Health and Observability

Observe mBedLM-core as layered signals: serving health, route behavior, contract quality, and workflow outcomes.

Observability Overview

Health is not a single metric. The platform is monitored as a set of linked indicators so operators can identify whether regressions come from serving, routing policy, contract shape, or domain adapters.

Signal Flow

graph TD
  A[Serving signals] --> B[Routing signals]
  B --> C[Contract signals]
  C --> D[Workflow signals]
  D --> E[Alerting and rollback]

Primary Signal Groups

• Serving Signals: endpoint readiness, latency, timeout rate, and backend reachability.
• Routing Signals: route decision distribution, fallback frequency, and escalation trends.
• Contract Signals: schema validation pass rate and malformed response incidents.
• Workflow Signals: plan/execute success rate, step failures, and completion latency.

Logging and Trace Model

• Correlate request/session identifiers across ingress, orchestration, and adapter boundaries.
• Emit structured lifecycle events for preflight, route, inference, and normalization stages.
• Preserve enough metadata for replay and regression investigation without leaking sensitive payloads.

Operator Checks

python -m cli.main system status
python -m cli.main workflow logs
python -m cli.main workflow status

Alert Strategy

• Page on sustained 5xx/timeout growth in production trust tier paths.
• Warn on elevated fallback rates and route drift before user-visible failures.
• Block promotion when contract integrity drops below accepted thresholds.

Curated Source Synthesis

• Health ownership is mapped per subsystem with explicit SLO and rollback triggers.
• Readiness checks include synthetic path validation, not just process liveness.
• Observability signals are tied to promotion gates and operational decision points.
• Quality guardrails and response normalization are monitored as first-class runtime health signals.

What Is Intentionally Not Included

• Private alert routing channels, pager schedules, and escalation rosters.
• Internal dashboard URLs and tenant-specific instrumentation settings.
• Environment-private threshold values and incident postmortem references.