This is a nasty one. It's not just conscription and vassalage that are involved - feudalism also has a key role. The problem is that it starts out without either feudalism or vassalage. EITHER of them (they are OR requirements) enable a couple of wonders (Field of the Cloth of Gold, and King's Tournament). When it's evaluating the civics it does so one category at a time (from the starting state), and it's counting the value of being able to build those wonders on BOTH Feudalism AND vassalage (they are in different categories). Both of them score quite highly due to these buildings, so it switches into (both of) them.
That switch is slightly flawed (it only needed one of them, and had it checked for combinations involving one but not the other it would have found better combos). However the real problem arises a few turns later when it next evaluates civics. Now the base state has those buildings enabled by ANOTHER category when it evaluates each category individually, so it doesn't include their value as something it would lose by switching in either evaluation (it only loses it if it makes both, but it can't see that). As a result vassalage and feudalism both look poor, and conscription comes out ahead.
Rinse and repeat.
To solve this robustly we'd need to evaluate sets of civics as a whole (not categories independently), but for each choice in a category you would then need to evaluate possible choices in every other category - so not just difference from base state. This rapidly leads to a combinatorial explosion in the amount of calculation to do, which would be utterly unacceptable in terms of turn processing time.
I think for now, pragmatically, I will simply consider any buildings that are civic-enabled by choices in more than one category (these may be the only ones - not sure - it's certainly rare though) as providing half value regardless of whether you already have them enabled or not, from a civics standpoint. Obviously this isn't true, but it's probably the least-worst solution for now.
Long term, the civic evaluation process needs to change since it's inherently non-linear due to these inter-category-couplings. Some sort of simulated annealing approach or similar (using background threads probably) is likely needed.