Applying machine learning to a guessing game?

Like you've said, this problem can be described with a HMM. You are essentially interested in maintaining a distribution over latent, or hidden, states which would be the true quantities at each time point. However, it seems you are confusing the problem of learning the parameters for a HMM opposed to simply doing inference in a known HMM. You have the latter problem but propose employing a solution (Baum-Welch) designed to do the former. That is, you have the model already, you just have to use it.

Interestingly, if you go through coding a discrete HMM for your problem you get an algorithm very similar to what you describe in your graph-theory solution. The big difference is that your solution is tracking what is possible whereas a correct inference algorithm, like the Virterbi algorithm, will track what is likely. The difference is clear when there is overlap in the 5% range on a domain, that is, when multiple possible states could potentially transition to the same state. Your algorithm might add 2 edges to a point, but I doubt that when you compute the next day that has an effect (it should count twice, essentially).

Anyway, you could use the Viterbi algortihm, if you are only interested in the best guess at the most recent day I'll just give you a brief idea how you can just modify your graph-theory solution. Instead of maintaining edges between states maintain a fraction representing the probability that state is the correct one (this distribution is sometimes called the belief state). At each new day, propagate forward your belief state by incrementing each bucket by the probability of it's parent (instead of adding an edge your adding a floating point number). You also have to make sure your belief state is properly normalized (sums to 1) so just divide by its sum after each update. After that, you can weight each state by your observation, but since you don't have a noisy observation you can just go and set all the impossible states to being zero probability and then re-normalize. You now have a distribution over underlying quantities conditioned on your observations.

I'm skipping over a lot of statistical details here, just to give you the idea.

Edit (re: questions):The answer to your question really depends on what you want, if you want only the distribution for the most recent day then you can get away with a one-pass algorithm like I've described. If, however, you want to have the correct distribution over the quantities at every single day you're going to have to do a backward pass as well. Hence, the aptly named forward-backward algorithm. I get the sense that since you are looking to go back a step and delete edges then you probably want the distribution for all days (unlike I originally assumed). Of course, you noticed there is information that can be used so that the "future can inform the past" so to speak, and this is exactly the reason why you need to do the backward pass as well, it's not really complicated you just have to run the exact same algorithm starting at the end of the chain. For a good overview check out Christopher Bishop's 6-piece tutorial on videolectures.net.

Because you mentioned adding/deleting edges let me just clarify the algorithm I described previously, keep in mind this is for a single forward pass. Let there be a total of N possible permutations of quantities, so you will have a belief state that is a sparse vector N elements long (called v_0). The first step you receive a observation of the sum, and you populate the vector by setting all the possible values to have probability 1.0, then re-normalize. The next step you create a new sparse vector (v_1) of all 0s, iterate over all non-zero entries in v_0 and increment (by the probability in v_0) all entries in v_1 that are within 5%. Then, zero out all the entries in v_1 that are not possible according to the new observation, then re-normalize v_1 and throw away v_0. repeat forever, v_1 will always be the correct distribution of possibilities.

By the way, things can get way more complex than this, if you have noisy observations or very large states or continuous states. For this reason it's pretty hard to read some of the literature on statistical inference; it's quite general.