Oliver Sherouse

If the cache is corrupted, it should be deleted instead of raising an error, to prevent problems like #62.

Allow sequences of length 2 for dates even though it should *really* be a tuple. Contributes to #71.

Corpus.builtins => analysis.nlp Estimator prediction => analysis.ml

https://github.com/QuantGov/quantgov/blob/5c22c6ee9cd1c702561b881bcf5dd38097d7ff99/quantgov/corpora/builtins.py#L123 We need the `term_counts` dictionary keys to be the actual regexes, not the string matches. We should map the one to the other and then generates results from the...

Use simplest model within 1 standard deviation of best score.

At the very least, should have a live check that warns if the index isn't unique. Should do this on loading when that's cheap.

Include a way to get raw text out of a variety of formats. [Textract](https://textract.readthedocs.io/en/stable/) may be the way, but I've had trouble with it in Anaconda.