Metrics & Observability

DriftGuard tracks runtime metrics for every review, record, prune, and graph mutation.

Get a snapshot

from driftguard import build_runtime

runtime = build_runtime()
snapshot = runtime.metrics_snapshot()

print(snapshot["counters"])
print(snapshot["gauges"])

Or via DriftGuard:

guard = DriftGuard()
# ... run some reviews and records ...
snapshot = guard.runtime.metrics_snapshot()

Counters

Counter	Description
reviews_total	Total before_step() calls (non-skipped)
review_warnings_total	Total warnings surfaced across all reviews
review_confidence_samples_total	Total reviews counted for average
reviews_blocked_total	Reviews that raised GuardrailTriggered
reviews_ack_required_total	Reviews that raised GuardrailAcknowledgementRequired
reviews_skipped_total	Reviews skipped due to record_only policy
records_total	Total record() calls
nodes_created_total	New nodes added to the graph
nodes_merged_total	Nodes merged into existing nodes
edges_created_total	New edges created
edges_reused_total	Existing edges incremented
prune_runs_total	Total prune() calls
prune_nodes_removed_total	Nodes removed by pruning
prune_edges_removed_total	Edges removed by pruning

Gauges

Gauge	Description
last_review_confidence	Confidence of the most recent review
review_confidence_total	Running sum (used for average)
review_confidence_average	Rolling average confidence across all reviews

Via MCP

guard_metrics()

Returns the full counters + gauges snapshot as a JSON object — available to any MCP client including Claude Desktop.

Example snapshot

{
  "counters": {
    "reviews_total": 42,
    "review_warnings_total": 18,
    "reviews_blocked_total": 3,
    "records_total": 12,
    "nodes_created_total": 36,
    "nodes_merged_total": 24,
    "edges_created_total": 24,
    "edges_reused_total": 18,
    "prune_runs_total": 5,
    "prune_nodes_removed_total": 6,
    "prune_edges_removed_total": 4
  },
  "gauges": {
    "last_review_confidence": 0.87,
    "review_confidence_average": 0.74
  }
}