By reference class I mean "the set of things that are like Subtext that I'd use to get my prior probability of success from, before updating on the specific merits of Subtext (or its flaws)".

I've acquired the term both from previous discussions here on LW and from slightly more formal training in forecasting, specifically participating in the Good Judgment Project.

There are perils of forecasting based on reference classes (more), but it can be a useful heuristic.