Formula candidates never get typeset in this Space, and that is the case the "ranking stays stable" note doesn't cover.

jev_render.py draws a formula as its LaTeX source in DejaVuSans.
So Jev sees the typeset crop beside a picture of the string \frac{a}{b}, not a fraction.
The module's own docstring says official training used Playwright + KaTeX.

That matters because the 0.70 slider is one threshold across text, formula and table (route() in visualize.py).
Ranking can hold inside each category and the pooled threshold can still break, if formulas take a renderer penalty that text doesn't.

Tables have a smaller version of the same issue. A cell of 40+ characters is clipped to 37 plus an ellipsis, so a correct long cell renders as a mismatch.

The Jev card says to pick the threshold from samples of your own application.
Have you scored the same formula crops through both renderers? Even 20 would show whether formulas need their own threshold.