A NaN That Was Quietly Picking K

I spent a chunk of the fortnight on a deep review of the HSI topic-modelling platform, and the most satisfying bug was the kind that hides in plain sight: the recommended number of topics, K, was being chosen by a max() over an array that secretly contained NaN.

The culprit was normalized pointwise mutual information. For a pair of topics with perfect co-occurrence, p_ij → 1, the NPMI numerator −log(p_ij) → 0, and you land on 0/0 = NaN. That NaN propagated into npmi_mean, and “argmax of an array with NaNs” silently fell through to K = 4 for several scenes — a fake recommendation that looked principled.

With NPMI handled as its analytic limit and non-finite values filtered, recommended_K became a genuine argmax again — and the honest result is that the score monotonically favors small K (perplexity rises and stability falls as K grows), so K = 4 is now earned rather than a NaN artifact.

The rest of the review was less dramatic but mattered more for users: I migrated the entire benchmarks subtree and the workspace wizard to react-i18next, so the EN/ES toggle now covers the whole explore-and-benchmarks surface (technical terms and symbols kept verbatim); fixed a HIDSAG explore crash from a type that didn’t match the served data; corrected several conceptual and equation errors a math review caught; and added the funding and data-provider acknowledgements where they belong. Deep reviews rarely produce a headline — they produce a product you can trust.