Prefekt | Prefekt is an Android SharedPreferences library for Kotlin | Android library

For non-polish readers: you are given those books only when preparing your solution, you won't have access to them during test. If you try to bundle them with binary somehow those would exceed 10kb limit hence you need to compress information somehow.

I would go for Naive Bayes classifier by default for a simple solution .

Due to time constraint I would go a little bit different route though.

Read all files in and tokenize them. Would be easiest with Python's split functionality (and whole program would be easiest, time constraint probably won't be a problem). Split on whitespace and punctuation as those are mostly noise and are not representative of texts.

Now calculate how often each of the tokens (words) occurs in each text, e.g. dog occured 15 times in first text and 3 times in another. Save those in three separate dictionaries, if the size of dict exceeds 10kb remove words occurring least frequently and adjust accordingly.

Use 3 unsigned long variables to keep results for each texts to keep overflow in check (it should be enough).

For every input text split it just like above.

For every word check in dictionaries how often those occured for each text and add this to one of 3 result variables. If it doesn't exist just add 0.

Finally return text which gathered "most points" this way. This should get quite a good score.

Naive Bayes with probabilities would work much better but given competition constraints I don't think it is a viable solution.

To do it, you would have to calculate probability of each word for each text and use log operstions during summation to avoid aforementioned overflow, just throwing it out for you to consider, doable but probably overkill.