I was reminded about the problem of determining canon in Tolkien's legendarium while reading about his sketchy treatment of the Blue Wizards (which may itself have been superseded in his later writings):
What success they had I do not know; but I fear they failed, as Saruman did, though doubtless in different ways; and I suspect they were the founders or beginners of secret cults and 'magic' traditions that outlasted the fall of Sauron.
Quoted in Encyclopedia of Arda, 'Alatar', note #3
One way of interpreting this talk of suspecting and not knowing is to treat it all as equivalent to 'I haven't decided', on the basis that Tolkien could have laid either option down by fiat had he chosen to. However, that may not be an accurate reflection of his creative process: his project, through all its revisions, its discarded ideas and its growth 'in the telling', had a certain recognisable consistency, in that he laboured under certain chosen constraints—with the result that ideas conducive to the development of a self-consistent, epic, myth-like world-history in which pleasing languages could be spoken was 'good', or perhaps 'fitting'; ideas conducive to kitchen sink drama, for example, were the opposite. When people argue about the 'true' identity of Tom Bombadil, or whether Balrogs have wings, they're generally doing it already under the broad understanding that whichever interpretation they favour is fitting (although of course what's fitting is also up for debate).
Now before I get carried away with trying to retrofit the constraints of creation into constraints of interpretation, there's an obvious rejoinder to the effect that any possible 'author as supreme arbiter' criterion is untouched whether the scaffolding stays up or comes down: had Tolkien decided to pit his Blue Wizards against Great Cthulhu and Lord Palmerston in a Sherman tank, this would still be true in and of his fictional world, however much we might wish otherwise. We should simply have to concede that he had abandoned or modified his conception of the fitting, or that we had failed to perceive what it was to begin with—i.e. the superstructure of fit is inferred according to the written facts, not vice versa.
Certainly, the prospects for a full epistemology of fictional worlds along these counterfactual lines look pretty grisly. A hypothetical 'perfected version of the author's project' which perfectly fits its chosen constraints, though perfectly conceivable as the end-point of an uninterrupted series of revisions and expansions, and hence as the culmination of the author's project (if we assume, not always plausibly, that that was unchanging), remains an object of speculation and at best very little certainty—so we risk the conclusion that there are truths about fictional worlds, but we can't know them, and they may contradict actual published statements. (I've previously shuddered at similar prospects when wondering about how we identify continuity errors and other such mistakes.) I suspect a more likely view, however, is that even if there are truths about the perfected world-version, typically we're not talking about that world-version but about the imperfect version(s) we encounter in actual writings: just as something may be explicitly true in a draft but explicitly false in its revision, so it seems reasonable to deny that any hypothetical perfected world-version should have any bearing on truth and falsity in imperfect works.
Nevertheless, I think there may be a legitimate, though more moderate, role for criteria of fit in the assessment of truth in fiction. Rich Cochrane has suggested a 'possible worlds' approach to fiction which makes contradictions permissible; though in conversation with him I wondered whether this might oblige us to swallow a wildly divergent fictional world as a 'possible' version of a familiar one, such as changing enough properties of Star Wars: A New Hope to turn it into Blade Runner. I think shifting or expanding authorial intention from the product of the author's thought itself to the product's 'fitness' may be of help here. We may have different judgments – even if 'we' are George Lucas at different stages of his life – about whether Han's or Greedo's shooting first would be more fitting, but we can surely agree that both are far more fitting than their embracing like brothers. Worlds in which this happens are ruled out as possible Star Wars worlds: it's false that Han and Greedo hug.
I think there are two likely objections to this line of thought. One is that it's redundant, because it still boils down to counterfactual inferences about authors' likely intentions (and if anything it invites looser ones): 'If Tolkien had ever made up his mind about the Blue Wizards, he would have decided...' The other is that it's too demanding, in that it adds a new layer of things for us not to be certain about.
The reason it isn't redundant is that it resolves what would otherwise be paradoxical: a legislative statement (given the assumption that what Tolkien asserted about his legendarium is necessarily true, at least until he changed his mind) that incorporates uncertainty. Tolkien said the Blue Wizards may have formed cults. His statement presupposes an in-world fact of the matter, a (fictional) truth one way or the other for him not to know. However, if our rule is that 'if Tolkien reports P as an event in his legendarium, then it is an in-world truth that P happened', we run into trouble if we take it to be an in-world truth P that 'the Blue Wizards may have formed cults'. While it's of course trivially true that either they formed cults or they didn't, the uncertainty of the matter can't itself be the in-world truth of the matter, because then there's nothing for Tolkien to be uncertain about: according to this interpretation of the rule, he's legislated an in-world indeterminacy. However, this conflicts with the apparent presupposition of his uncertainty that there is an in-world fact of the matter, one way or the other.
Now the obvious move, for anyone not keen to abandon a binary legislative/non-legislative distinction, is to invoke multiple (possible) fictional worlds: there's a version of Tolkien's legendarium in which the Blue Wizards formed cults, and a version in which they didn't, and what Tolkien was uncertain about was not so much which was 'true' (for both sprung from the mind of the same author) but which to endorse as canon and continue developing. This is probably a reasonable characterisation of Tolkien's predicament, but it isn't wholly faithful to his words: he spoke as though there were a single fact of the matter which he had not fully uncovered—which of course suggests it as an in-world truth that there is (Tolkien being the legislating author). But how could there be, if Tolkien hadn't made up his mind about it?
A notion of 'fit' may help make sense of the matter. We ordinarily read with a certain imaginative grasp – reminiscent of Gadamer's 'fore-understanding' – which gives us some pretty reasonable prima facie assumptions, e.g. that a building described as a palace has more than one room. Authors may play games with our expectations, or confound them accidentally, but that doesn't render them unreasonable. Now of course, when it comes to wondering whether Balrogs have wings we find ourselves far out of our depth and have to rely on the wording of the text; and perhaps there are subtle questions of best fit which even the author behind the fictional world in question can't answer without exercising a measure of creative judgment—so in practice, I grant, the notion is of rather limited epistemic use. But it does help to explain Tolkien's uncertainty: there was a sort of hypothetical 'best fit' which had some shadowy reality, even though he was free to choose to overwrite it.
So as for its being too demanding—well, it does raise the spectre of there being facts about fictional worlds, albeit merely provisionary ones, which aren't wholly accessible even to the worlds' authors. That does seem quite a lot to swallow, but consider the case of video game worlds: one point I'm fond of making with regard to efforts to create a timeline for the Legend of Zelda series is that these efforts tend to assume the successful end of each game is the 'true' continuation of the storyline. (Whereas when considering games which are set up to offer 'multiple plots' one just has to deal with multiple possibilities. Given that one of Chrono Trigger's ending scenes is accessible only if the party loses to Lavos, the final boss, is that somehow more significant than a common-or-garden in-game death?) In fact, Link will be killed by some wandering Octorok far more often than he vanquishes all the game throws at him; the Zelda II death screen even gave a storyline hint with its 'Game Over—Return of Ganon'. Which is the 'true' outcome? One answer: all of them; they happened to us; we remember reloading our saved games. Another answer: the outcome which can best be called the 'true' one is that which continues the story of the hero to its 'proper' conclusion, rather than letting it be unceremoniously cut short.
Even so, I imagine that presenting anything as apparently evaluative as 'best fit' in fictional worlds as a kind of quasi-objective reality may very reasonably invite objections...