“What would count as evidence about whether the author wrote the conclusion down first or at the end of his analysis?”:
Past history of accuracy/trustworthiness;
Evidence of a lack of incentive for bias;
Spot check results for sampling bias.
The last may be unreliable if a) you’re the author, or b) your spot check evidence source may be biased, e.g. by a generally accepted biased paradigm.
In the real world this is complicated by the fact that the bottom line may have only been “pencilled in”, biased the argument, then been adjusted as a result of the argument—e.g.
“Pencilled in” bottom line is 65;
Unbiased bottom line would be 45;
Adjusted bottom line is 55; - neither correct, nor as incorrect as the original “pencilled in” value.
This “weak bias” algorithm can be recursive, leading eventually (sometimes over many years) to virtual elimination of the original bias, as often happens in scientific and philosophical discourse.