies? Well, no, and here is why. That is the function of the precedent and the literature review.
This is also useful because it allows rerunning the narrative, if we think of it as a conceptual program, at a later point in time with altered insights. If the dating shifts, if the source becomes discredited, then we can see what that means for the remainder of the argument. Perhaps this is then no longer the stance of X, who may no longer be able to participate in the argument; but it may approximate the stance that X could have taken, as best as we know, given the new information.
The drawback of all of this is of course that the abductive process repeats and re-occurs at the various levels of the interpretation, whether it be the text as a whole, the individual claims, the specific meaning of the words used, or the scripts employed to make progress on the problems.
The problem is also that we primarily wish to rule out narratives, but the number of wrong narratives is of course too large to manage properly. So we need succinct ways of explaining plausible candidate narratives that we can then rule out equally succinctly, as a representational problem.
