Eve library

Kuan Yin — Monitoring & Observability

Guide

Revision 1

On this page

Technical doc 32.16.1.14. Target stack in §32.17.2; signals from @kuanyin/dharma-analytics and performance-sla.ts.

Two classes of signal must be watched: system health (is the platform up and fast?) and community health (is the platform doing its job — keeping people safe and helping them grow?). Kuan Yin is unusual in treating the second as a first-class, alertable metric.

System metrics (Prometheus)#

  • Latency percentiles per path (intent, toxicity, live-stream protection) from computeLatencyPercentiles; alert on SLA breach (slaAlert).
  • Throughput, error rate, circuit-breaker state, degradation level.
  • Event-bus health (evtImplementEventBusIntegration → healthScore, latencyP99Ms, deadLetterQueueSize).
  • Cache hit rate (CacheRepository.getStats), rate-limit throttle rate.
  • Model inference latency and drift (PSI/KL/JS) from the evaluation harness.

Community-health metrics#

From dharma-analytics/community-health-metrics.ts: compassion index, mindfulness score, connection depth, conflict-resolution rate, rehabilitation success rate, community coherence, positive/harmful ratio, newcomer retention correlation, creator satisfaction. Surfaced on dharma dashboards (materialized views) and in analytics.report events.

Tracing & logs#

Distributed tracing (Jaeger) spans the gateway → engine → repository path so a single moderation decision can be followed end-to-end. Logs are structured and correlation-id stamped; decision logs record the rationale (InterventionRecord.contextSnapshot, AuditLog.description) for transparency and appeals.

Alerting#

Condition Severity
Crisis detected, not routed to a human P1
Shield failed to activate for a flagged target P1
Intent/protection SLA breached sustained P2
Community temperature crosses intervention threshold P2
Drift exceeds threshold on a live model P2
Dead-letter queue growing P3

Predictive wellness (predictive-wellness.ts) raises proactive community.alert events for users- or communities-at-risk before an incident, so on-call can intervene early rather than clean up after.

Dashboards#

The performer wellness dashboard, community temperature dashboard, and dharma analytics dashboard are the operator-facing views; see the user docs for the performer- and moderator-facing equivalents.